Menu
Reference only (do NOT invoke as an action): the RICE scoring convention for the goal-loop backlog. Read by the loop/make-plan skills.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Reference only (do NOT invoke as an action): token-reduction conventions (subagent isolation, artifact compaction). Applied throughout the other skills.
Implement a single planned issue, optionally inside a bounded autonomous loop. Use to execute one unit of an approved plan.
Run a continuous goal-loop: work a prioritized RICE-scored backlog toward a standing goal, cycle after cycle, consuming/emitting signals. Use for ongoing autonomous improvement toward a goal; not for a single discrete task (use run).
Run the enabled compliance checklists (any of HIPAA / HITRUST / ISO 27001 / SOC 2) against the diff, separating code-verifiable findings from organizational controls. Always run before committing a sensitive change.
Emit or consume cross-loop signals (observations / frictions / opportunities) in the shared signals/ store. Use to record a finding for other loops or teammates, or to read relevant ones before deciding. PHI must never go in a signal.
Verify UI changes by capturing Playwright screenshots and attaching them to the PR (GitHub-native, no third-party hosts). Use when a change touches the frontend.
| name | backlog-scoring |
| description | Reference only (do NOT invoke as an action): the RICE scoring convention for the goal-loop backlog. Read by the loop/make-plan skills. |
How the goal-loop (/loop) and /make-plan score and re-score backlog items so the loop always works the highest-leverage task. Scores live in each Claude Code Task's metadata (the backlog source of truth).
For each backlog item the model estimates four factors and computes a score:
score = (reach × impact × confidence) / effort
| Factor | Meaning | Scale |
|---|---|---|
| reach | How much of the goal / how many users/files/cases this item moves | 1–10 (relative) |
| impact | How strongly it advances the goal when done | 0.25, 0.5, 1, 2, 3 (massive) |
| confidence | How sure we are about reach × impact | 0.5 (low), 0.8 (med), 1.0 (high) |
| effort | Estimated cost to complete (person-equiv units) | ≥ 0.5 |
The estimates are model-set (the model fills them from the item description + investigation), and user-overridable — if the user sets any factor, keep it and don't overwrite it on re-score.
On each Task, in metadata:
metadata: {
goalId, reach, impact, confidence, effort,
score, // recomputed; (reach*impact*confidence)/effort
scoreSource, // "model" | "user" per factor that was overridden
cycle, // cycle last scored
learnings // rolling 1-line digest from the last attempt
}
/make-plan during a goal run, backlog-gen inside the loop, or a consumed signal promoted to a task — see .agents/skills/signals/SKILL.md; carry the signal's confidence into the RICE confidence factor and tag the task with the signal id). Set all four factors + score.effort; a newly-found high-leverage gap gets high reach/impact), recompute score. Never overwrite a user-set factor.score among open, unblocked tasks for the active goalId.effort (smaller wins first), then lowest Task id.blockedBy) are never selected regardless of score.