AGENT SKILL

bs-score

Use when auditing a repo, PR, branch, PR review, AI skill, agent config, or one or more prompts for claims vs truth and real defects: follow the canonical bs-score SKILL.md, emit findings JSON, then run bs-score (higher = more bullshit).

License 许可
MIT
所属仓库
alexhawat/bs-score
最近更新
2026-11-14
Stars
★ 3
Forks
0

01 前言 Front Matter

name
bs-score
license
MIT
description
Use when auditing a repo, PR, branch, PR review, AI skill, agent config, or one or more prompts for claims vs truth and real defects: follow the canonical bs-score SKILL.md, emit findings JSON, then run bs-score (higher = more bullshit).
allowed-tools
Read, Grep, Glob, Bash

02 资源文件 1

03 SKILL.md 正文

Bullshit Score (bs-score) — Claude Code

Install

Two pieces: the skill files, and the bs-score command that scores them.

The skill — clone the repository into Claude Code's skills folder, so the root SKILL.md and the checklists/ it links to land together:

git clone --depth 1 https://github.com/alexhawat/bs-score ~/.claude/skills/bs-score   # personal
git clone --depth 1 https://github.com/alexhawat/bs-score .claude/skills/bs-score   # project

~/.claude/skills/bs-score applies to every project; .claude/skills/bs-score is checked into one repository. Keep the folder name bs-score — it is the skill's name.

The command — one line, and nothing here vendors it:

uv tool install git+https://github.com/alexhawat/bs-score   # puts bs-score on PATH

Canonical pack

This wrapper belongs to alexhawat/bs-score.

  1. Read SKILL.md at the repository root — the two-pass workflow, the per-review-type checklists, and the hard rules live there.
  2. Emit findings JSON matching findings.schema.json (no score / points).
  3. Score it:
bs-score findings.json --repo-root .          # installed
uv run bs-score findings.json --repo-root .   # from a clone of the repo

Every install option is in the repository README.

Quotes are verified against the files they name before anything is scored, so --repo-root must point at the tree you audited. For prompts and review bodies, add --sources (bs-score below is the same command):

bs-score findings.json --sources prompts/                    # prompt audit
bs-score findings.json --repo-root . --sources review.json   # review audit

From a clone, python3 score.py findings.json and uv run bs-score … do the same thing.

What it audits

target_kind Scope Weight profile
repo Docs claims and deep code baseline
pr / branch The diff and its callers landing defects weigh more
review A PR review vs the diff false claims and misses weigh more
skill Skill MD + the code it names unimplemented claims weigh more
agent Agent / persona configs unsafe instructions weigh most
prompt One or more prompts injection and secrets weigh most
docs One document, claims only wrong claims weigh most

Fixture examples

Run from a clone of the repo:

uv run bs-score examples/findings.valid.json  --repo-root examples/fixture-repo   # 17
uv run bs-score examples/findings.review.json --repo-root examples/fixture-repo \
    --sources examples/fixture-review/review.json                                # 10
uv run bs-score examples/findings.skill.json  --repo-root examples/fixture-repo   # 7
uv run bs-score examples/findings.agent.json  --repo-root examples/fixture-repo   # 6
uv run bs-score examples/findings.prompt.json --sources examples/fixture-prompts  # 18
uv run bs-score examples/findings.hallucinated.json \
    --repo-root examples/fixture-repo                                            # 0

That last one is the point: a fabricated audit scores zero because none of its quotes exist.

Never invent the score — report only what bs-score printed.

04 同仓库更多 Skills