skills
AGENTS.md11 starsChanged 14 days agoMITView on GitHub
Public, Markdown-first skills library. CLAUDE.md is a symlink to this file. This repo owns reusable Skill implementations only. Device inventory, consumer enablement and routing, automations, personal memory, and install evidence belong to private oldwinter/general-tasks; machine and harness config belongs to private oldwinter/dotfiles. Every skill is exactly one of two kinds, and the split is mechanical so a new skill cannot drift: Both markers always agree — a skill is user-invoked in every harness or in none — and just check…
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 repository19 other files it gives its agents.
- obsidian-skills/AGENTS.mdAGENTS.md
- add-just-doctorSkill
- agent-autoresearchSkill
- agent-harness-repairSkill
- change-evidenceSkill
- ci-speedSkill
- verifySkill
- deck-style-gallerySkill
- doneSkill
- longform-fictionSkill
- loop-kitSkill
- loop-nightly-reviewSkill
- loop-triage-githubSkill
- loop-watch-socialSkill
- orch-claude-pi-codexSkill
- skills-doctorSkill
- tdd-pstackSkill
- teach-pstackSkill
- windows-leftover-cleanupSkill
The file
We split this AGENTS.md into the rules it gives an agent
2 rules. About 1 is common, 1 is distinctive, and 0 are very specific to this project.
≈1
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Common: line 16
Keep every skill usable outside this owner's machines: no credentials, private sessions, personal memory, private hostnames, or hardcoded …
- Distinctive: line 22
Every skill is exactly one of two kinds, and the split is mechanical so a new skill cannot drift:
≈ “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
AGENTS.md
Public, Markdown-first skills library. CLAUDE.md is a symlink to this file.
This repo owns reusable Skill implementations only. Device inventory, consumer enablement and routing, automations, personal memory, and install evidence belong to private oldwinter/general-tasks; machine and harness config belongs to private oldwinter/dotfiles.
Layout
- A skill is a directory named by its identifier (hyphen-case) with a required
SKILL.mdand optionalreferences/,rules/,scripts/,agents/(harness metadata),tests/,examples/; forked skills also carryLICENSEandNOTICE.md. - Skills live at the repo root, except Obsidian ones under
obsidian-skills/. Only skills in use stay onmain; third-party skills are installed from their upstream, not copied here. - Retired skills are removed from
mainand listed under README → 归档.archive/2026-10-13is the pre-cleanup snapshot tag; anything retired later lives in git history. Restore withgit checkout <ref> -- <path>. - Adding, removing, or renaming a skill means updating the README list;
just checkenforces it.
Skill rules
- Keep every skill usable outside this owner's machines: no credentials, private sessions, personal memory, private hostnames, or hardcoded device paths; use placeholders or public interfaces in examples.
SKILL.mdfrontmatter is parsed by scripts:---delimited,name:anddescription:as single-line scalars (no>block scalars), no angle brackets indescription. A description containing:must be quoted: an unquoted one is invalid YAML, and the install CLI skips that skill silently, so the install looks like it worked while the skill is missing.agents/openai.yaml, when present, advertises the display name, default prompt, and implicit-invocation policy to harnesses that read it. The checker validates it when present:interface.display_name/interface.short_description/interface.default_prompt(which must invoke$<skill-name>) and a booleanpolicy.allow_implicit_invocationthat must agree with the SKILL.md marker (see Skill invocation).
Skill invocation
Every skill is exactly one of two kinds, and the split is mechanical so a new skill cannot drift:
- 自动引用 / model-invoked — the model may fire it on its own. Omit
disable-model-invocationfromSKILL.mdfrontmatter and setpolicy.allow_implicit_invocation: trueinagents/openai.yaml. - 手动引用 / user-invoked — only the human can reach it, by typing
/nameor$name; no other skill can call it. Setdisable-model-invocation: trueinSKILL.mdfrontmatter andpolicy.allow_implicit_invocation: falseinagents/openai.yaml.
Both markers always agree — a skill is user-invoked in every harness or in none — and just check enforces the pairing, including that a user-invoked skill ships an agents/openai.yaml. This mirrors the mattpocock/skills convention.
Decide the kind when you add or edit a skill, by one test: could the model usefully reach for this autonomously? If yes, it is model-invoked (the default). Mark it user-invoked only when auto-firing would be wrong or unsafe — the skill performs an irreversible or outward-facing action (commit, push, merge, post, spend), starts a long unattended run, or must be a command the human starts on purpose.
The description follows the kind: a model-invoked skill keeps trigger phrasing ("Use when …") so auto-invocation fires; a user-invoked skill reads human-facing — a one-line summary, with the trigger list stripped. Group both in the README under 自动引用 / 手动引用.
Verify
just checkafter any change: the README list and statistics match the tree, frontmatter name/description/resource links are valid, no personal machine paths appear, and the check scripts' fixture tests pass. On Windows run shell checks throughjustor Git Bash; thebashfirst on PATH may be WSL, which does not share the filesystem.just doctorchecks the local toolchain (python3, bash, git, just) and theCLAUDE.mdsymlink;PASS/WARN/FAIL/SKIProws, nonzero exit on anyFAIL. No env file or services exist in this repo, so the doctor's service inventory is intentionally empty..claude/skills/verify/is project-local verification tooling, not a published skill: it drives the real install path (.scratch/workspace, evidence under.scratch/verify-evidence/) and the gates above, with its feature map infeatures/..agents/skills/verifyis a local junction so DeepSeek Harness discovers the same bundle; it is not committed.
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.

