Use when creating a new ultra-* skill, editing the body of an existing ultra-* skill, splitting or renaming an ultra-* skill, or when ultra-planner surfaces a phase that has no matching sub-skill yet. NOT for authoring superpowers:* skills (use superpowers:writing-skills directly) and NOT for narrative-only edits to DESIGN.md or BUILD_PLAN.md.
Authors and edits ultra-* skills under TDD-for-skills. Extends superpowers:writing-skills with orchestration-specific pressure scenarios, planner dispatch-table fit checks, and the ecosystem updates an ultra skill must land with (catalog, README, BUILD_PLAN, planner dispatch table).
Core principle: A new ultra-* skill is a test-driven artifact inside a coupled ecosystem. If you did not watch a subagent fail without it, you cannot know what it should teach. If you did not update the planner dispatch table, catalog, README, and BUILD_PLAN, it is orphaned.
REQUIRED BACKGROUND: Invoke superpowers:writing-skills and superpowers:test-driven-development via the Skill tool at session start. Reading them as reference is insufficient.
| Signal | Use? |
|---|---|
Creating a new ultra-<name>/SKILL.md | Yes |
| Editing any existing ultra-* SKILL.md body | Yes |
| ultra-planner needs a phase-skill that doesn't exist yet |
| Yes |
| Splitting an ultra skill that grew past ~750 words | Yes |
| Renaming or rescoping an ultra-* skill | Yes |
| Writing a superpowers:* skill (not ultra-*) | No — use superpowers:writing-skills directly |
| Editing DESIGN.md / BUILD_PLAN.md narrative only (no skill change) | No |
Operate in order. Do not batch or skip.
Leader-only steps: 5, 9, 10 require dispatching subagents — only the main session can do this in Claude Code. If invoked as a subagent, describe the scenarios + expected rationalizations + verification test concretely in BUILD_PLAN.md and defer dispatch back to the leader.
Load required skills. Invoke superpowers:writing-skills and superpowers:test-driven-development via the Skill tool. Confirm both loaded.
Verify naming. Must match ultra-<verb-ing> (e.g. ultra-decomposing) or ultra-<noun-phrase> (e.g. ultra-planner). Reject vague nouns and non-ultra prefixes. Record in BUILD_PLAN.md before drafting.
Verify dispatch-table fit. Open ultra-planner/SKILL.md. Identify which phase the new skill serves. If none fits, either rescope or queue a planner update as a P0 interview item and stop.
Design a pressure scenario. Orchestration skill → must trigger checkpoint surfacing, multi-session resumability, or dispatch decisions. Pattern skill → force recognition under ambiguity. Reference skill → force retrieval + correct application. Document in BUILD_PLAN.md.
Dispatch RED baseline subagent. (LEADER-ONLY) Fresh subagent, scenario, WITHOUT the skill loaded. Never draft SKILL.md before this runs. Outputs → /tmp/ultra-<name>-red/.
Analyze baseline. Document rationalizations, dropped steps, and gaps verbatim in BUILD_PLAN.md. These are the tests the SKILL.md must pass.
Draft minimal SKILL.md addressing those gaps only. Match ultra-planner/ultra-decomposing style: Overview, When to Use table, numbered Procedure, Red Flags, Common Mistakes. Target 500-750 words body for orchestrators; <400 for pattern skills. No example code or templates.
Verify CSO compliance. Description states triggering conditions only — NO workflow summary. Start with "Use when…". Under ~500 chars.
Dispatch GREEN verification subagent. (LEADER-ONLY) Fresh subagent, same scenario, WITH the new SKILL.md loaded as a skill (not read as a file). Outputs → /tmp/ultra-<name>-green/.
Refactor loopholes. (LEADER-ONLY for re-running GREEN.) If GREEN surfaces new rationalizations, add explicit counters to Red Flags / Common Mistakes and re-run. Iterate until bulletproof.
Update the ecosystem (all four, same commit): DESIGN.md catalog row, BUILD_PLAN.md test-results table, README.md status row, and ultra-planner dispatch table if dispatchable.
ultra-<verb-ing> or ultra-<noun-phrase>[ -d ~/src/ultra-skills ] && printf '\n---\n*Dogfooding: patch this skill in place when you find gaps.*\n'