Backlog scoring — sophisticated, model-set prioritization

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).

Model: RICE

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.

Where it's stored

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
}

When scoring happens

• Initial — when items enter the backlog (/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.
• Re-score — every Review & learn step: update factors from what the last cycle taught (e.g. an item that proved harder gets higher effort; a newly-found high-leverage gap gets high reach/impact), recompute score. Never overwrite a user-set factor.
• Selection — the loop's "next item" trigger picks the highest score among open, unblocked tasks for the active goalId.

Notes

• Keep estimates cheap — this is prioritization, not precision. A 1-line rationale per factor is enough; do not write essays into metadata.
• Ties broken by lowest effort (smaller wins first), then lowest Task id.
• Blocked tasks (open blockedBy) are never selected regardless of score.