desk-research
Skill27 starsChanged 40 days agoMITView on GitHub
Answer a science question by reading existing sources (papers, software documentation, reviews, database documentation) and keep the answer in the repo's shared library, .science/library/. Use when framing a study needs reading, or when a question needs reading before a probe can be chosen or run.
4 other projects use a version of this.
Adopted by
Most recently taken up by
- teodstr/garagetheir own version, first seen Nov 15, 2026
- henrique-simoes/Istaratheir own version, first seen Sep 27, 2026
- PierreAtrocca/claude-skillstheir own version, first seen Jul 22, 2026
- strzalex/claude-skillstheir own version, first seen Apr 24, 2026
Other versions4 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 4
- henrique-simoes/Istara★ 15same name
- strzalex/claude-skills★ 1same name
- PierreAtrocca/claude-skills★ 0same name
- teodstr/garage★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository45 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- ask-mattSkill
- codebase-designSkill
- code-reviewSkill
- diagnosing-bugsSkill
- domain-modelingSkill
- grill-with-docsSkill
- implementSkill
- implement-specSkill
- improve-codebase-architectureSkill
- prototypeSkill
- prSkill
- researchSkill
- retroSkill
- setup-matt-pocock-skillsSkill
- tddSkill
- to-specSkill
- to-ticketsSkill
- triageSkill
- wayfinderSkill
- wizardSkill
- chief-of-staffSkill
- claude-handoffSkill
- loop-meSkill
- setup-ts-deep-modulesSkill
- software-factorySkill
- writing-beatsSkill
- writing-fragmentsSkill
- writing-shapeSkill
- git-guardrails-claude-codeSkill
- migrate-to-shoehornSkill
- scaffold-exercisesSkill
- setup-pre-commitSkill
- grillingSkill
- grill-meSkill
- handoffSkill
- teachSkill
- to-questionnaireSkill
- wait-whatSkill
- writing-for-agentsSkill
- ask-scienceSkill
- experimentSkill
- lab-notebookSkill
- promoteSkill
- setup-scienceSkill
The file
We split this skill into the rules it gives an agent
6 rules. 0 are common, 4 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 3 edited versions we read keep these.
4
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 15
Read .science/SETUP.md for the domain.
- Distinctive: line 64
search with the domain's own vocabulary and several queries per question.
- Distinctive: line 67
check retractions and corrections where the tool reports them.
- Distinctive: line 69
If no literature tool is available, use web search on primary sources and say so under Gaps.
- Very specific: line 73
Every question starts here.
- Distinctive: line 80
Check each report: every claim cited, every search saved as returned, every data source the answer points to named with its release and …
“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Desk Research
Desk research answers a question by reading what already exists: academic papers, software and tool documentation, reviews, database documentation, standards. It is the sister of the bench work in probes: nothing is run, things are found, read and saved.
Everything it finds goes into the library, .science/library/: one shared collection for the whole repo, belonging to no study. Studies link to library pages; the pages themselves never name a study, so any later study can reuse them. The library follows the spirit of Karpathy's LLM Wiki in its simplest form: an index read first, an append-only log, and markdown pages the agent writes and keeps current, with the raw search outputs saved beside them.
Desk research reads; it downloads no data. When the answer points to data a probe will load (a database export, a published dataset, a model checkpoint), the page names it with its source, release and licence, and the probe that needs it fetches it into its own folder (see Data in lab-notebook).
Read .science/SETUP.md for the domain.
The library
.science/library/
index.md one line per page, grouped by topic; read first, before any new search
log.md append-only record of every search and page change
<slug>.md one page per research question or topic
searches/
<YYYY-MM-DD>-<slug>/ one per search run with a tool
query.md tool, exact query, filters, date, number of hits, what was kept and why
results.json the tool's output, saved as returned (or .csv, whatever it returns)
references.bib the kept papers, with DOIsPage, named by slug, flat at the library root:
# <title>
**Question:** <what this page answers>
**Updated:** <YYYY-MM-DD>
## Answer
<the answer, each claim cited to its source by DOI or URL>
## Evidence
- [searches/<date>-<slug>](searches/<date>-<slug>/query.md): <what that search contributed>
- <a database, dataset or checkpoint the answer points to>: <source, release, licence, how to fetch it>
## Gaps
<what could not be found or verified, and secondary sources used in place of primary ones>A page grows: a later question on the same topic extends it instead of starting a new one, and Updated moves.
index.md groups pages under ## <topic> headings, one line each: - [<title>](<slug>.md): <one-line answer>.
log.md is append-only, one entry per event, so grep '^## \[' log.md lists the history: ## [YYYY-MM-DD] search | <slug>: <tool>, <hits> hits, <kept> kept (for <study>), ## [YYYY-MM-DD] page | <slug>: new, ## [YYYY-MM-DD] update | <slug>: <what changed>. The (for <study>) note is the only place a study is named.
Primary sources first: the paper, the official database, the tool's own documentation, never a write-up of them. Follow every claim back to the source that owns it, and name data by the database or the authors' release, at a pinned version. A secondary source is used only when no primary one is reachable, and is marked as such under Gaps.
Literature tools
When a question is bibliographic (finding, reading or checking papers), use the literature tools this harness exposes before generic web search: literature MCP servers (Scite, Semantic Scholar, publisher servers such as Wiley, and the like) and skills or APIs such as OpenAlex. Check which are available at the start of every run, since they differ by harness, and name them in each subagent's brief.
With those tools:
- search with the domain's own vocabulary and several queries per question;
- save every search run under
searches/, the tool's output exactly as returned; - cite only papers the tools actually returned, by DOI;
- check retractions and corrections where the tool reports them.
If no literature tool is available, use web search on primary sources and say so under Gaps.
Look up
Every question starts here. Read index.md and open the pages that look relevant. When a page already answers the question, report it with its link, and stop: nothing is searched twice. When a page answers it in part, the research below extends that page.
Research
- Collect the questions the look-up left open.
- Start one research subagent per question, all at once, in the background, each with a fresh context window. They work in this checkout, on this branch, and leave git to you. Each writes only its own
searches/<date>-<slug>/folders, so they never collide. Brief each with context pointers, not copies: the question and slug,.science/SETUP.md, the library layout above, any existing page it extends, and the literature tools available. Each reports back a draft answer with its citations. - Keep working while they read. The caller goes on with whatever does not depend on the answer.
- Check each report: every claim cited, every search saved as returned, every data source the answer points to named with its release and licence. Send anything missing back to the same subagent.
- Write the page yourself. You alone write pages,
index.mdandlog.md, so parallel subagents and parallel studies never collide on the shared files. Write or extend the page, add or update its index line, append the log entries. - Commit the library:
git add .science/library, thengit commit -m "library: <slug>" -- .science/library(see Commits and tags inlab-notebook).
Done when every question has a committed page and an index line. Report each page to the caller, one line with its link. The caller links it from its study: a Background line in STUDY.md, a probe's Builds on, or a line under the notebook's Library.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

