llm-wiki-pm
AGENTS.md7 starsChanged 9 days agoMITView on GitHub
Behavioral rules for all agents. Operation details live in skills/llm-wiki-pm/SKILL.md and sub-skill files. This file governs how agents behave, not what they do. Always resolve this before any wiki operation. Never hardcode paths. If none of these resolve to a valid directory, surface: "No wiki path configured. Run /llm-wiki-pm:set-wiki-path ~/path/to/wiki." Before any write (ingest, update, archive, supersede), complete the core skill's Orient steps ①-④ (read SCHEMA.md, index.md, recent log.md, overview.md) in the current session. Narrow read-only queries may skip it.…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 32,671 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository6 other files it gives its agents.
- llm-wiki-briefSkill
- llm-wiki-maintainSkill
- llm-wiki-pmSkill
- llm-wiki-researchSkill
- llm-wiki-reviewSkill
- set-wiki-pathSkill
The file
We split this AGENTS.md into the rules it gives an agent
4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
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 14
Always resolve this before any wiki operation.
- Distinctive: line 18
Before any write (ingest, update, archive, supersede), complete the core skill's Orient steps ①-④ (read SCHEMA.md, index.md, recent log.md, …
- Distinctive: line 22
Every non-obvious factual claim in wiki pages must carry an inline source marker:
- Very specific: line 40
No silent writes: Every write is snapshotted (the pre-write hook), logged in log.md, and reported to the user.
≈ “Common” is an estimate: at least 1 in 5 of the 41,581 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
LLM Wiki PM — Universal Agent Contract
Behavioral rules for all agents. Operation details live in skills/llm-wiki-pm/SKILL.md and sub-skill files. This file governs how agents behave, not what they do.
Wiki Path Resolution
WIKI=$(cat .wiki-path 2>/dev/null | tr -d '[:space:]')
CONF='${user_config.wiki_path}'; case "$CONF" in '${'*) CONF= ;; esac # plugin option, if set
WIKI=${WIKI:-${CONF:-${WIKI_PATH:-}}}
[ -z "$WIKI" ] && [ -f SCHEMA.md ] && WIKI=$(pwd) # cwd only if it is already a wikiAlways resolve this before any wiki operation. Never hardcode paths. If none of these resolve to a valid directory, surface: "No wiki path configured. Run /llm-wiki-pm:set-wiki-path ~/path/to/wiki."
Orient Protocol (mandatory before writes)
Before any write (ingest, update, archive, supersede), complete the core skill's Orient steps ①-④ (read SCHEMA.md, index.md, recent log.md, overview.md) in the current session. Narrow read-only queries may skip it. Full detail: core llm-wiki-pm SKILL.md → "Orient Every Session".
Source Attribution
Every non-obvious factual claim in wiki pages must carry an inline source marker:
[source: raw-slug, location]
Where location is a page number, section name, or timestamp (e.g. [source: gartner-mq-2026, p.12]).
The page's frontmatter sources: field lists all sources for the page. Inline markers anchor individual claims. Both are mandatory — frontmatter alone doesn't anchor which source backs which claim.
Core Operations Summary
Operation procedures (Ingest, Query, Update, Lint, Crystallize) live in the core llm-wiki-pm SKILL.md → "Core Operations". This file governs only the cross-cutting behavioral constraints below.
Behavioral Constraints
Maintain by default; report after. Keep the wiki current without asking. Add or correct a fact that a primary source supports (read this session, or stated by the user), fix a stale claim and its variants on other pages, file decisions and open questions, create pages that meet the thresholds, and repair links and the index. After the answer, list every page changed: what changed and the source. Ask first only when a write would change a stated strategy, bet or earlier decision; contradict something the user said, with no primary source to settle it; delete, archive or supersede a page; touch 10+ pages; or add shareable: true. Never write a claim whose only support is wiki prose or memory; record it as a gap instead.
Review before reporting. Every answer or page built from the wiki goes through the llm-wiki-review sub-skill (a separate reviewer checks each statement against its sources and searches for what is missing) before the user sees it. The review-gate Stop hook enforces this; the user can say "no review" to skip it for one turn.
No silent writes. Every write is snapshotted (the pre-write hook), logged in log.md, and reported to the user.
No schema drift. New tags require a SCHEMA.md update before use. Tags not in taxonomy are a lint error.
No orphan pages. Min 2 outbound [[wikilinks]] per page. After creating a page, add inbound links from related pages.
Snapshot before destructive ops. Copy to _archive/<slug>-<YYYY-MM-DD>.md before overwrite, archive, or supersede.
Verify writes. Re-read after writing. If frontmatter is malformed or write failed, do not update index.md or log.md.
No raw/ mutations. Layer 1 sources are immutable. Corrections live in wiki pages.
Privacy by default (allowlist). Every page is private and excluded from exports unless it carries shareable: true. Do NOT add a private: flag — it is noise the tooling never reads. Leave customer names, deal sizes, and 1:1 content unflagged (they stay private automatically); add shareable: true only to genuinely public-safe pages.
Supersede explicitly. Materially replacing a page requires supersedes: / superseded_by: frontmatter + archive + link rewrite. Silent replacement is not allowed.
Confirm mass updates. 10+ pages touched → show list, get user sign-off before writing.
Dedup ambient captures. Grep the wiki before adding a fact. If it already exists, update that page; don't create a new one.
Model Routing
| Task | Model |
| File reads, grep, web fetches, writing pages, log appends | Sonnet (I/O worker) |
| Synthesis, crystallize, coverage audit, conflict resolution | Opus (reasoning) |
Workers write output ≥ 2K tokens to /tmp/{task}-{YYYYMMDD}.md and return short status + path.
Skill Architecture
The plugin uses a modular sub-skill design. Each sub-skill is a separate SKILL.md with
scoped when_to_use triggers. The core skill (llm-wiki-pm) serves as the fallback for
any operation not handled by an installed sub-skill.
| Sub-skill | Handles | Core fallback |
llm-wiki-brief | Daily/weekly briefs, tag digests, coverage brief | §10, §11, §12 |
llm-wiki-research | Research sprints, competitive deep dives, stub enrichment | — |
llm-wiki-maintain | Daily-maintenance loop (sweep all sources → ingest → brief → tidy), interactive/autonomous modes | run steps manually |
Sub-skills are additive. Install only what you need. The core skill works standalone.
Agent Surfaces
This plugin ships Claude Code support:
- Core skill:
skills/llm-wiki-pm/SKILL.md - Sub-skills:
skills/llm-wiki-{brief,research,maintain}/SKILL.md - Hooks:
hooks/(session start, prompt recall, write guards, session end) - Wiki search: bundled obsidian-mcp (
wiki-searchMCP server)
This file is the portable behavioral contract for any future agent surfaces.
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.

