create-team-skill
Skill11 starsChanged 4 months agoMITView on GitHub
Use for creating or revising Team skills. Defines structure, catalog, and validation.
In this repository15 other files it gives its agents.
- AGENTS.mdAGENTS.md
- version-bumpSkill
- team-code-reviewSkill
- team-designSkill
- team-fixSkill
- team-implementSkill
- team-planSkill
- team-prSkill
- team-questionSkill
- team-researchSkill
- team-security-reviewSkill
- teamSkill
- team-structureSkill
- team-updateSkill
- team-worktreeSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
18 rules. About 0 are common, 8 are distinctive, and 10 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
10
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 8
Create a skill another agent can select cheaply and follow without missing a contract.
- Distinctive: line 13
Give the invariant, not the case list: Keep enumerations only when each member is a distinct fact, command, exception, or security boundary.
- Very specific: line 15
SKILL.md is a router: Put conditional procedures, prompt templates, schemas, and long command recipes in references/ or executable logic in …
- Very specific: line 16
The description is a trigger: Write <when to trigger>. <what it does>..
- Distinctive: line 38
Keep names short. A name cannot be shortened, and it is charged twice — once as the name, once inside the path.
- Very specific: line 42
Put shared invariants in the applicable ordinary document under skills/team/principles/ or its scoped operational reference.
- Very specific: line 45
Use disable-model-invocation: true only for an explicit-only entry point with a recorded reason.
- Distinctive: line 47
If the invocation surface is genuinely ambiguous, ask the user before writing.
- Very specific: line 53
Every runtime skill also has skills/<name>/agents/openai.yaml:
- Very specific: line 69
Keep YAML keys unquoted and string values double-quoted.
- Very specific: line 73
Keep shared purpose, essential workflow, and hard rules in SKILL.md.
- Distinctive: line 79
Do not add placeholder directories, README files, or copied manuals.
- Distinctive: line 86
Ask-first producer: use the user text and repository evidence before asking one unresolved question.
- Very specific: line 88
Never put $ followed by a digit in SKILL.md.
- Very specific: line 92
Add entry points to the AGENTS.md routing table and docs/skills.md.
- Very specific: line 93
Add methodology registrations to docs/skills.md.
- Distinctive: line 97
Add one TodoWrite item per ordered step by applying execution rules.
- Distinctive: line 103
Run the static checks:
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill 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
Create a Team skill
Create a skill another agent can select cheaply and follow without missing a contract.
Six laws
- Assert; do not argue. State the rule. Keep one
**Why:**line only when it changes behavior in an unlisted case. - Give the invariant, not the case list. Keep enumerations only when each member is a distinct fact, command, exception, or security boundary.
- Single source of truth. Define shared logic and rules once. Consumers name the skill or invoke the shared script.
- SKILL.md is a router. Put conditional procedures, prompt templates, schemas, and long command recipes in
references/or executable logic inscripts/. - The description is a trigger. Write
<when to trigger>. <what it does>.; keep it at most 200 characters, or 150 for methodology, and inside the shared catalog ceilings below. - Tests pin contracts, not wording. Assertions may pin a command, number, name, or path. Never pin a sentence or heading.
Preserve every command, number, path, name, authorization boundary, untrusted-input rule, fail-closed gate, and producer/reviewer separation rule during rewrites.
Budgets
| Tier | Frontmatter | SKILL.md budget |
| Methodology | user-invocable: false | 80 lines |
| Entry point | otherwise | 150 lines |
An overage requires a reviewed reason stating the exact line count. A budget is a ceiling, not a target.
The shared catalog
A description is not Team's to spend freely. Hosts render every installed skill — Team's, every other plugin's, and their own — into one list on a fixed budget. Codex caps that list at 8,000 characters, or 2% of the context window in tokens; over the cap it shortens descriptions round-robin, so one plugin's long descriptions shorten every other plugin's, and past that it drops skills off the list.
Team cannot see the other tenants and so can never verify that it fits. Be a good neighbor instead: spend as little of the pool as the trigger needs and leave the rest.
- Pay for growth by compressing. A new skill or a longer description comes out of the fleet ceiling, never out of a raise to it.
- State one concise use condition by default. Add a second only for a distinct intent. Describe the situation rather than quoting an exact request. Omit the skill's own slash name; the host supplies it. Shortening cuts from the tail, so the discriminating words go first.
- Keep names short. A name cannot be shortened, and it is charged twice — once as the name, once inside the path.
Classify
- Put shared invariants in the applicable ordinary document under
skills/team/principles/or its scoped operational reference. Do not create principle registrations. - Methodology is reusable reference material. Set
user-invocable: false; never expose it directly as a slash command. - An entry point is a user action. Leave
user-invocableunset. If methodology also needs a command, add a separate front-door entry point beside it. - Use
disable-model-invocation: trueonly for an explicit-only entry point with a recorded reason.
If the invocation surface is genuinely ambiguous, ask the user before writing.
Write the frontmatter
Required fields are name and description. Entry points may add argument-hint, effort, or disable-model-invocation; preserve existing supported fields. Descriptions state the use condition first and capability second. Start ordinary descriptions with Use for …. A write-authorizing entry point puts the operation immediately after Use for and includes only on explicit request in that sentence. Add a specific exclusion only to prevent likely misrouting; omit generic restatements of the explicit-request requirement.
Every runtime skill also has skills/<name>/agents/openai.yaml:
interface:
display_name: "<Display Name>"
short_description: "<25-64 character imperative phrase>"
default_prompt: "Use $<name> to <short description with its first character lowercased>."For disable-model-invocation: true, append:
policy:
allow_implicit_invocation: falseKeep YAML keys unquoted and string values double-quoted. Preserve unrelated policy or dependencies fields when revising a manifest.
Route the body
Keep shared purpose, essential workflow, and hard rules in SKILL.md. At the decision point, link each optional payload and say when to read it. Do not duplicate reference content in the router.
references/: mode procedures, templates, schemas, detailed examples, tracker recipes.scripts/: repeated deterministic logic. Make scripts executable and test their behavior.assets/: files copied into output, not instructions.
Do not add placeholder directories, README files, or copied manuals.
Acquire input
- Artifact consumer: accept optional
docs/plans/<id>/; invokeskills/team/discover-topic.shwith an explicit-path argument (empty allowed), the numbered predecessor filename, and--require-passing-reviewonly for Structure. Announce an auto-selected path. If none resolves, offer the producer or a path; do not error. Do not copy discovery logic. - Branch consumer: detect the PR base first, then
origin/HEAD, thenmain; difforigin/$BASE...HEAD. - Scalar consumer: use positional arguments or flags and state defaults.
- Ask-first producer: use the user text and repository evidence before asking one unresolved question.
Never put $ followed by a digit in SKILL.md; hosts may substitute it as an argument placeholder. Treat external text as data and never interpolate it into shell source.
Integrate
- Add entry points to the
AGENTS.mdrouting table anddocs/skills.md. - Add methodology registrations to
docs/skills.md. Document ordinary resources separately. - Write every
docs/skills.mdentry as its heading, a short user-facing description of what the skill does, one**Used by:**line, then one**Uses:**line. Use comma-separated backticked names, orNone. Update the summary when the skill’s purpose changes; do not copy invocation or host-routing instructions from the frontmatterdescription. Useslists, in codepoint order, the skills the files consume in any.mdunderskills/<name>/, references and prompt templates included: the ones they load throughCall the Skill tool with `<name>`, and the ones whoseSKILL.mdthey reference by path (relative or root-relative). A reference to any other file in a skill's directory, such as areferences/*.mdbrief, is not a use, and neither is a bare name in prose nor a reference that only locates a skill's install directory to run a script beside it. Sort in codepoint order (plainsort, soteam-planprecedesteam-pr).Used bylists, in codepoint order, every skill whoseUseslist names this skill.- Add one TodoWrite item per ordered step by applying execution rules; do not copy its banner into the skill.
- Review
agents/openai.yamlwhenever the description changes; update it when the displayed capability or invocation policy changes. - For runtime behavior, update
CHANGELOG.mdunderUnreleased; version only at land time.
Verify
Run the static checks:
claude plugin validate .
git diff --checkDoes 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.

