| name | self-scoring-loop |
| description | Iteratively improve a non-code deliverable (spec, plan, document, design) by writing an explicit rubric, scoring the work 0-100 against it, rewriting the weakest parts, and repeating until the score plateaus. Use when a deliverable has a quality bar but no test suite โ e.g. before handing over a spec, proposal, architecture doc, or report. |
Self-Scoring Loop
Turn a single-shot draft into a converging loop: score the work against an explicit rubric, fix the weakest parts, and repeat until improvement stalls. Scoring and rewriting are different acts โ critiquing the work before rewriting turns blind editing into directed search.
This is the quality loop for deliverables that have no test suite. For code, use the code-scoring-loop sibling skill โ the same loop with the diff scored by the framework's specialist agents, run before (never instead of) the review gates (code-review-gatekeeper, peer-review-critic).
When to Apply
- A spec, plan, proposal, or document is "done" but has not been judged against anything.
- The
/agentic-framework:spec command finishes a draft spec (self-score it before presenting it).
- A deliverable will be consumed by someone else (or another agent) and errors are expensive downstream.
- The user asks to "polish", "tighten", or "make this good" without concrete criteria.
The Loop
- Write the rubric first โ before judging. 5โ7 criteria specific to this deliverable, each with a weight (weights sum to 100). Generic criteria produce generic scores; a spec's rubric should weigh things like testable acceptance criteria and explicit out-of-scope fence, not "clarity" alone. The rubric is fixed for the rest of the loop โ do not move the goalposts between iterations.
- Score 0โ100 against the rubric, with a one-line justification per criterion. Score honestly: an inflated score ends the loop early and defeats it.
- Name the 1โ2 weakest criteria and exactly why they lost points.
- Rewrite only the weak parts. Keep what scored well โ a full rewrite discards verified quality and restarts convergence.
- Rescore and decide:
- Improvement < 3 points, or 3 iterations reached โ stop. More loops past a plateau produce churn, not quality.
- Otherwise โ back to step 3.
- Report the final version, the final score with per-criterion breakdown, and the score trajectory (e.g. 71 โ 84 โ 88) so the reader can see the loop worked.
Rules
- Rubric before judgment, always โ scoring without a rubric is vibes.
- The rubric must be visible in the report; a hidden rubric is unfalsifiable.
- Never revise the rubric mid-loop to make a score look better.
- If the deliverable scores โฅ 90 on the first pass, say so and stop โ the loop is for improvement, not ceremony.
- This loop is single-agent and advisory; it does not replace the framework's review gates for code โ code takes the
code-scoring-loop variant.