Claude
d28dfdc39b
Bare `/skill`-style prose ("run the `/grilling` skill") relies on the model
reading a slash-prefixed name as a command to fire. Naming the tool directly
gets a higher hit rate, and it's actually more harness-neutral than the old
convention since it drops the Claude-Code-specific `/` trigger syntax
entirely rather than assuming it. Updates the documented convention in
.agents/invocation.md and every operative cross-skill invocation site in
skills/**/SKILL.md to match.
Co-Authored-By: Claude <noreply@anthropic.com>
3.1 KiB
Model-invoked vs user-invoked
Every SKILL.md in this repo is a skill. The one axis that splits them is invocation — who can reach it:
- User-invoked — reachable only by the human typing its name. Set
disable-model-invocation: truein the frontmatter (Claude Code) andpolicy.allow_implicit_invocation: falseinagents/openai.yaml(Codex). Thedescriptionis human-facing: a one-line summary read by a person browsing slash-commands. Strip trigger lists ("Use when the user says…"). - Model-invoked — reachable by model or user. The default: omit
disable-model-invocationand thepolicyblock fromagents/openai.yaml. Thedescriptionis model-facing and keeps rich trigger phrasing ("Use when the user wants…, mentions…, asks for…") so auto-invocation fires. The test for whether a skill should stay model-invoked: could the model usefully reach for this autonomously? (Reuse is the reason to extract a skill, not the test.)
Each harness excludes a user-invoked skill from the model's reach in its own way, so nothing but the human can fire it — no other skill can. A user-invoked skill may invoke model-invoked skills, but it can never reach another user-invoked skill.
Every skill also carries an agents/openai.yaml beside its SKILL.md. It holds Codex UI metadata — interface.display_name and interface.short_description for the skill picker — and, for user-invoked skills, the policy.allow_implicit_invocation: false that pairs with disable-model-invocation. Keep the two in sync: a skill is user-invoked in both harnesses or neither.
Bucket README.mds and the top-level README.md group entries into User-invoked and Model-invoked.
Dependencies between them
Dependencies are expressed as an explicit instruction to call the Skill tool for the named skill ("Call the Skill tool for grilling"), not deep ../other-skill/FILE.md cross-references, and not a bare /skill-style mention left for the model to interpret. Naming the tool is what gets it fired: most harnesses expose skill invocation as a tool the model calls, and spelling that out gets a higher hit rate than dropping a /name into prose and hoping it's read as a command. Dropping the leading / also keeps this harness-neutral rather than less — a skill name on its own carries no assumption about which harness's trigger syntax it belongs to. Shared reference docs live inside the skill that owns them; other skills reach that material by calling the Skill tool for it, not by linking across folders.
This is about operative instructions — a skill's own steps telling the agent to go run another skill right now. Router prose that just names skills for a human to pick from (ask-matt, bucket README.mds) isn't invoking anything, so it keeps /skill-style names as plain labels.
Passive vs active domain work
Merely reading CONTEXT.md for vocabulary is a one-line prose pointer, not the domain-modeling skill. Only the active build/sharpen discipline (challenge terms, edge-case scenarios, write ADRs, update CONTEXT.md inline) is domain-modeling.