Getting started, four tabs, source suggestions, files, follow-ups and 32 copyable requests.
Reviewed · 2026-10-12
On this page · 61
1. Get started
- Open Research Consultant → Research spaces. For a new space, choose a lab and research question; link a parent grant or project when needed.
- Open new research from Overview and write one or two sentences. A title, type and status are not prerequisites.
- Use Add materials to find evidence or attach files. Review the request, then choose Send to AI.
- For ongoing research, open its entry in Research work and continue in the same composer. Reading and selecting evidence uses no AI; sending follows current usage policy.
2. Repeated work and roles
- Continue or derive discussions, literature studies, experiments and manuscripts as needed. Keep follow-up questions for the same purpose inside that work; link evidence when creating different work.
- New questions and notes start privately. Linking to team work does not automatically publish a private conversation.
- Authorship and access are separate, and reviewing one version does not approve later revisions.
3. Existing services
- Materials and results shows authorized parent project and grant materials. Upload files, public links or connected cloud files to the existing parent scope; the global Materials menu exposes the same record ID.
- Spaces do not duplicate files into a separate store. External connections reuse existing connection and authentication screens.
- Members must also have parent-scope access; space membership does not expand source permissions. Manuscripts retain their editing origin and version.
- Live Overleaf, DOCX and HWPX editing synchronization requires separate adapters.
4. Continue across AI surfaces
- Use ▾ beside the send button in either a new or follow-up request to choose: Answer in this work, Continue in AI chat, or Continue in ChatGPT.
- Selecting a destination runs no AI. Answer in this work sends the request. Continue in AI chat prepares your final question, selected evidence and files in the consultation panel; review and send there to start.
- Continue in ChatGPT prepares handoff text. Copy it into ChatGPT with LUNA connected. Copying alone neither transfers nor saves anything externally.
- Reuse the same work and conversation for the same purpose. New work is created on actual send; switching a prepared request’s destination must not run it twice.
- Only delivered and saved content and files are preserved. Check the receipt’s work ID, files and download links. Choosing a destination does not broaden private conversation access or sharing.
5. Research decisions and reproducibility
- Record multiple environments, datasets, metrics, baselines and repeat conditions. Existing workflow and reproducibility import uses the approved ReproKit result contract.
- Producer-reported results, researcher decisions and independent verification remain distinct. This space does not automatically dispatch remote experiments or paid AI.
6. Use the four tabs
- Overview: review work to continue, recent outputs, upcoming events and suggested research. Research work: choose work from the list or map and continue its questions, conversations and results in one view.
- Materials & results: search sources and saved outputs and pin important materials. Open AI source suggestions here too.
- Schedule & tasks: create or link events, GitHub tasks and report requests. Open Research team and space settings from the space menu.
7. Core materials and AI drafts
- Search and pin important sources in Materials & results. Pinning and unpinning do not copy, delete or change source access.
- Private pins enter only your own AI context; team sources still require current source permissions. Attach relevant evidence and choose Ask AI to save a request on the same work.
- Choose Save note to preserve a private memo without AI; it does not replace the current result. Waiting, running and failure reflect actual request state; saved stages remain preserved.
- AI requests follow administrator policy and personal limits. Reading, saving notes and pinning make no new generation call.
- Pins and AI drafts are not factual verification or researcher approval.
8. Members and manuscript authors
- Edit a member inline to load their current role and access. Only researchers with current lab and parent access can join.
- Permission changes retain exact previews. In a manuscript record, Authorship for this manuscript stores multiple roles and order separately.
- Authorship is versioned without changing space access. Existing space roles are not automatically copied to every manuscript.
9. Search, pins and space-specific schedules
- Use the single search row and expand Advanced filters only when needed. Search and page through all accessible parent materials in Materials & results, then pin them there.
- The pin button shows saving; the row and pin count update together after success. Use the active pin to remove it in place.
- On failure, review the current state and error before retrying. Pinned materials shows the full list independently of the bounded AI evidence bundle, without deleting files or history.
- Schedule & tasks defaults to this research space. Search Link existing event or task and save a private or team reference.
- Select the full parent scope in Advanced filters. Calendar and task links retain the space filter.
- + Event and + GitHub task open the standard editors. Request, review and apply an AI draft there; edits made after a request are preserved.
- The existing save/send confirmation creates the real item and retains its research-space reference. Reads, links and pins incur no AI usage; generation follows administrator policy.
- The More menu guide links directly to each step. Web, personal AI and ChatGPT use the same space IDs, references, versions and permissions; this does not imply copying an entire external conversation.
10. Find earlier analyses and results
- Existing goals, code, reports and discussions retain their original IDs and permissions while being linked into one work list. Open work or search Materials & results for originals and downloads.
- The current revision and latest saved checkpoint represent a conversation; earlier versions remain in detailed history. Distinct reports are not merged by matching titles.
- If evidence is inaccessible, the owner may see only a preserved-result notice. Restore current connection and permissions before opening content or downloads.
- Never infer or claim to have read blocked content. A classification-only change with identical content, version, connection and current access does not require regenerating saved results.
11. One work list and research map
- Choose List or Research map in Research work. Selecting work opens the right panel. Use ↗ or Ctrl/⌘-click to open the same work on its own page with a stable URL. The default map groups results inside their owning work; Expand outputs reveals exact report and code-result cards.
- Renaming work changes its display name without rewriting source titles, content, versions or evidence, or duplicating work. Distinct works are never merged by title.
- Solid workflow/derivation links and dashed reference/validation links are both visible by default. Multiple origins, references and sections retain their exact report IDs and versions even when output cards are folded.
- Counts distinguish the works, outputs and materials currently shown, not the total saved collection. Use the next work range or a focused neighborhood to navigate.
- Renaming or reordering the list does not reposition the same works and relationships. Goal settings, literature, strategy analysis and specialized code tools remain available from that work’s settings or detail screen.
- Source changes require deliberate evidence review and refresh. Disconnecting does not delete the source or earlier history.
- Maps, search and opening a work URL call no AI and create no work or conversation. The panel and separate page use the same work, current permissions, conversation, results, files and history.
- While drafting, use ↗ to open a new tab. Unsent text and selected files remain in the original tab. Open in research space returns to that work. Use the original tab to resume its search, map position and draft.
12. Trash and a new research cycle
- Review the target and effects before moving work or a space to Trash. Saved conversations, outputs, attachments and external originals remain preserved.
- A new cycle archives earlier work while retaining materials and members. Restore from Trash or previous cycles without automatically restoring work individually trashed earlier.
- Related local pending or running requests stop, but a call already sent to an external provider cannot be guaranteed recalled. Restoration itself queues no AI.
- Original automation policies remain preserved, so an enabled policy may execute at its next scheduled time after restoration.
13. Navigate quickly and inspect history
- Switch between List and Map and sort by recent activity, creation, execution state, attention or title. Changing work loads its detail in place and retains the draft question for each work.
- Open Records and filter All, Conversation, Results and versions, Executions or Links. An answer and its result are grouped.
- A conversation started from an earlier report or selected passage retains its original thread. Load earlier records on demand.
- No execution requested, generating and needs attention describe execution separately from saved outputs. Downloads require actually preserved files.
14. Save naturally from a conversation
- Ask the AI consultation or ChatGPT: “Save this conversation and its results to the current research work.” LUNA prioritizes an explicit work, the bound conversation and a previous save receipt. Choose by title when several targets match, or preserve privately in the personal inbox before selecting a destination.
- New work requires an explicit request; retrying the same request does not duplicate it. Review the destination, content and files before saving.
- The personal inbox supports downloads and linking to existing work. Saving text makes no additional platform model call.
- Files can be uploaded in resumable chunks up to 8 MB per bundle; larger materials may be saved as accessible source references. A source link is not a preserved binary copy.
- Check the completed save receipt: uploading means only received parts are preserved. External messages and files that were never delivered are not automatically collected.
- Team sharing and external publication retain separate scope reviews.
15. Full-width map and images
- The map uses the full work area. Derivation and dependency relations guide its left-to-right layout, including independent work and loops.
- Select a node for detail and return to the same map position. More offers PNG, JPEG and SVG exports of the currently loaded range, not a claim to cover every record.
- No AI call is made.
16. Attach files and use code results
- Use Add materials for files, images and platform evidence; paste links into your question. Upload up to 50 files totaling 100 MB through resumable chunks.
- Originals are preserved while AI excerpts are prepared in the background. Preparation is not a full read of every original; extraction failures and text or image limits are shown separately.
- Actually completed ChatGPT file transfers can be attached with the same permission checks. Missing code output has no download; execution history shows the failure.
- Import preserved text code or a validated ZIP as a code changeset and inspect each diff. ZIPs are never executed or automatically installed.
- Review repository, branch, baseline commit and changes before approving GitHub publication. Reconcile uncertain responses to avoid duplicate posts.
17. Keep existing work in My saved items
- Star existing work, conversations or results to keep a shortcut in My saved items. This differs from a preserved copy of delivered conversation content and files.
- Shortcuts reopen the same work and version; source permission revocation blocks content access. Web, AI consultation and ChatGPT use the same bookmarks and save receipts.
18. One library and pinned materials
- Use All, Favorites and Saved conversations and files in one personal library. Old favorite URLs and folders remain valid.
- Unstarring does not delete preserved originals. Materials and results initially shows pinned items; search includes all accessible materials.
- Toggle the same pin to unpin and clear search to return to pins. Pinning does not expand source permissions.
19. Private work and team discussion
- Click Only me on a work item to manage sharing. Review the exact body and selected files before sharing.
- Earlier private discussions and personal memory are not shared automatically. Continue in team view to accumulate attributed discussion under the same canonical research work.
- Sharing withdrawal, membership and source access are rechecked; private originals remain preserved. ChatGPT uses the same work ID and current permissions and reflects only delivered content.
20. Add schedules and export maps
- Use Add to choose an event, GitHub task or report request and then create or link existing. Creation opens the standard form with optional AI drafting.
- Linking does not duplicate the original or create an external item. Select one image format before downloading a map.
- Dashed reference and validation relations identify evidence or verification links, separately from derivation order.
21. Report links and research team settings
- Use the work or report share icon to copy its link, copy its title and link, or share to another app. A work link opens the same work on its own page; use the exact result or version link when referring to a specific output. Copying uses no AI, grants no access and still requires the recipient’s current permission.
- Review recipients, permission changes and Notion publication in sharing settings. In Research team, choose Add researchers, search names and select several candidates.
- Review the access level and add them to the list. Authorship is optional and permits multiple roles.
- Edit roles and access from the member row. Granting management or ending participation requires an impact confirmation; the owner retains management.
- Basic settings holds the name, description and parent scope. If candidates are missing, check active lab membership and parent grant or project participation.
- On a save conflict, retain your input and refresh membership. Private conversations, memory and source permissions are not automatically shared.
22. Work menu and next research plans
- Open the work menu to rename a work item, manage multiple references, view history or change relevant settings. Display names do not alter original titles or result versions.
- Find conversations, executions and revisions in one panel. Save a report next action as a research plan; private research replies can also suggest a plan within the same AI call.
- Review the rationale, expected output, completion and stop criteria before adopting or deferring. Adoption prepares follow-up work; it does not execute AI or verify a hypothesis.
- Continue plan with AI explicitly starts a request under current policy and allowance. Changed evidence requires review.
- Report Next actions starts with three suggestions. Use View all for more and chapter links to inspect their evidence. Grouped suggestions retain original candidates and sources.
- Distinguish started, deferred or dismissed suggestions and genuinely different experiment conditions. Print view and PDF omit the interactive Next actions section while preserving the report body and citations.
23. ChatGPT files and accurate save coverage
- Ask to save the delivered discussion and generated files. Reuse the linked work, or keep an unresolved destination in your private inbox.
- Only content and bytes actually delivered are preserved. Up to 50 originals totaling 100 MB use resumable upload; already preserved files are reused by ID.
- Preserved does not mean fully read by AI. Check downloads and preparation states in the receipt.
- Broader sharing, publication and GPU execution are not implied by a save request.
24. Refine a request through conversation
- Write a few sentences in New research or Continue this work, then select sources and files in Add materials. Put links in the text. Both composers use the same refinement, attachment and destination controls.
- Refine request opens a conversation on the right. It checks the current request, work and selected evidence, suggests a research kind and identifies missing conditions. It reuses known conditions and asks one or two important questions at a time.
- A choice fills the reply box. Add or edit details, then choose Send reply. You may leave a condition undecided or ask for a recommendation. Repository candidates are authorized connection metadata, not proof of write permission or server availability.
- Preview the prepared request and conditions, distinguishing your input, existing evidence, AI suggestions and unknowns. The composer changes only when you choose Apply to request. Undo restores your earlier input.
- Applying does not execute research. Edit the result if needed, choose this work, AI consultation or ChatGPT from the destination menu, then send. You may also skip refinement and send directly.
- Closing and reopening preserves the saved conversation. On mobile, closing returns to the original composer. An older response never automatically overwrites text you changed.
- Starting refinement and sending each additional reply request one personal AI attempt each. Page reads, saved conversation reads and applying use no AI. Closing the panel does not cancel an already accepted AI job.
- Administrator policy, remaining allowance, provider and time limits apply. Failures retain input and conversation without automatic paid retries. Up to 50 files totaling 100 MB are supported; uploading is distinct from analyzing all original content.
- Selecting a saved suggestion only fills the request and evidence; Save memo stores a record without AI. Team requests use only currently authorized shared evidence.
25. Continue request drafts and results in ChatGPT
- You can hand off a rough request without finishing refinement on the website. Choose Continue in ChatGPT from the destination menu for either a new request or a follow-up.
- The handoff includes the latest composer text, selected sources, prepared attachments and existing work, conversation and result versions. An applied refinement also carries the same request session and accepted revision. Preparing and copying the handoff calls no AI, creates no work and transfers nothing externally.
- Paste the text into ChatGPT with LUNA connected and check that the host exposes the needed tools. Ask: “First clarify the request by asking for missing conditions; proceed after we settle it.”
- Preserve the same request session, work and evidence versions when saving a request refined in ChatGPT. Saving externally prepared requests and results does not rerun platform AI. Draft storage, applying a request and executing research remain separate actions.
- Ask: “Save the discussion delivered so far and the generated files to this work, then show download links.” The platform preserves only conversation content and files actually delivered through its tools, not the entire external chat automatically.
- A filename or temporary link is not proof of preservation. Check the receipt destination, file list, downloads and readiness. Use web upload if the host cannot deliver files.
- Confirm a destination for new work and retain the existing work, conversation and exact result version for follow-ups. Continue through the stable work link; include the exact result or version link when reviewing a specific output. Opening a link does not save drafts, external conversations or files. Private conversations and evidence do not become team-shared automatically. Publication and paid GPU execution have separate scopes.
26. AI source suggestions
- Open AI source suggestions in Materials & results. It works without pins and first reads saved suggestions without AI.
- For fresh suggestions, check the source scope and expected usage, then choose Generate suggestions · 1 AI attempt or Suggest again · 1 AI attempt. Review titles, reasons, sources and whether full text or only summaries were checked.
- Pin useful materials. Research with selected materials opens the composer with selected evidence and an objective.
- Actual AI research starts when you review and send the question.
Start with a short question
First question
Explain the attached paper’s main question, method and limitations in plain language.
Select the paper in Add materials before sending.Clarify an idea
I want to study how to reduce bias in recommender systems. Suggest three questions to narrow the idea.
Leave unknown data and equipment as questions.Refine wording only
Refine “I want to improve a model with little data” into a request with a clear objective and expected output.
Refine through conversation, then Apply to request. Starting and each additional reply use one AI attempt; research execution is separate.Save a note
Save this as a private note in the current work: first check uncertainty caused by the small sample.
Saving a note makes no additional platform AI call.One next step
Identify the most important uncertainty in this work and recommend one next step with a reason.
Adopting a suggestion and executing it are separate.Clarify a paper goal together
I want to prepare a paper. First check this space’s objective and selected literature. Reuse known research questions, venue, deadline, time and resources; ask only one or two missing questions at a time. Use my answers to prepare a final request covering the literature scope and desired outputs. Do not start research or execution before I settle it.
Use Apply to request to put the accepted draft in the composer. Refinement and actual research execution remain separate.Prepare a repository-aware implementation request
I want to apply this method to the project code. Show authorized connected repository candidates and help choose a repository and base branch. Ask only for necessary environment, implementation scope, small-example checks and completion criteria. Distinguish actually reading a repository from checking connection metadata. Do not commit yet.
A repository connection does not prove write permission or server availability. Review actual commits separately.Find literature and core sources
Find materials
Find materials about contrastive learning in this space and show their titles and sources.
Searching saved materials is a read within current permissions, with no new AI generation.Recommend core sources
Recommend core sources for this research. Show saved suggestions first, and tell me the AI usage before generating new ones.
Reading saved source suggestions uses no AI; generating suggestions uses one AI attempt.Pin selected sources
Pin the first and third materials I just reviewed as private core sources in this space.
Check the selected sources and audience; source permissions stay in effect.Compare papers
Compare the questions, data, methods, metrics and limitations of the three selected papers. Distinguish full text you read from summaries.
A preserved file is not proof that its entire text was read.Start from selected sources
Use the selected sources to identify unresolved questions and propose the most feasible research direction.
Use Research with selected materials, review the evidence and objective, then send.Find references across sources
Search this space’s literature collection for contrastive learning papers. Include collected and pinned Zotero references, showing each paper once. Do not start a new search run or AI summary.
Read the saved collection under current access permissions.Read the original abstract
Show the selected paper’s original abstract, DOI, authors and publication date. Keep AI summaries separate from the original abstract.
If no abstract is available, say so rather than inventing content from the title.Explicitly request a saved summary
Request a saved LUNA AI summary and separate relevance to this space’s purpose and questions, using this paper’s accessible abstract or preserved body. First check source coverage and AI usage, then show the summary when it actually finishes.
A host answer and a paid platform summary request are separate. Queued does not mean completed.Correct bibliography
I want to correct the selected reference’s author name and year. Show the current bibliography first. Apply only the values I supply to this space without modifying the Zotero original.
Review field-level changes and the current revision. Recollection preserves corrections.Export selected references as BibTeX
Export only the three selected papers as BibTeX, including authors, year, DOI and original abstracts. Do not substitute AI summaries for abstracts. Tell me if you cannot create a real downloadable file.
Claim a download only when the host actually creates a file.Export all search matches
Export all “graph learning” matches in this space’s literature as RIS across every page and report the count. Do not upload anything to Zotero.
Export is a standard bibliographic snapshot, separate from Zotero synchronization.Restore the source bibliography
Restore only the selected paper’s publication date to the source value. Preserve other edits and its work references.
A null field removes that override and restores the source value.Use literature as implementation evidence
Link the two selected papers as references for a new model-design task. Check readable source coverage first, distinguishing applicable method details from unknowns.
Retain paper IDs, versions and permissions, and state what source text was read.Register references as materials
Register these DOIs and paper links as references in this research space. Reuse existing materials and check the destination project/grant and audience. Do not generate new AI summaries.
Bibliography registration, original storage and AI analysis are separate; registration requires actual host tools.Register only missing papers
Check each paper from my earlier registration request and register only items not yet accepted. Do not recreate queued materials. Distinguish storage failures from missing PDFs.
Reuse existing request receipts and material IDs.Attach a PDF to an existing reference
Attach this published PDF to the already registered paper without creating another reference; preserve any earlier preprint. Tell me when it is stored and downloadable, and report text extraction and AI analysis separately.
Verify completion from delivered bytes and the platform preservation receipt.Find bibliography in Materials
Find the references collected in my research space through Materials. Search by author and DOI and show each paper once after checking duplicate source identities.
Search only currently accessible references without exposing another space’s private corrections.Export corrected bibliography
Check that these bibliography corrections made through Materials also appear in the research space, then export only the selected references as BibTeX. Preserve Zotero originals and other spaces’ corrections.
Read and export the same reference IDs and space corrections under current permissions.Check original and extraction status
Check original-storage status for the registered papers. Distinguish PDFs, archived HTML, extraction failures and missing originals, and show download links for preserved files. Do not start AI analysis.
An archived webpage is not presented as an acquired PDF.Work through research stages
Form a hypothesis
Propose two testable hypotheses grounded in the selected literature and observations that could refute each.
Separate source claims from newly proposed interpretations.Plan an experiment
Design an experiment for the first hypothesis, including data splits, baselines, metrics, repetitions and stop criteria.
Planning does not start an experiment or a GPU run.Plan implementation
Inspect the linked repository’s current structure, then propose small implementation steps and the necessary checks.
Label proposed code, synthetic checks and actual execution results separately.Interpret results
Calculate changes from the baseline and uncertainty in the attached results, separating supported conclusions from claims needing more experiments.
Select the actual result file and check the calculation evidence.Draft a manuscript section
Draft a results section using only reviewed findings in this work. Do not invent numbers or citations.
The draft does not automatically change the manuscript source or author approval.Plan training with unknown resources
I want to prepare training for this model. First check access to the selected data, splits, baselines and metrics. I do not yet know whether a GPU server is connected and available. Do not assume missing resources exist; ask about possible environments and time/cost limits. Save only the execution plan for now; do not start training.
Unverified environments and permissions remain unknown. Saving a plan does not execute a paid GPU run.Inspect paper-code provenance
Inspect the paper’s linked code repository and report whether it is official, the inspected commit, license and environment requirements. Do not commit to my project or run it on a server yet.
Distinguish official links, inspected code, reuse rights and validation results.Edit saved research defaults
Show this work’s research conditions. Change the default evaluation metrics to the ones I confirm while preserving earlier request/result settings. Save settings only, without execution.
Check the kind-specific fields and current settings revision.Continue existing work
Follow up in the same work
Analyze the sample-size limitation in this work’s latest result in more detail.
Open the existing work before sending to continue its conversation and results.Continue an exact report
Use section 3 of the selected report version to prepare the next analysis plan.
Keep the selected version as evidence rather than silently substituting the latest result.Review all suggestions
Show all saved suggestions for continuing this work so I can compare their reasons and expected outputs.
View all reads saved suggestions without a new AI call.Derive separate work
I want to start separate research comparing another dataset while referring to this result. Preserve the link to the original work.
Keep same-purpose questions in existing work; derive work for a different objective.Check status and partial results
Show this request’s actual state, saved partial results and files available for download. Do not rerun it.
When failed or quota-waiting, inspect saved results and resume conditions first.Append a request/result pair
Continue within the current work without creating another item. Check two assumptions in the latest result and preserve this request and answer in the same work.
Use continuation_mode=same and never attach an earlier answer to a pending request.Derive from an exact result
Create separate implementation work from the selected output version. Keep the original conversation intact and link this literature and result as evidence.
continuation_mode=new creates the new work and originating evidence on actual save/send.ChatGPT, files and preservation
Hand off the final request
Check the final question and selected evidence in the copied LUNA request, then continue this work. Check attachment readiness first.
Verify the LUNA tools actually available in ChatGPT. Copying alone does not transfer anything.Save discussion and result files
Save the discussion delivered so far and the actual generated files to the linked LUNA work, then show download links.
Verify actual content/file delivery and a completed receipt. Saving does not require another AI run.No work selected yet
Save this conversation and its generated files in my private inbox. I will link them to research work later.
Actual uploads support up to 50 files totaling 100 MB. Filenames or temporary links alone do not establish preservation.Check original downloads
List the originals preserved in this work with their readiness states and provide downloads for completed files.
Use the website upload when the host cannot deliver the file.Refine and save from ChatGPT
Help refine the supplied LUNA request in this conversation. Preserve its request session, work and evidence versions, and ask about missing conditions. After saving the settled request, preserve the results I ask you to prepare here and the files actually generated in the same work. Do not regenerate them with platform AI. Show the saved work link, file list and download availability.
The host must expose the required LUNA tools. Check draft/result receipts and actual delivered files; do not describe this as automatic synchronization of the entire external chat.Team, relationships and schedules
Rename work
Rename the current work to “Small-data comparison experiment” while preserving its results and references.
Renaming does not create a new work item.Link several references
Link the two selected works and the report as references for this work, then show their relationships on the map.
Reference, derivation and prerequisite are distinct relationships.Share selected results
I want to share only this summary and the selected result files with the research team. Show the recipients, body and files first.
Share only the reviewed scope; private conversation history and memory are not automatically exposed.Link an existing schedule
Link the existing next research meeting to this space without creating a duplicate event.
Reuse the existing calendar or GitHub task; review any external creation separately.33. Automatic recommendations and usage
- Open space settings in Research team to turn automatic research recommendations on or off. Existing and new spaces default to on, preparing suggestions from existing reports and consultations.
- Opening pages, reading suggestions, searching and sorting make no new AI call. Relevant saved results refresh recommendations in the background; the first view shows saved suggestions.
- Additional automatic AI enrichment defaults to zero attempts. Set a daily cap if needed; the space owner’s allowance and administrator policy also apply. Changing settings never resets consumed usage.
- Turning off stops subsequent automatic recommendation generation while preserving existing suggestions and history. Pending work rechecks the setting and permissions immediately before execution.
- Private consultation or source recommendations are not automatically shared with other members. Changed or revoked evidence requires revalidation. Preparing a suggestion does not execute research, prove success or constitute human review.
34. Compact cards and work search
- Open titles in the Projects and Research spaces sections of a grant or project. The directory uses the same cards without repeating a parent already shown.
- Work and result counts use currently accessible entries. A work’s conversation, report and execution history are not counted as separate work. A + marks a bounded count.
- Choose List or Map and type to search Research work. Search waits for composition to finish; Enter searches immediately. Use the compact sort menu for recent activity, creation, status and other criteria.
- Searching and changing views preserve the draft and selected work. Renaming work or adjusting map relations retains the same work ID and history.
35. Research kinds, requests and results
- A research kind describes the activity: planning, literature review, model design, experiment design, code implementation, training, validation, reproducibility, analysis, discussion, writing, reporting or scheduling. Reports, code files and literature tables are result formats.
- AI proposes a kind during refinement. Direct requests receive an initial classification; use Change research kind in the work menu to correct it. Manual choices are never overwritten automatically.
- Every work uses Request, Results and Continue this work, in that order. Expand long requests and open or download results. Find initial requests, conversation, execution and earlier versions in History under the work menu.
- Changing kind preserves sources, results and permissions; it does not authorize server execution or GitHub writes. Distinguish a training plan from actual training and a validation plan from completed validation.
36. Continue earlier goals and literature
- Earlier paper-goal settings and discovered literature can be organized as a separate preparation work. Read its research objective and conditions as the request, and open the literature bundle as its result.
- The literature result includes search and observation history, selection state and sources. Open or download it, distinguishing reorganized existing records from new AI analysis.
- Earlier strategy and detailed-report IDs, conversations, reports, citations and downloads remain intact. Linking preparation as the logical starting point never invents historical runs or conversations.
- Earlier settings and literature URLs lead to the preparation work. Its automatic execution settings retain the earlier policy; reorganization does not resume stopped AI or incur additional AI charges.
- Renaming work never creates another map node for the same work. Work and result artifacts remain distinct; expand result versions when needed.
37. Browse the space literature collection
- The Space literature card in Materials & results combines papers collected by literature work with pinned Zotero references. Its compact view shows three references with pinned items first.
- View all opens a full-width literature page scoped to this research space. Each title opens its permanent detail page. Going back preserves the search context. Search by title, author or DOI and move between pages.
- Sources with the same DOI share one collection entry while retaining original sources, versions and access. Similar titles alone never merge papers. Your entire Zotero library is not automatically published to the team.
- Add materials in new and follow-up requests searches these same references. Selecting one links evidence without executing AI. Unpinning does not delete the paper or its original files.
38. Original abstracts and on-demand AI summaries
- Original abstract shows the provider-supplied abstract, or explicitly states that none is available. Browsing, collecting, pinning and exporting never automatically generate an AI summary.
- Choose AI summary and relevance when needed. One request summarizes the source and separately assesses its relevance to this space’s purpose and questions. One request follows current usage and access policy; queued, running and completed states remain distinct. Repeated clicks do not duplicate the same active request.
- Use an available abstract first, otherwise an accessible preserved body or excerpt. If no text is available, add an abstract or readable source. A title is not evidence for inventing a paper summary.
- AI summaries show actual coverage, such as abstract or body, without replacing the original abstract. AI relevance separately shows direct, partial or not-established relevance and its reason. Relevance sorting uses only current saved assessments, never inventing scores for unevaluated references. Source assessments remain separate. Source changes mark older summaries as based on an earlier source; regeneration requires another explicit request.
39. Edit standard bibliographic metadata
- Use the pencil in reference details to edit title, authors, publication date, publication, DOI, original URL and abstract. Add, remove or move authors up to change order. Expand More bibliographic fields for volume, issue, pages and publisher.
- Save applies a space-specific correction without overwriting Zotero or other external originals. Cancel returns to saved values; Restore original bibliography removes corrections. Recollection preserves manual edits.
- Bibliography uses CSL-JSON types, creators, dates and identifiers. DOI and date formats are validated. A conflicting or invalid DOI raises an error rather than silently merging papers.
- Current search, Add materials and new exports use edited metadata. Earlier report citations, files and source versions remain preserved. If another screen saved first, refresh the latest version before editing again.
40. Export selected or all references
- Use the card download icon to choose BibTeX, RIS or CSL JSON. Export individually checked references, the entire collection, or all current search results.
- Select all covers matching references across every page. Uncheck individual items to exclude them and verify the actual export count. References whose access has been revoked are not exported.
- Files contain current bibliography and original abstracts. AI summaries and private research relevance never replace original abstracts. Import a downloaded BibTeX or RIS file through Zotero File → Import.
- Export creates a point-in-time copy, not two-way Zotero synchronization or PDF transfer. Check for duplicates with existing Zotero items after import.
- For direct delivery, open Send to Zotero, review the destination and permission to send selected citations, and apply that binding first. Select 1–20 individual references, preview the citation fields, then choose Apply reviewed changes. Existing Zotero bibliography is preserved; only new citations and collection membership are sent. PDFs, notes and deletions are excluded.
- verified describes remote readback at the displayed verification time. For partial, uncertain or conflict results, inspect the same request instead of blindly submitting a new one. Import new items adds collected references to the space; it neither authorizes unattended AI-to-Zotero writes nor enables collection scheduling.
41. Continue the same work or derive new work
- In follow-up requests, Continue in this work appends paired requests and results to the same work. New derived work creates a separate item only on actual send and retains the originating work or selected output version.
- Report suggestions also offer Continue in this work or New derived work. Continuation and delivery destination are separate choices. Choose Save as a note in the send menu to save without running AI. Web, AI consultation and ChatGPT retain the same request ID, evidence and settings versions. Preparing, copying or changing a destination creates no work or paid execution.
- Research flow puts earlier work below later causal work. Reference links are not silently treated as prerequisites.
- Edit kind-specific defaults in Research conditions under the work menu. Applied refinement conditions and saved defaults are reused, while each request preserves its settled snapshot. Editing settings does not authorize GitHub writes or server execution.
42. Connect literature to code implementation
- Select references in Add materials when requesting model design or implementation. AI uses the current authorized source and exact version, distinguishing abstract-only access from inspected body text.
- Future code links retain official-versus-third-party status, supporting source, repository, commit, license, data and environment. A URL alone proves neither official authorship nor permission to reuse code.
- The intended trace is literature → selected method → source code version → target repository/branch → patch → checks and experiments. The collection preserves evidence and extension fields for this flow.
- This literature feature does not automatically enable training, server execution or external commits. Distinguish generated code, synthetic checks, actual experiments, independent reproduction and scientific conclusions.
43. Literature results, printing and collection settings
- Literature results focus on references and expandable abstracts/reference information. Change priority or exclusion under each item’s ⋯ without deleting the paper. PDF links appear only when a known URL is available.
- Print view opens every reference matching the current filter, not just the visible page. Search, settings and action controls are excluded from printing.
- Manage recurring collection and analysis in the work’s ⋯ → Automation settings. Collect literature now requests one additional collection without enabling automation; current access, query conditions and queued status are rechecked.
- Collection updates metadata and source observations. It never automatically generates per-paper AI summaries and preserves earlier collection and analysis history.
44. Find and edit literature in Materials
- Open References in Materials or search by title, author or DOI. References saved in a research space are discoverable in the same catalog; do not register another copy.
- Select a card title to open its wide permanent detail page; View all opens the dedicated literature list. Translation, sharing, favorites and print are at the top, followed by authors, venue, publication date, abstract and source links. Original abstracts and AI summaries are separate. Generate an AI summary only with the explicit button or request.
- Edit only the required bibliography fields. Materials and the space use the same reference ID and correction. Check the applicable research space; other spaces and Zotero originals are not silently overwritten.
- Select references and choose BibTeX, RIS or CSL JSON under Export. Distinguish selected references, all search matches and the complete current scope. All includes every page in that scope, not just the visible rows.
- Use File → Import in Zotero to read the exported file. This is a bibliographic snapshot, not automatic two-way synchronization. Unpinning does not delete the original material.
45. Register references and request source storage in conversation
- Tell native AI or ChatGPT the research space and the DOI, source URL or bibliography to register. Example: “Register these papers in this space, reuse existing references and do not generate new AI summaries.”
- Check the destination project/grant, audience, new/existing references and original attachments. The host must actually expose the required tools. Search-only tools cannot prove registration.
- Check each returned material link and status. With a linked research work, the actual registration request and receipt also appear in its history. Ask for a destination if work context is unclear; never invent earlier conversation that was not delivered. Bibliography registration, pinning, source storage, text extraction and AI analysis are distinct. Reuse the same request ID when retrying to prevent duplicates.
- Add a PDF later from the reference detail page. Active authorized participants can contribute files, preserving contributor and file versions. Identical files are not linked twice and earlier PDFs are not overwritten. Distinguish published papers, preprints and supplements. In conversation, identify the existing reference before attaching a PDF. Example: “Attach this PDF to the existing reference without creating another paper.” Only delivered files or accessible originals can be stored.
- A ZIP or download link generated in ChatGPT is not proof of platform storage. Check file delivery and the platform receipt. Registration, reading, pinning and exporting do not themselves generate AI summaries.
- The tool sequence is consultant_material_prepare → consultant_material_commit → consultant_material_status. Review the destination and audience, then pass the unchanged preview_token and review. A cached host can use the existing consultant_thread_prepare save / materials envelope for the same service. If neither route is exposed, refresh the connection tool list.
- A request accepts up to 50 items and 100 MB of combined attachments. Distinguish “register bibliography only” from “also store an accessible PDF.” A paper landing page is not reported as a saved PDF; supply an actual PDF file when needed.
46. Understand storage, PDF and extraction states
- Once registered, bibliography can be searched and edited before source collection finishes. Do not create duplicate references because earlier files are still processing.
- Queued and running storage are distinct. Stored means a verified preservation receipt exists. An archived HTML page is not a PDF.
- If the original is stored but text extraction fails, it remains downloadable. Combined source attachments may total 100 MB; safe text extraction supports up to 20 MB per file. Larger originals are preserved with an extraction-limit status. Scans, protected documents or unsupported formats may not yield readable text.
- Retryable storage failures use bounded attempts and backoff. For persistent failures or access requirements, inspect the reason and supply an accessible file. Access and login restrictions are not bypassed.
- A PDF downloadable from its reference is distinct from source archiving and text extraction. Linking a PDF does not prove external synchronization, extraction or AI analysis. AI analysis is a separate request with its own usage. Stored originals or extracted text do not mean AI analysis has completed.
Bibliography and PDF visibility
New public-paper bibliography is visible by default to active signed-in platform users, not anonymous visitors. PDF originals, private analysis and research spaces retain their existing access. Reusing an existing private source does not publish it. Edit bibliographic fields and grant/project/researcher metadata on the permanent literature page. External Zotero originals are not overwritten.
Find and attach PDFs from ChatGPT
Example: “Find references in this space that lack a PDF. Locate public originals, verify the title and DOI, and attach them to the existing references while preserving earlier files.” Tools import a public HTTPS PDF or an actually transferable ChatGPT file and verify bytes, size and hash. They never bypass login or paid access. Original preservation, text preparation and AI analysis are separate states; an unavailable file transfer is not a saved result.
Compare literature and save to the same work
Example: “Compare these three papers for the Paper Goal and Literature Survey work. State the PDF version and actual coverage; tabulate model structure, data, metrics and limitations. Save the delivered analysis as another request/result pair in this same work and attach the comparison file.” The existing work ID and earlier results are preserved. An excerpt is never labelled a complete-paper reading. Specify explicitly when a new derived work is intended.
File transfer and retry
Example: “Save the actual image and BibTeX files just created as outputs of this work. Report each saved status and download link.” Attach up to 50 files, combined 100 MB. Reuse the request key for retries; reselect files to renew expired transfer links. The file tool must actually be available in the ChatGPT host. If unavailable, transfer the files through Add sources on the platform. A filename or temporary link is not a preserved file.
Edit, export and apply requests
Search, select required references or all references, then use the download icon beside search for BibTeX, RIS or CSL JSON. Example: “Correct the selected authors and DOI while preserving the original.” AI summaries run only after a button action or explicit request. Applying the refinement preview updates both the current composer and its local draft, ready to inspect before sending. Dashed graph edges indicate references; solid edges indicate research flow.
Start paper analysis in the same work
In the existing work composer, use Add sources to select one reference or PDF. Enter a goal-specific question, choose Paper analysis and discovery in the Send to AI menu, and send. Source reading, backward/forward/topic search, registration, synthesis and manuscript drafts are checkpointed. Search uses the public paper title or your explicit public query, not private goals. Verified references and final recommendations are registered automatically; PDF collection is queued with actual storage status. Missing originals limit analysis to the abstract. Forward and similar searches cover the first 25 results each; request further searches separately. Paywalls and institutional login are not bypassed; unavailable originals retain a reason. Registration, original collection, AI analysis and author review are separate states.
Goals, notifications and collaborative review
Use work menu → Research conditions to edit the question, venue, deadline and constraints. Expand Collaborative guidance to opt into a reviewed goal snapshot and next-action notifications. Only the published goal version is shared; private edits do not silently replace it. Notifications open the same suggestion ID used by the composer and do not execute AI. Deferred or dismissed suggestions are not resurrected. Before joint use, check work sharing, access to originals and the shared goal separately. Approved methods also require use and provider-processing permission. Participants use the research team destination for follow-up analysis. Private results are not published automatically. A researcher confirms submission or goal completion.
Improve v1 into v2
Use work menu → Know-how settings to select exact approved versions. The default direction and relevant common/activity guidance are pinned; four stages are not required for every work. Existing work-only know-how remains available in collapsed history. Select currently accessible, non-deleted request/result evidence and distinguish paper-specific fixes from reusable improvements. Compare against the actual active v1 using matched input/model/work conditions, held-out input and human review receipts before considering v2 promotion. Promotion changes future defaults and is reversible; it does not rewrite existing results. The built-in method before initialization is distinct from a reviewed and evaluated v1. Method versions, manuscript revisions and model IDs are distinct. Output-only use, source reading and derivation are separate permissions.
Delete and restore request/result pairs
The author or an authorized space manager can move an accessible request/result pair to trash through its menu. Removed content is excluded from new analysis, method improvement and downloads; other researchers’ follow-ups and existing quotations remain. Restore through Deleted requests and results; this never restarts AI. Management does not grant access to another person’s private conversation. External copies and provider costs already incurred are unaffected.
ChatGPT, consultation, Overleaf and costs
Continue in ChatGPT preserves the same work ID and selected evidence, saving only delivered results. Survey request/status/resume/external, method reads/improvements and pair delete/restore reuse consultant_thread_prepare/commit reviews. Independent host analysis never claims to have executed a protected method. Core instructions are excluded from ordinary results, history and exports. Provider processing permission is reviewed separately; this does not guarantee that outputs cannot reveal or imply knowledge. Download the draft and BibTeX, review them, then incorporate them in Overleaf; there is no automatic bidirectional sync. Per-request, daily and provider limits pause work while preserving completed stages. Inspect saved stages and usage before explicitly resuming. Team consultation tools currently support requests and status; use the web for resuming and method management. Host tool availability, new server PDF collection, AI quality and Lightsail latency require separate verification.
Keep only the know-how you choose from a discussion
In the same work composer, ask: “Separate the literature-analysis and manuscript-writing know-how from this discussion. Propose only the changes to each existing method.” Only delivered, saved requests and results are evidence. External ChatGPT or Codex can prepare structured proposals using returned evidence IDs and current permissions, then use the existing reviewed save contract. On the proposal page, select only the drafts you want to keep and review the save. Unselected proposals and active versions remain unchanged. Preparing or saving a draft does not validate quality or approve v2.
Apply several methods together
Use the applied know-how count below the work title, or work menu → Know-how settings, to manage defaults. No empty badge appears for zero items. Search names, handles or activities and review exact versions and roles before saving. Add to a work on the independent detail page lets you choose a space and work and pin an exact version. Type @ in a follow-up, new research request or personal AI consultation to search researchers and know-how together. Choosing know-how adds it only to that request without changing defaults. [[ source citations keep their existing behavior. Multiple selections never auto-merge assets or adopt newer versions. Web, consultation and ChatGPT handoff check the same IDs and current permissions.
Split, merge, originals and shared evidence
One discussion can produce independent proposals that reference the same segment as shared evidence. Drafts start privately. Separate paper-specific corrections and unverified scientific conclusions from reusable guidance. Change organization prepares a split or merge request in the existing composer for review and sending. A merge proposes a new asset linked to exact parent versions. It does not delete or overwrite the versions, history or active state of assets A and B. Changing how a proposal is organized requires a newly reviewed proposal, not an unreviewed browser-side rewrite.
Version inheritance and evaluation
Derived work inherits the exact selected versions of independent assets. For existing work-specific methods, the collapsed Work know-how controls retain version and candidate selection in the original work. Review generalized instructions and create a separate asset before reuse in other work. Execution applies modules relevant to the current stage, such as literature, synthesis, writing or review, while retaining links to other stages. New versions are not adopted automatically. Initial approval reviews the actual instructions and their generalization; it does not claim measured quality improvement. Improvement candidates require actual runs with matched questions, inputs, models and effort, held-out cases and human review before adoption. A candidate trial is a one-off selection that does not change the default and uses the normal AI allowance.
Protection, storage and execution boundaries
An independent detail page shows introduction, ID, handle, authors, usage, current permissions and versions. Editing the introduction or image link does not change execution guidance; image access is checked separately. The default direction is protected guidance. Write a draft or propose a direction improvement, review it and select which drafts to keep. Saving an introduction or candidate does not change the active version. Ordinary catalogs and proposal cards exclude protected instructions. Source review requires current permission and AI-provider processing permission is separate. Reads, selections and structured proposal saves do not run AI, while sending analysis or reorganization requests may use your allowance. Independent external analysis does not prove protected-service execution. Instructions belong in encrypted backups, not code or static files. AI quality, actual ChatGPT tool availability and server speed require separate validation.