Paxel Skill — Auto-avaliação Local de Sessões Claude
skoppisetty/paxel-skill
Score suas sessões Claude localmente com a rubrica Paxel, sem upload.
Ver no GitHubSkill que avalia suas sessões Claude Code localmente usando a rubrica exata do YC Paxel (5 eixos: Execution Leverage, Steering, Engineering Quality, Product Thinking, Planning), sem enviar dados para servidor externo. Utiliza Claude Haiku para scoring determinístico e oferece transparência total do pipeline de avaliação.
Instalação
npx convez-ai skoppisetty/paxel-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Auto-avaliação de qualidade e efetividade em projetos Claude Code
- Análise de padrões de steering e engineering quality em workflow pessoal
- Feedback estruturado sobre planning e product thinking em sessões de coding
Veredito

Ferramenta sofisticada e transparente para quem quer entender profundamente sua própria performance em Claude Code usando a métrica de referência do YC Paxel, mas exige familiaridade com Python e pipelines de dados. Implementação fidedigna do rubric original, com a troca sensata de gpt-5.5 para Haiku.
README
A tradução para o português deste README ainda está na fila.
paxel-skill
A local self-assessment that scores your own Claude Code and Codex CLI sessions on YC Paxel's exact 5-axis rubric — without uploading anything to YC.
It bundles Paxel's verbatim scoring and narrative prompts (extracted from the public ghcr.io/yc-software/paxel-client image) and its band-cut thresholds. The one change: every narrative/scoring call is forced to Claude Haiku 4.5 (model: haiku), not YC's gpt-5.5-none proxy. Nothing leaves your machine.
What it does
your ~/.claude/projects/**/*.jsonl + ~/.codex/sessions/**/*.jsonl + your git repo
→ condense.py (deterministic: drop tool/file bodies, scrub secrets, chunk;
Codex rollouts auto-detected & normalized to the same form)
→ events.py (deterministic: events, session_signals, user highlights,
plan files, dispatch/return pairs — EventExtractor port)
→ gitdata.py (deterministic: commit grouping + session→episode linking
with real git LOC — CommitGrouper/EpisodeLinker port)
→ decisions.py (extract → Haiku classifier, verbatim v5 prompt + 49-law
catalog → finalize with in-session outcome analysis)
→ narrative prompt (Haiku, verbatim Paxel SYSTEM_PROMPT) → per-session notes
→ episodes.py (deterministic: byte-faithful build_episode_input — all
blocks, caps, Code volume, dispatch ratio)
→ scoring prompt (Haiku, verbatim Paxel rubric) → 5-axis scores
→ aggregate.py (verbatim band cuts + approx rollup) → overall + band
→ analytics.py (optional: the upload-only payload Paxel's server would
receive — 14 code-quality dims, velocity, steering traces)
Axes: Execution Leverage, Steering, Engineering Quality, Product Thinking, Planning (1.0–10.0). Bands: WEAK / LIMITED / STRONG / ELITE / EXEMPLAR.
What's faithful vs approximate (read this)
| Layer | Fidelity |
|---|---|
| Scoring rubric + calibration | Verbatim from the client image |
| Narrative prompt | Verbatim |
| Condensing (drop bodies, scrub, caps) | Faithful port of the client pipeline |
| Codex CLI ingestion | Faithful port of CodexNormalizer (boilerplate filter, apply_patch → per-file edits, update_plan → plan signal), plus one deliberate extension: the newer custom_tool_call apply_patch shape, which the archived client misses |
Event extraction + session_signals | Faithful port of EventExtractor / SessionSignalExtractor (verbatim regexes, caps, Ruby string semantics) |
| Commit grouping + episode linking | Faithful port of CommitGrouper / EpisodeLinker / LinkingStrategy (pr 1.0 / sha 0.9 / branch 0.7 / ±1h timestamp 0.5; first-match-wins; real numstat LOC) |
| Decision exchanges | Faithful port of extractor + classifier (verbatim v5 prompt + 49-law catalog, Haiku-classified) + deterministic in-session outcome analysis; chains not ported (need YC's embedding service — the client itself degrades identically when embeddings fail) |
Episode input (build_episode_input) | Byte-faithful port — all blocks, caps (50K/30K/10K/5K), Code volume gating, committed-return ratio. ## Code Reviews absent exactly as in Paxel's own local pipeline (server-side only) |
| Upload-only analytics (14 code-quality dims, velocity, steering, parallelism) | Faithful port, exposed as a transparency report (analytics.py) — never feeds scores, same as in Paxel |
| Band cut thresholds (<4/<6/<8/<9/≥9) | Verbatim |
| Per-axis scores | Faithful (model differs — see below) |
| Overall score + band | APPROXIMATION — YC's rollup is server-side, not in the image |
| Scorer model | Claude Haiku 4.5 (forced via model: haiku), not gpt-5.5-none (Paxel's actual model) — directional, not identical |
LLM scoring is nondeterministic; re-runs vary. The overall number is a calibration target, not YC's verdict. Full gap analysis: reference/GAPS.md.
Known local deviations: decision-exchange chains (need embeddings); pr_number recovered from in-session gh pr create output instead of Paxel's gh-CLI sidecar; committed-return ratio via the parent-commit-after-return branch (no child subagent session records locally); Gemini CLI / Cursor / opencode ingestion not ported (only Claude Code + Codex CLI here).
Install (as a Claude Code skill)
Clone straight into your Claude Code skills directory:
git clone https://github.com/skoppisetty/paxel-skill.git ~/.claude/skills/paxel-skill
Start a new Claude Code session (skills load at startup), then invoke /paxel-skill — or just ask "score my Claude Code sessions on the Paxel rubric." Update later with git -C ~/.claude/skills/paxel-skill pull.
Hacking on it? Clone anywhere and symlink instead, so your edits are live:
git clone https://github.com/skoppisetty/paxel-skill.git
ln -s "$(pwd)/paxel-skill" ~/.claude/skills/paxel-skill
Requirements: Claude Code + Python 3 (standard library only — nothing to pip install).
Note: scoring is per-project — it defaults to the logs for the repo you're currently in (
~/.claude/projects/<encoded-cwd>/, plus Codex sessions whosecwdmatches), not your entire history. Point it at another project's log dir to score that one instead.
Run the scripts directly (no skill)
python3 scripts/condense.py ~/.claude/projects/<encoded-cwd> > sessions.jsonl
# narrate + score each session with the prompts/ files (any LLM), collect to episodes.json
python3 scripts/aggregate.py episodes.json
Tests (stdlib only): python3 scripts/test_scripts.py.
Provenance & IP
Prompts and thresholds are extracted verbatim from YC's publicly pullable client image for personal, interoperability/assessment use. They are YC's text, not original work here — do not redistribute this as your own rubric. If you publish or share a derivative, reimplement the rubric in your own words.