← ClaudeAtlas

testing-skills-with-subagentslisted

TDD applied to skill/process documents — pressure-test a discipline-enforcing skill with subagents BEFORE deploying it, by running realistic multi-pressure scenarios without the skill (RED, capture rationalizations verbatim), with it (GREEN), then closing each loophole (REFACTOR) until agents comply under maximum pressure. Use when creating or editing any gate/discipline skill in this repo — especially before /ms.sync broadcasts the edit to every registered project repo. Skip for pure reference skills with no rules to violate.
beomeodev/specter · ★ 0 · Testing & QA · score 70
Install: claude install-skill beomeodev/specter
<!-- Source: adapted from obra/superpowers `writing-skills/testing-skills-with-subagents` (MIT), trimmed and wired to SPECTER's ms.sync. --> # Testing Skills With Subagents **Testing skills is TDD applied to process documentation.** If you didn't watch an agent fail *without* the skill, you don't know whether the skill prevents the right failures. **SPECTER wiring**: this repo's `ms-*` skills are broadcast to every registered project by `/ms.sync` — a bad revision propagates everywhere at once. Pressure-test any edit to a discipline-enforcing skill (gates, TDD rules, surgical-scope rules, parallel/worktree rules) before syncing. Reference skills (ms-lang-*, command crib sheets) don't need this — nothing in them can be rationalized away. ## The cycle | TDD phase | Skill testing | Success criterion | |---|---|---| | RED | Run scenario WITHOUT the skill | Agent fails; rationalizations documented verbatim | | GREEN | Write/edit skill addressing those exact failures | Agent complies with skill present | | REFACTOR | Close each new loophole | Agent still complies; no new rationalizations | ## RED: baseline first Dispatch a subagent on a realistic task engineered to make it WANT to violate the rule — without the skill. Document its choices and excuses word-for-word. Combine **3+ pressures**: | Pressure | Example | |---|---| | Time | deploy window closing, "5 minutes left" | | Sunk cost | 3 hours / 200 lines already written | | Authority | "the senior said skip it" | | Exhaus