From auto
Turn a plan into a model-judgeable goal doc the user binds with `/goal <doc-path.md>`. Use when the user says "generate a goal for this plan", "make a goal doc", "write a /goal for <plan>", "write a goal doc for this run", or wants auto to phrase an exit predicate from a plan. This skill AUTHORS the goal doc and writes it to disk; it does NOT run `/goal` — binding is the user's action (auto can neither arm nor clear a native goal). The doc is written to TRACK auto's own exit predicate so the model-judged goal and auto's deterministic Stop hook are as aligned as possible.
How this skill is triggered — by the user, by Claude, or both
Slash command
/auto:auto-author-goalThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
A **goal doc** is a saved, model-judgeable exit predicate derived from a plan.
A goal doc is a saved, model-judgeable exit predicate derived from a plan.
The user binds it with /goal <doc-path.md> (a file-referenced goal — the
model judges the file's whole content as the "done" condition). This skill
writes that doc; it never runs /goal.
The user runs /goal .claude/auto/goals/<slug>.md themselves. The model then
judges the entire doc as the goal at each stop attempt. Two consequences
drive how the doc is written:
lib/on-stop.py); a native /goal is
model-judged and shares NO verdict with it. They can DIVERGE — a goal judged
stricter than the ledger keeps re-prompting "Goal not yet met… continuing"
even after auto reaches done (the spam class the Stop-hook fix addressed,
now self-inflicted via the goal). We can't guarantee agreement; we write the
doc to TRACK auto's predicate to minimize divergence. /goal clear is the
escape — name it in the doc. Auto never arms or clears this goal; the user
does. See skills/auto/SKILL.md §1.The criteria the model judges should each be:
/goal — it cannot be
one; recommend a checklist or a test the agent can drive instead.<plan>" and assume the judge opens the plan (that's an unverified second-file
read). Name the plan as provenance, but spell out the concrete done-states the
judge can check from repo/transcript without leaving the doc.tests/run.sh is green"), NOT as separate bullets ANDed on top.
Any independent gate ANDed onto the predicate can only make the goal STRICTER
than the ledger — re-introducing the divergence (re-prompt-after-done) from
the binding model above. If you DO want a stricter gate deliberately (e.g.
keep the session going until tests pass, beyond auto's review-loop), that is a
valid choice — but say so explicitly, because the extra "continuing" prompts
are then INTENDED, not the spam failure mode.docs/plans/*.md). Read it. Pull: the requirements trace (R1…), acceptance
examples (AE1…), scope boundaries, deferrals, and any explicit success /
"done when" criteria./goal and
suggest a checklist instead.<repo>/.claude/auto/goals/<plan-slug>.md using
the template below (create goals/ if absent). .claude/auto/ is gitignored
runtime state — the doc is a local artifact, not a committed file./goal .claude/auto/goals/<slug>.md ← run this to bind/goal clear ← run this to release itThe whole file is the condition — keep it criteria-focused.
# Goal: <plan title>
Source plan: `<relative/path/to/plan.md>`
This goal is met when auto's exit predicate holds: every unit in the plan is
implemented and reviewed, and only P3 review findings remain (P0/P1/P2 all
resolved). Concretely, that state means:
- <inlined acceptance outcome 1 — e.g. "a circle shape can be drawn and persists across reload">
- <inlined acceptance outcome 2 — observable, spelled out, not a citation>
- `tests/run.sh` exits 0.
Out of scope (do NOT hold this goal open on these): <deferred items, and any human-gated/manual-QA steps the agent can't perform itself, or "none">.
<!-- note: model-judged goal, independent of auto's Stop hook; release with `/goal clear`. -->
/goal. Author and surface only; the user binds./goal re-prompts
"continuing" forever. Human-gated steps go out-of-scope; a goal that is ONLY a
manual test must not be authored as a /goal at all./goal clear is the escape.npx claudepluginhub shawnroos/autoGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.