agentleFS
Sign inStart free
udecodeudecode/better-convex/.agents/skills/research-wiki/SKILL.md

research-wiki

Skill452 starsChanged 23 days agoApache-2.0View on GitHub

Operate the docs/research compiled layer in either full or maintain mode. Use for research-lane creation, research upkeep, authority-sensitive corpus work, source summaries, concepts, systems, decisions, and open-question pages.

We split this skill into the rules it gives an agent

27 rules. 0 are common, 15 are distinctive, and 12 are very specific to this project.

  • 0

    Common

    Most of the 8 edited versions we read keep these.

  • 15

    Distinctive

    This author's own choices. Worth a look.

  • 12

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use them as sub-workflows, not as dead docs.

    line 32
  2. Distinctive:

    Keep the workflow lossless relative to the support docs above.

    line 37
  3. Distinctive:

    Update the research layer itself, not just the chat answer.

    line 38
  4. Distinctive:

    If the topic is authority-sensitive and spans multiple likely corpora, do a

    line 39
  5. Very specific:

    If evidence is thin or contradictory, say so explicitly and create an open-questions/ page instead of flattening uncertainty into fake law.

    line 62
  6. Distinctive:

    create a new research lane from scratch

    line 96
  7. Very specific:

    When full is chosen for a surface that obviously touches multiple editors or source families, do not stop after finding one useful source.

    line 107
  8. Distinctive:

    Do the full corpus-level pass:

    line 110
  9. Distinctive:

    run an official-source discovery step per corpus

    line 115
  10. Distinctive:

    create a per-corpus evidence ledger

    line 117
  11. Very specific:

    update index.md and log.md

    line 120
  12. Distinctive:

    Do not end a full pass with silent corpora.

    line 153
  13. Distinctive:

    If a likely corpus was scoped and produced no useful proof, say that directly.

    line 155
  14. Very specific:

    Before recording raw gap, use official-source discovery to confirm that the missing coverage is real and not just a stale or incomplete …

    line 158
  15. Very specific:

    Before recording evidence gap, read the strongest local raw hits that your searches surfaced for that corpus.

    line 161
  16. Very specific:

    If raw evidence is missing for a scoped corpus after that discovery step, record a raw gap explicitly instead of quietly letting another …

    line 164
  17. Very specific:

    If compiled pages are missing but raw exists, record a compile gap and create the missing source summary when safe.

    line 168
  18. Very specific:

    If evidence exists but still does not answer the behavior question, record an evidence gap or contradiction gap instead of flattening it …

    line 171
  19. Distinctive:

    If official discovery and local raw search already surfaced an obviously relevant file, you do not get to stop at "broad grep over corpus …

    line 175
  20. Distinctive:

    Do not do “research” that only rephrases the first source you found.

    line 193
  21. Very specific:

    Do not call something full if you only spot-checked one corpus when the topic obviously spans several.

    line 194
  22. Very specific:

    Do not close full mode without an official-source discovery step for each scoped corpus.

    line 196
  23. Distinctive:

    Do not treat grep hits as the same thing as inspected evidence.

    line 198
  24. Distinctive:

    Do not record a corpus as thin, partial, or unresolved when the likely answer is already sitting in local raw unread.

    line 199
  25. Very specific:

    Do not close full mode without a per-corpus evidence ledger.

    line 201
  26. Distinctive:

    Do not bypass the compiled layer and patch law straight from raw evidence unless the task is tiny and the evidence is already obvious.

    line 202
  27. Very specific:

    Do not leave index.md or log.md stale after adding research pages.

    line 204

“Common” is measured: kept by at least 6 in 10 of the 8 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Research Wiki

Handle $ARGUMENTS. This is the reusable workflow for operating the docs/research compiled layer without turning research into one-off chat mush.

<task>#$ARGUMENTS</task>

Purpose

Turn the existing research command docs into one reusable skill entrypoint.

This skill should decide whether the job is:

  • a full ingest-and-compile pass
  • a maintain pass over an existing research lane

The support docs remain canonical detail:

  • docs/research/commands/full-pipeline.md
  • docs/research/commands/maintain.md

Use them as sub-workflows, not as dead docs.

Core Rules

  • docs/research is the compiled layer; ../raw is the evidence layer.
  • Keep the workflow lossless relative to the support docs above.
  • Update the research layer itself, not just the chat answer.
  • If the topic is authority-sensitive and spans multiple likely corpora, do a full corpus-level ingest across the relevant source families instead of a narrow spot check.
  • In full mode, verify corpus completeness via an official-source discovery step before declaring a corpus missing, thin, or only partially evidenced.
  • In full mode, once official-source discovery confirms that the relevant raw family already exists locally, exhaust the strongest local raw hits before calling evidence gap or contradiction gap.
  • In full mode, every scoped corpus must end in an explicit disposition:
    • strongest evidence found
    • raw gap
    • compile gap
    • synthesis gap
    • freshness gap
    • evidence gap
    • contradiction gap
    • structure gap A full pass is not complete until every scoped corpus has one.
  • Do not confuse:
    • raw evidence
    • compiled source summaries
    • synthesized decisions
    • open questions
  • If evidence is thin or contradictory, say so explicitly and create an open-questions/ page instead of flattening uncertainty into fake law.

Read First

Always read:

  1. docs/research/README.md
  2. docs/research/index.md
  3. docs/research/log.md

Then choose one support doc:

  • docs/research/commands/full-pipeline.md for full
  • docs/research/commands/maintain.md for maintain

Mode Selection

Use full when:

  • the topic is broad or thinly covered
  • the authority question is large enough that trusting the visible local slice would be reckless
  • raw evidence is missing, stale, or uneven
  • the surface spans multiple likely corpora
  • trigger behavior, input behavior, or product claims need a full corpus-level ingest

Examples:

  • compare Typora / Obsidian / Milkdown on one under-researched trigger surface
  • rebuild a weak authority lane
  • create a new research lane from scratch

Use maintain when:

  • the lane already exists in docs/research
  • the goal is contradiction cleanup, freshness, backlinks, or stronger compiled synthesis
  • you need upkeep, not a new ingest push

Full-Mode Requirement

When full is chosen for a surface that obviously touches multiple editors or source families, do not stop after finding one useful source.

Do the full corpus-level pass:

  1. scope the relevant corpora
  2. inspect existing compiled coverage
  3. inspect raw coverage across the relevant corpora
  4. run an official-source discovery step per corpus
  5. exhaust the strongest local raw hits per corpus
  6. create a per-corpus evidence ledger
  7. classify gaps per corpus
  8. fill what is safe
  9. update index.md and log.md
  10. return the remaining gaps honestly

For editor-behavior authority work, that often means some subset of:

  • Typora
  • Obsidian
  • Milkdown
  • Google Docs / GitHub Docs / Notion

depending on the actual surface

Per-Corpus Evidence Ledger

For full mode, explicitly close each scoped corpus with:

  • corpus name
  • compiled pages inspected
  • raw paths inspected
  • direct raw files actually read
  • official source entrypoints checked
  • strongest evidence found
  • disposition:
    • evidenced
    • raw gap
    • compile gap
    • synthesis gap
    • freshness gap
    • evidence gap
    • contradiction gap
    • structure gap
  • next action if still unresolved

Do not end a full pass with silent corpora.

If a likely corpus was scoped and produced no useful proof, say that directly. That is still a result.

Before recording raw gap, use official-source discovery to confirm that the missing coverage is real and not just a stale or incomplete local raw mirror.

Before recording evidence gap, read the strongest local raw hits that your searches surfaced for that corpus. Broad grep is routing, not proof.

If raw evidence is missing for a scoped corpus after that discovery step, record a raw gap explicitly instead of quietly letting another corpus carry the whole answer.

If compiled pages are missing but raw exists, record a compile gap and create the missing source summary when safe.

If evidence exists but still does not answer the behavior question, record an evidence gap or contradiction gap instead of flattening it into a partial yes.

If official discovery and local raw search already surfaced an obviously relevant file, you do not get to stop at "broad grep over corpus X". Read the file and cite it, or explain why it was not actually relevant.

Outputs

A successful pass should leave behind durable artifacts such as:

  • source summaries
  • entity pages
  • concept pages
  • system pages
  • decision pages
  • open-question pages
  • updated index/log entrypoints

Blunt Rules

  • Do not do “research” that only rephrases the first source you found.
  • Do not call something full if you only spot-checked one corpus when the topic obviously spans several.
  • Do not close full mode without an official-source discovery step for each scoped corpus.
  • Do not treat grep hits as the same thing as inspected evidence.
  • Do not record a corpus as thin, partial, or unresolved when the likely answer is already sitting in local raw unread.
  • Do not close full mode without a per-corpus evidence ledger.
  • Do not bypass the compiled layer and patch law straight from raw evidence unless the task is tiny and the evidence is already obvious.
  • Do not leave index.md or log.md stale after adding research pages.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

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.

Check your setup, free