phaselisted

# Phase One keyword, two or three sentences, and the skills that already exist for it. ## Why this exists `[measured 2026-08-25]` Across **558 transcripts over seven days**, **4 of this plugin's 45 user-invocable skills fired at all.** The other 41 fired zero times. The split matters more than the total. Of those four, exactly **one** was reached by the model choosing it (`rule-diagnosis`, once). The other three were reached only because a person typed a slash command: `brain` 17 times, `audit` 11, `auto-brain` 3. A few skills are reachable by hand; the model-initiated channel is effectively dead. Those are different problems with different fixes. An earlier version of this paragraph said "exactly one of 37 invocable skills". That read only the Skill-tool channel and missed slash commands entirely, which is the same defect this file is about: a measurement that answers a neighbouring question and gets reported as the whole picture. They are not bad skills. They are unreachable ones: a skill nobody remembers to invoke is indistinguishable from a skill that does not exist, and the same sessions spent that week re-deriving things those skills already encode. The fix is not more skills. It is a **cheap entry point**: a word you can say without deciding anything, which then names what matters now. ## The phases Each is: the two or three sentences that actually change the work, then the skills that already exist for it. Read the sentences; invoke the skills only if the wor