Przejdź do treści
Biblioteka skilli
AI i agenci · Analiza

conductorscore

/conductorscore

Ocenia skill Claude Code na podstawie lokalnych transkryptów: wpływ na pracę agenta, jakość wykonania i antywzorce.

Repozytorium
conductorscore ↗
Licencja
Apache-2.0
Ostatnia zmiana repozytorium
25.10.2026

Ocena i bezpieczeństwo

100/ 100Klasa: dobry
Format25 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność25 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
62
Skrypty
60
Komentarze HTML
0
Adresy URL w SKILL.md
2
Przestarzałość

Brak sygnałów. Repozytorium zmienione 25.10.2026, bez odwołań do wycofanych modeli i API.

Licencja

Apache-2.0, otwarta: można używać i zmieniać, z zachowaniem licencji i informacji o zmianach.

Co sprawdzić przed instalacją
  • 60 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.

Skrypty: scripts/__init__.py, scripts/_http.py, scripts/agents/__init__.py, scripts/agents/base.py, scripts/agents/claude/__init__.py, scripts/agents/claude/config.py, scripts/agents/claude/discovery.py, scripts/agents/claude/events.py, scripts/agents/claude/taxonomy.py, scripts/agents/codex/__init__.py, scripts/agents/codex/config.py, scripts/agents/codex/discovery.py, scripts/agents/codex/events.py, scri

Co robi i kiedy

Ocenia skill Claude Code na podstawie lokalnych transkryptów: wpływ na pracę agenta, jakość wykonania i antywzorce.

Oryginalny opis (EN)

Score your Claude Code skill — measure agent leverage, craft, and anti-patterns from your local transcripts. Use when the user types /conductorscore.

Jak działa

  1. How to run
  2. Progress relay rule
  3. ASK relay rule
  4. Final report
  5. Debug / visualize a session — /conductorscore debug
  6. Device-flow login
  7. Privacy & data handling
Sekcje z nagłówków SKILL.md, w oryginale · 163 linii, 1 244 słów
Podgląd SKILL.md
---
name: conductorscore
description: Score your Claude Code skill — measure agent leverage, craft, and anti-patterns from your local transcripts. Use when the user types /conductorscore.
---

ConductorScore reads your local coding-agent transcripts and uploads only the
resulting numbers (your score) — never transcript text.

## How to run

When the user invokes this skill, say **"Calculating your ConductorScore"** and,
on the next line, the skill version the script reports (its first output line is
`Version X.Y.Z`) — e.g. show `Version 0.5.0` beneath the heading. Run the
orchestrator from **this skill's own directory** — the folder that contains this
SKILL.md, with `run.py` in its `scripts/` subdirectory. Don't assume a fixed
location: depending on the agent and installer the skill may live under
`~/.claude/skills/conductorscore/`, `~/.codex/skills/conductorscore/`,
`~/.cursor/skills/conductorscore/`, or `~/.agents/skills/conductorscore/`.
Use the `scripts/run.py` next to this file:

```
python3 <this-skill-dir>/scripts/run.py --step
```

On Windows, if `python3` isn't found, use `python` or `py -3` instead. Always
pass `--step` when running `run.py` for a scan, including the re-runs below.

The script drives the flow. Don't narrate the steps. Your jobs are: (a) run the
script, (b) relay its progress (below), (c) handle the `CONDUCTORSCORE_ASK`
lines below, and (d) when the scan finishes, render the **final report** from
the `CONDUCTORSCORE_RESULT` line (see "Final report").

## Progress relay rule

The scan runs in the background, and each `run.py --step` call returns after a
few seconds. If the output ends with `CONDUCTORSCORE_PROGRESS <percent> "<text>"`,
the scan is still running: write `<text>` to the user as its own short message
(just that text, e.g. `Scanning your transcripts… 42%`, nothing else), then
immediately run `run.py --step` again with no other flags. Keep going until the
output has no `CONDUCTORSCORE_PROGRESS` line. Don't sleep between calls; the
script already waits.

## ASK relay rule

If a printed line starts with `CONDUCTORSCORE_ASK <id> "<question>" [opt A] [opt B] …`,
the script has STOPPED and is asking the user to decide. Present `<question>` as a
multiple-choice using the bracketed options (use the question UI; don't add extra
prose), then re-run `run.py` with the matching flag:

- `login` → the script printed a GitHub URL + code. Show them, let the user
  authorize in their browser, and when they confirm ("I've authorized") re-run
  `run.py --step` with NO new flag — it resumes the login and continues. On
  Cancel, stop.
- `providers` → re-run `run.py --step` with `--providers=all` (All),
  `--providers=claude` (Claude Code), `--providers=codex` (Codex), or
  `--providers=cursor` (Cursor). On Cancel, stop.
- `update` → asked BEFORE any scanning, when a newer ConductorScore is
  published. Present EXACTLY the two printed choices, **Update now** and
  **Not now**. On **Not now**, re-run `run.py --step --skip-update` (it scans
  with this version). On **Update now**:
  1. Run the command from the `CONDUCTORSCORE_UPDATE_CMD <command>` line
     exactly. Its output may mention other copies ("no GitHub metadata") —
     ignore that; don't judge success from the output.
  2. Read the `VERSION` file in this skill's directory (next to this SKILL.md).
  3. If it now shows the new version, re-run `run.py --step` (the updated
     script now runs the scan).
  4. Otherwise this copy isn't managed by `gh skill`, so the update can't
     reach it. Tell the user to reinstall it once with
     `gh skill install https://github.com/flatironconsulting/conductorscore conductorscore --agent <claude-code|codex|cursor> --scope user --force`
     (the agent you're running in), then re-run `run.py --step --skip-update`
     so they still get a score now.
- `daily` → present EXACTLY two choices, **Yes** and **No** (no Cancel, no other
  options). Re-run with ONLY `--daily=yes` or `--daily=no` — do NOT add
  `--providers`; this enables/declines the once-per-day hook and does NOT re-scan.

## Final report

When the scan finishes the script prints a single line:
`CONDUCTORSCORE_RESULT {"score":N,"url":"...","sessions":N,"verified_github":bool}`.
Don't show that raw line. The `--daily=yes|no` re-run prints it again with a
`"daily":true|false` field. Render the final report from that line **in the same
reply as the daily re-run, before anything else**.
Never end the run without showing the user the report and its URL. Use exactly
this shape (keep the ✓ and 📊 and the bullets):

```
✓ ConductorScore: <score>

📊 View your full breakdown: <url>

- Scanned <sessions> sessions from your local transcripts
- <Daily auto-refresh enabled — your score updates once per day | Daily auto-refresh not enabled>
- Only your score (the numbers) was uploaded — never any transcript text
```

## Debug / visualize a session — `/conductorscore debug`

`/conductorscore debug` renders a session locally as a self-contained HTML
timeline plus the Interactive / Background leverage table and opens it in the browser. It runs entirely on
the user's machine and **uploads nothing — not even the numbers a scan sends.**
Use this path when the user types `/conductorscore debug …`, or asks to
**visualize / view / debug a session**, show its **timeline**, or see the
**leverage breakdown** (Interactive / Background, HITL/AFK/Tool/Idle).

The slash interface IS the `session_viewer.py` interface: take whatever the user
put after `debug` and pass it **straight through** to the Python CLI — every
flag below maps 1:1, so `/conductorscore debug <ARGS>` runs
`session_viewer.py <ARGS>`.

```
python3 <this-skill-dir>/scripts/session_viewer.py [ARGS]
```

(`session_viewer.py` sits next to `run.py` in this skill's `scripts/`, wherever
the skill is installed. On Windows use `python` or `py -3` if `python3` isn't
found.)

Supported arguments (the full Python CLI surface):

- **(no args)** → the session the user is in. The SessionStart hook records the
  current transcript, so a bare run visualizes the active session. To inspect an
  earlier one, the user navigates to it first with the built-in picker
  (`claude --resume` / `/resume`, or `codex resume`), then runs
  `/conductorscore debug` — don't reinvent that navigation.
- **`<path.jsonl>`** → render that explicit transcript.
- **`--list`** → print discoverable local sessions (index, date, id, project) and
  exit.
- **`--pick N`** → render the Nth session from `--list` (newest first).
- **`--no-redact`** → show the real transcript text. Default is **redacted**:
  bubbles show the wire-equivalent view (hashes + token counts, tool names only —
  no raw text), matching exactly what a scan would upload, so the page is safe to
  share. `--no-redact` is safe locally (nothing leaves the machine) and best for
  hands-on debugging.
- **`--out <path>`** → write the HTML somewhere other than
  `~/.cache/conductorscore/`.
- **`--no-open`** → render without launching the browser.

The script prints a **local-only debug summary** — the source transcript, the
redaction mode, message/turn/streak/leverage counts, and an explicit
"nothing is uploaded" assurance — then writes the HTML and opens it. **Relay the
script's console summary and its `Rendered … → <path>` line as-is; don't
re-narrate or restate the numbers.**

## Device-flow login

First run prints a `https://github.com/login/device` URL and a code, then STOPS
(it does not poll in the background). Show the URL and code to the user, let them
authorize in their browser, and once they confirm, re-run `run.py --step` — it resumes
the pending login, finishes authentication, and proceeds to the scan.

## Privacy & data handling

ConductorScore runs entirely on your machine. It reads your local coding-agent
transcripts to compute metrics, then uploads only:

- **Numbers** — counts, token totals, timing, and derived scores.
- **Names you already surface by using them** — the skills, slash-commands, MCP
  servers, plugins, and model IDs invoked in your sessions (these appear on your
  public profile). Just the names — no arguments, no contents.

It never uploads transcript or message text, prompts, code, file contents, or file
paths. Message and file text is reduced on-device to a truncated SHA-256 hash (used
only for dedup/similarity) before any metric is computed. The exact wire payload is
documented and inspectable in `WIRE_FORMAT.md` (`wire_format_sample.json` is a real
sample). Uploads happen only after you pair this device, only for the agents you
consent to scan, and never on an unpaired machine.
Pliki w paczce (62)
  • SKILL.md 8 KB
  • VERSION 1 KB
  • scripts/__init__.py 1 KB
  • scripts/_http.py 3 KB
  • scripts/approval_counter.py 12 KB
  • scripts/auth_store.py 4 KB
  • scripts/commit_counter.py 4 KB
  • scripts/config_scanner.py 1 KB
  • scripts/cron_classifier.py 2 KB
  • scripts/daily.py 9 KB
  • scripts/delta_upload.py 6 KB
  • scripts/device_flow.py 4 KB
  • scripts/edit_counter.py 4 KB
  • scripts/events.py 4 KB
  • scripts/feature_cache.py 7 KB
  • scripts/frustration_detector.py 3 KB
  • scripts/output_schema.py 21 KB
  • scripts/plan_signals.py 10 KB
  • scripts/prompt_similarity.py 4 KB
  • scripts/reauth.py 4 KB
  • scripts/revert_detector.py 5 KB
  • scripts/run.py 43 KB
  • scripts/scan.py 19 KB
  • scripts/scanner.py 34 KB
  • scripts/session_viewer.py 102 KB
  • scripts/session_window.py 3 KB
  • scripts/status_writer.py 1 KB
  • scripts/time_accounting.py 57 KB
  • scripts/tool_counter.py 33 KB
  • scripts/turn_classifier.py 32 KB
  • scripts/version_check.py 3 KB
  • scripts/agents/__init__.py 1 KB
  • scripts/agents/base.py 9 KB
  • scripts/agents/consent.py 14 KB
  • scripts/agents/registry.py 3 KB
  • scripts/agents/claude/__init__.py 1 KB
  • scripts/agents/claude/config.py 8 KB
  • scripts/agents/claude/discovery.py 5 KB
  • scripts/agents/claude/events.py 38 KB
  • scripts/agents/claude/taxonomy.py 3 KB
  • scripts/agents/codex/__init__.py 2 KB
  • scripts/agents/codex/config.py 4 KB
  • scripts/agents/codex/discovery.py 7 KB
  • scripts/agents/codex/events.py 38 KB
  • scripts/agents/codex/taxonomy.py 16 KB
  • scripts/agents/cursor/__init__.py 2 KB
  • scripts/agents/cursor/cli_events.py 32 KB
  • scripts/agents/cursor/config.py 4 KB
  • scripts/agents/cursor/discovery.py 16 KB
  • scripts/agents/cursor/events.py 14 KB
  • scripts/agents/cursor/store.py 9 KB
  • scripts/agents/cursor/taxonomy.py 10 KB
  • scripts/core/__init__.py 1 KB
  • scripts/core/config_counts.py 2 KB
  • scripts/core/edits.py 3 KB
  • scripts/core/normalized.py 9 KB
  • scripts/core/paths.py 1 KB
  • scripts/core/prompt_origin.py 3 KB
  • scripts/core/shell.py 3 KB
  • scripts/core/text.py 3 KB
  • scripts/core/thresholds.py 3 KB
  • scripts/core/timestamps.py 2 KB

Wersje

W bazie nie ma innych skilli o nazwie conductorscore.

Migawka bazy z 13.11.2026 · wersja paczki 36385448fcf0 (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 86measure
    Mierzy, jak użytkownik pracuje z Claude Code na podstawie lokalnych transkryptów i pokazuje raport, bez zmian w plikach.
  • 100skill-craft
    Ocenia, naprawia, tworzy i audytuje skille Claude według ośmiomodułowego systemu punktacji, wykrywając antywzorce i błędy.
  • 100skill-review
    Przegląda i audytuje skille Claude Code, analizując strukturę, jakość opisu, projekt workflow, oszczędność tokenów i antywzorce względem dobrych praktyk.
  • 100claude-md
    Pomaga pisać skuteczne pliki CLAUDE.md dla projektów Claude Code: strategia treści, struktura, antywzorce, stopniowe ujawnianie informacji i hierarchiczne rozmieszczenie.
  • 100leverage-cloud-agents
    Konfiguruje Leverage Computer, czyli wspólne miejsce pracy agentów w chmurze dla Claude Code i Codex, i uruchamia pierwsze zadanie użytkownika.
  • 100k-tokens
    Obniża zużycie tokenów w Claude Code: audyt lokalnych transkryptów pokazuje, na co idą tokeny, a zastosowanie zmian dopisuje klucze do settings.json i reguły dla agentów do CLAUDE.md.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.