Repository Score Review
SkillDev toolsScore is a monitoring subagent that designs honest evaluation frameworks for machine learning models. Once added, your AI can pick metrics that reflect real business costs and compare models with statistical rigor. It is useful whenever you need to choose evaluation metrics or decide between competing models.
Use Repository Score Review in Claude, ChatGPT or Ahel Desktop
Free. Sign in, add Repository Score Review and connect your AI. About a minute.
Also: Claude Code · Cursor · Codex
Then ask your AI: use the Repository Score Review skill
Details
Instructions available. Your AI can read the instructions. Execution depends on the setup they require.
Account requirements not reviewed. Check the skill instructions before use; Ahel provides instructions and does not run this skill.
No other account needed.
Add Score, then ask your AI to design a model evaluation framework or to compare your models statistically.
What your AI can do with it
- Select evaluation metrics matched to your business cost functions
- Run statistical significance tests when comparing models
- Check model calibration
- Analyze confusion results to see where a model gets predictions wrong
- Design a complete evaluation framework for a model
What this skill tells your AI
The instructions your AI receives, as published by esposter/esposter in .agents/skills/score/SKILL.md and read by Ahel’s review.
SCORE.md at the repo root is a periodic self-audit of the repository across its areas, each with a score, terse notes, and explicitly accepted trade-offs. It is the reference for "how healthy is this repo right now", and the README badge advertises the overall number.
The three things that must stay in sync
Every score review updates all three in the same commit:
SCORE.md— theLast revieweddate, theNuxt compatibilityDatemirrored in the header line, the overall total, each area score/note, and the per-area sections.README.md— the[badge-score]shields URL (score-<total>%2F100-<color>). The number must equal SCORE.md's overall.apps/web/configuration/compatibilityDate.ts— bumped to the review date.
Why compatibilityDate moves with the review
Nuxt's compatibilityDate opts into the framework behaviour as of a given date. Left alone it silently drifts years behind the installed Nuxt, so new defaults never activate and the eventual bump becomes a big-bang change. The score review is the natural checkpoint: it is the moment we deliberately re-read the repo, so it's the moment to advance the date, verify nothing broke, and record the new date in the SCORE.md header. Set it to the review date, check it ("Process" on this page, step 2), and if something breaks that's a finding for the review, not a reason to revert silently — fix it or note it as an accepted trade-off with the older date kept.
Process
- Re-audit each area against the code as it exists today — versions, workflows, and CSP/security posture all go stale. Volatile counts (test files, routers, stores) do not need re-auditing, because they are written as magnitudes rather than readings ("Writing style" on this page). Re-check one only when its magnitude has plausibly moved.
- Bump
compatibilityDateto today, then runpnpm typecheckfromapps/web/. The suite is CI's — a full local run is thetestingskill's ban (references/running-the-suite.md) — so a behaviour the new date changes surfaces there, and is a finding for the review. - Rewrite
SCORE.md: header line (date + compatibilityDate + overall), summary table, and the section bodies. Keep it terse — notes are one line each. - Update the README badge number and color.
- Commit all three together so the badge never advertises a stale number.
Writing style
The shared writing rules are the docs skill's. Three are specific to this file:
- The scores are exact, their evidence is a magnitude. A score is the artifact rather than a reading of the tree, so it stays precise; the counts cited to justify one are wrong within weeks, and a stale count discredits the score it was cited for. Held by
apps/web/content/docs/getPreciseCounts.test.ts, which fails on a precise count of anything the repo can count — an enumeration is its own count. - Accepted trade-offs are first-class. A score below maximum must say what the trade-off is and why it was accepted, not just what's missing. An area at maximum with a known compromise (e.g.
skipLibCheck) still records it. - No aspiration. SCORE.md describes the repo as it is, never as it will be. Planned work belongs in the relevant
docs/<area>/roadmap.md.
Signals
- GitHub stars
- 23
- Forks
- 3
- Last commit
- Oct 2026
Advanced
- Item type
- skill
- Key
score- Source
- github.com/esposter/esposter