pulser — SKILL.md Linter for Claude Code

Website · npm · 한국어

"Check my skills." Diagnose, classify, prescribe, and fix — without leaving the conversation.

$ pulser

  pulser v0.4.0

  54 skills scanned · Score: 89/100
  ✓ 48 healthy  ⚠ 4 warnings  ✗ 2 errors

  Top issues:
    cardnews    — No Gotchas, no allowed-tools
    geo-audit   — 338 lines, single file

  💊 Rx #1 — cardnews
  [GOTCHAS] Add Gotchas section
    Why: Anthropic's highest-ROI improvement
    Template:
      ## Gotchas
      1. Validate output against conventions
      2. Check scope — don't over-generate

  Fix type: AUTO

What it does

pulser scans your SKILL.md files against 8 diagnostic rules derived from Anthropic's published principles in "Building Claude Code: How We Use Skills":

Rule	What it checks
frontmatter	Required name and description fields
description	Trigger keywords, "Use when" pattern, length
file-size	SKILL.md under 500 lines
gotchas	Gotchas section with failure patterns
allowed-tools	Tool restrictions appropriate for skill type
structure	Supporting files for large skills
conflicts	Trigger keyword overlap between skills
usage-hooks	Skill usage logging hook installed

Each skill is auto-classified by type (analysis, research, generation, execution, reference) with confidence scoring. Prescriptions are tailored to the detected type.

Install

npm install -g pulser-cli

On install, pulser registers itself as a Claude Code skill — say "check my skills" or /pulser to run it conversationally.

Usage

In Claude Code

Just say it:

check my skills

Or use the slash command:

/pulser

Claude runs the diagnosis, summarizes results, and offers to fix issues — all within the conversation.

In terminal

# Scan default path (~/.claude/skills/)
pulser

# Scan a specific directory
pulser ./my-skills/

# Scan a single skill
pulser --skill reasoning-tracer

# Auto-fix issues with backup
pulser --fix

# Rollback the last fix
pulser undo

# JSON output (for CI/automation)
pulser --format json

# Markdown report
pulser --format md

# Treat warnings as errors
pulser --strict

# Disable TUI animation
pulser --no-anim

Eval — Skill Testing

New in v0.4.0: test your skills against real inputs.

pulser eval

Write eval.yaml next to your SKILL.md:

tests:
  - name: "catches bugs"
    input: "Review: function add(a,b) { return a - b }"
    assert:
      - starts-with: "The bug"
      - contains: "subtract"
      - min-length: 30

pulser runs each test through claude -p, checks assertions, and tracks regressions automatically.

$ pulser eval

  reviewer (2 tests)
    ✓ catches bugs         320ms
    ✓ passes clean code    280ms

  2 passed · 0 failed · 0.6s

Supported assertions: contains, not-contains, starts-with, ends-with, min-length, max-length, matches (regex).

Exit Codes (eval)

Code	Meaning
0	All tests passed
1	Failures found
3	Regression detected (previously passing test now fails)

Core Pipeline