goal

# /goal — autonomous goal loop You prompt yourself a proactive executor: state a GOAL plus how to verify it, and the agent runs check→fix→check on its own, returning ONLY with the finished result. Frees you from sitting at the monitor nudging every step. This instruments four principles (think before code · simplicity · surgical edits · drive through a verifiable target) into a reusable loop. Not a replacement for a heavy decomposing orchestrator — `/goal` is for ONE mid-sized goal. The design phase (goal / verification / control rubrics) is cherry-picked from Looper (ksimback) — the framework wasn't installed, the techniques were. ## When to use - A clear goal with a **verifiable done-criterion**: a bug (test reproduces → green), research (facts verified), content (critic ≥ threshold), audit (findings adversarially checked), refactor (tests + lint green). - NOT for: trivial work (just do it), pure discussion, and ⚠️ NOT for irreversible/production mutations without confirmation (see guards). ## Goal template (fill in BEFORE starting — design phase, don't skip) ``` Goal: <a concrete OUTCOME + the artifact/state that proves completion; NOT "improve X"> Sharpness test: what counts as "done" if two competent agents disagreed? Subjective → measurable proxy. Scope: included <…> / excluded <…> / max depth <…> Context: <path to project/data, symptoms, logs — what to READ, not assume> Checks (typed, programmatic FIRST — see "Verification"): - [programmatic] <test / compile / ...