| name | my-colleague-task |
| description | REPLACE THIS — one-line trigger summary that tells Claude when to load this skill. Write it like a tool docstring, not a title. Concrete trigger phrases beat abstract descriptions. Example "Triggers when the user asks about <specific recurring task in your domain>; expects <input shape>; produces <output shape>." |
My Colleague Task
Fork this template, change four things, ship it. See the bottom of this file for the "Things to change" checklist.
When to invoke
Describe the specific situations where this skill should fire. The clearer this is, the better Claude routes to it.
Concrete trigger examples (rewrite for your task):
- "User pastes a stack trace and asks 'why does this fail'"
- "User asks to triage a Jira ticket"
- "User asks to draft a PR description"
- "User mentions one of:
<list of keywords specific to your task>"
Inputs the skill needs
Before producing output, gather these inputs (ask the user if missing):
- [input 1] — what kind of artifact, where it lives
- [input 2] — context the skill needs to do its job well
- [input 3] — optional but helps
Walkthrough — what the skill actually does
This is the steps the skill follows. Write it as instructions to Claude, not as documentation.
Example for a code-review skill:
- Read the diff fully before commenting
- Find the highest-risk change — the one that, if wrong, breaks production
- Check that change against
reference.md patterns
- If anything matches a known anti-pattern, cite it
- Default to silence on style issues; only flag substance
- Output: 1-3 bullets, each citing a file:line
Output format
Describe what the skill produces. Show an example.
## Review summary
**Highest risk:** path/to/file.ts:42 — comment explaining the risk
**Other notes:**
- path/to/other.ts:99 — secondary issue
What this skill does NOT do
Naming non-goals makes the skill sharper.
- Doesn't refactor (just reviews)
- Doesn't run tests (just reads code)
- Doesn't engage with style preferences (substance only)
Linked resources (progressive disclosure)
This skill bundles reference.md with deeper anti-pattern catalog. Claude loads it on demand when reviewing a substantial diff. Keep these references focused — under 500 lines each. Long references defeat the progressive-disclosure design.
See also
reference.md — anti-pattern catalog
scripts/check.sh — optional verification helper (loaded on demand)
Things to change before shipping
Verification (after install)
In your colleague session:
> tell me what skills you have
Your skill should appear by name. Then test:
> [type a trigger phrase from your "When to invoke" section]
Does the colleague route to this skill? If not — the description isn't routing well. Tune it.
Why this works (the canonical framing)
Per Bayram (Webinar 3, line 515-527):
"Скиллы — претендент на повторяющийся путь экономии энергии."
A skill is a repeating pattern crystallized into a callable. The test: did you copy-paste this between sessions before? If yes → skill candidate. If you only did it once → keep it in CLAUDE.md.
Cross-runtime note: this same SKILL.md works in Claude Code (.claude/skills/), Codex CLI (.codex/skills/), Anthropic Developer Platform, and ChatGPT. Skills are the portable layer of harness engineering.