| name | goal |
| description | Outcome contract then autonomous loop. Use when the user runs /goal, /goal-loop, “capture goal”, “GOAL.md”, “until done”, or the work spans many files/steps and success must be falsifiable. Writes GOAL.md before wandering. /goal-loop is a mode of this skill (OBSERVE→PLAN→ACT→VERIFY→RECORD), not a third always-on skill. Not /loop (session interval ticks). Not /better (quality pass on an existing deliverable). |
/goal — Contract, then loop
Slash: /goal (optional project command file)
Templates: reference.md
Interval ticks (different skill): Cursor loop skill — not this file
Turn an outcome into a falsifiable contract, then (when asked) run until the criteria pass or a hard blocker needs the user.
Say once: “Running /goal.” (or “Running /goal-loop.”)
Modes
| Invocation | Behavior |
|---|
/goal … | Author or update GOAL.md. Stop at plan if asked (plan-only). |
/goal-loop | After a contract exists (or write it first): OBSERVE → PLAN → ACT → VERIFY → RECORD until criteria pass or blocked. |
stop goal / pause loop | Halt after the current cycle. Leave GOAL.md accurate. |
When to use
- Outcome spans many files/steps or more than one session
- User says
/goal, “track this goal”, or wants autonomous completion criteria
- Ambiguous “build X” where success must be checkable
Skip: one-line mechanical edits, a named /better pass on an existing artifact, /scale volume (use quality-at-scale).
GOAL.md contract (minimum)
Create or update GOAL.md at the project-agreed path (default: repo root GOAL.md). Fields:
- Objective — one clear outcome sentence
- Success criteria — measurable checks (commands, UI states, tests)
- Out of scope — explicit non-goals
- Constraints — stack, safety, no-touch files, deadlines
- Evidence plan — how each criterion will be proven
- Current status —
not started | in progress | blocked | done
- Next action — single next step only
Do not mark done until success criteria are verified with fresh evidence.
Template: reference.md.
/goal-loop cycle
- OBSERVE — Read
GOAL.md, repo state, errors, prior evidence. No edits yet if state is unclear.
- PLAN — Smallest next action that advances a success criterion. Max-depth reasoning.
- ACT — Execute only that action. Smallest safe diff.
- VERIFY — Project verify path; capture exit codes / logs / screenshots.
- RECORD — Update
GOAL.md status, evidence, blockers, and the new Next action.
Loop rules
- Continue until all success criteria pass or a hard blocker needs the user
- On failure: diagnose, adjust plan; do not thrash the same failed approach >2 times without a new hypothesis
- On blocker: stop; report exact blocker + hypotheses + what was tried
- Never claim done without VERIFY evidence
- Do not spawn a second parent orchestrator. Delegate by skill (
review-results, workers) when the procedure says so.
Compose
| Need | Use |
|---|
| Quality pass on current artifact | /better |
| Recurring interval in this session | /loop (skills-cursor/loop) — not this skill |
| Parallel slices | /swarm — parent synthesizes and verifies once |
| Track B / high-signal OSS | addvalue after the contract names the target |
| Sensory ship | /reviewresults (review-results) |
Must not
- Start
/goal-loop with no success criteria
- Mark
done without VERIFY evidence
- Duplicate
/loop processes for the same purpose
- Swarm dependent steps onto the same mutable files
- Expand risk (push, secrets, paid APIs) without user rules
- Paste this skill into every turn as always-on text — Read it when
/goal fires