jayden-dang The Rule
Invoke relevant or requested skills BEFORE any response or action — before clarifying questions, before exploring the codebase, before checking a single file. If the skill turns out not to fit, you may set it aside after reading it.
Announce "Using [skill] to [purpose]", then follow the skill exactly. If it has a checklist, create one todo per item.
Anti-drop (OVERRIDE 2026-09-08). After the entry skill is named, copy its
mandatory sequence (Iron Law / Steps / Done-when chain) into the visible todo
list before acting. Never delete a step silently — write
skip: <reason> on that todo, or keep it open. Lead pressure, standup clocks,
and "skip ceremony" do not erase reproduce / design / prove steps.
Priority: process skills first, then implementation skills. When the
entry point is unclear, load docs/guide/process/on-ramps.md — that table
is the one home; do not invent a second router.
Participant boundary: never infer skill-set membership from roster, CODEOWNERS, or PR authorship; only skill-mediated actions are enforced or recorded. Treat supplied evidence as supplied — do not invent mediation that did not happen.
Handing off
Once the entry point is named, how you reach it depends on the target:
- Model-invocable (no
disable-model-invocationin its frontmatter) — invoke it and let it take over. - User-invoked (
disable-model-invocation: true) — you cannot invoke it. Name it for the user to run: "run/triage". Telling the agent to invoke one is a dead-end hand-off, a real bug rather than a style nit, andscripts/lint-write-handoffs.pyfails it.
Do not start executing a chosen flow from inside this skill. Route, then hand over.
Context hygiene: keep discovery through plan in one unbroken context window.
If the window is filling before the plan is done, tell the user to run
/write-handoff. Execution sessions are context-isolated per task by design.
Red Flags — you are rationalizing
| Thought | Reality |
|---|---|
| "This is just a quick question" | Questions are tasks. Check for skills. |
| "Let me look around first" | Skills define HOW to look. Check first. |
| "This is too small for process" | Small things grow. The tier system handles size — the skill decides, not you. |
| "I remember what that skill says" | Skills change. Read the current text. |
| "I'll do this one step, then check" | The check comes before the first step. |
| "Being helpful means answering fast" | Being helpful means following the process that works. |
Precedence
The user's explicit instructions (including CLAUDE.md) override skills; skills override your defaults. Skip a skill's workflow only when the user has explicitly told you to.