Project Health skill

Project Health

One skill for everything about your project's Claude Code configuration. Run it at the start, middle, or end of a project — it figures out what's needed.

Goal: Zero permission prompts, well-organised context files, no cruft.

When to Use

You say...	What happens
"project health" / "check project"	Full audit: permissions + context + docs
"setup project" / "kickoff" / "bootstrap"	New project setup from scratch
"tidy permissions" / "clean settings"	Fix permissions file only
"capture learnings" / "update CLAUDE.md"	Save session discoveries
"add python" / "add docker permissions"	Add a preset to existing settings
"audit context" / "audit memory"	Context-focused audit only

Architecture: Sub-Agents

Heavy analysis runs in sub-agents to keep the main conversation clean. The main agent orchestrates; sub-agents do the scanning and return summaries.

Agent 1: Permission Auditor

Launched with Task(subagent_type: "general-purpose"). Prompt:

Read .claude/settings.local.json.

**Discover connected MCP servers**: Use ToolSearch (search "mcp") and extract unique
server prefixes from tool names (e.g. mcp__vault__secret_list → vault).

**Discover installed skills**: Use the Skill tool or ToolSearch to list available skills.
For each skill that has scripts/ in its directory, note what Bash patterns it needs
(python3, env var prefixes like GEMINI_API_KEY=*, etc.). Check the SKILL.md for any
MCP tools the skill references (e.g. mcp__vault__secret_get).

Report:
1. MCP servers connected but NOT in settings (missing)
2. MCP servers in settings but NOT connected (stale)
3. Skill permissions: Bash patterns and MCP tools that installed skills need but aren't approved
4. File access: check for Read/Edit/Write patterns for .claude/** and //tmp/**
   in project settings, and ~/Documents/**/~/.claude/** in global settings
5. Leaked secrets: entries containing API keys, tokens, bearer strings, hex >20 chars, base64 >20 chars
6. Legacy colon syntax: entries like Bash(git:*) instead of Bash(git *)
7. Junk entries: shell fragments (Bash(do), Bash(fi), Bash(then), Bash(else), Bash(done)),
   __NEW_LINE_* artefacts, loop body fragments (Bash(break), Bash(continue), Bash(echo *))
8. Duplicates: entries covered by a broader pattern (e.g. Bash(git add *) redundant if Bash(git *) exists)
9. Missing presets: based on files present, suggest presets from [permission-presets.md]

Prefer Read/Glob/Grep tools over Bash. If you need to scan multiple files or
run 3+ commands for one analysis, write a Python script to .jez/scripts/
and run it once (mkdir -p .jez/scripts first).

Return a structured summary, not raw data.

Agent 2: Context Auditor

Launched with Task(subagent_type: "general-purpose"). Prompt:

Audit the project context landscape at [repo-path]:

1. Find all CLAUDE.md files. For each:
   - Count lines (target: root 50-150, subdirs 15-50)
   - Score quality on 6 criteria (see quality-criteria.md)
   - Check for stale file/path references
   - Flag oversized files

2. Find .claude/rules/ topic files. Check sizes (target: 20-80 lines).

3. Detect project type from files present (see project-types.md).
   Check expected docs exist (ARCHITECTURE.md, DATABASE_SCHEMA.md, etc.)

4. Find public markdown (README.md, LICENSE, CONTRIBUTING.md).
   Check for overlap with CLAUDE.md content.

5. Check auto-memory at ~/.claude/projects/*/memory/MEMORY.md

6. If Cloudflare project: find all wrangler.jsonc/wrangler.toml files.
   Check each has "observability": { "enabled": true }. Flag any missing it.

Prefer Read/Glob/Grep tools over Bash. If you need to scan many files or
aggregate data across the repo, write a Python script to .jez/scripts/
and run it once rather than running many individual bash commands
(mkdir -p .jez/scripts first).

Return: project type, quality scores, missing docs, stale refs, overlaps,
size violations, observability gaps, and total markdown footprint.

Parallel Execution

For a full health check, launch both agents in parallel:

Task(subagent_type: "general-purpose", name: "permission-audit", prompt: "...")
Task(subagent_type: "general-purpose", name: "context-audit", prompt: "...")

Both return summaries. The main agent combines them into one report and proposes fixes.

Mode 1: Full Health Check

The default. Run this anytime.

Steps

1. Launch Permission Auditor and Context Auditor agents in parallel

2. Combine findings into a single report:

   ## Project Health Report
   
   **Project type**: [detected type]
   **CLAUDE.md quality**: [score]/100 ([grade])
   
   ### Permissions
   - Missing MCP servers: [list]
   - Leaked secrets: [count] found
   - Legacy syntax: [count] entries
   - Missing presets: [list]
   
   ### Context
   - Oversized files: [list]
   - Stale references: [list]
   - Missing docs: [list]
   - Overlaps: [list]
   
   ### Recommended Fixes
   1. [fix 1]
   2. [fix 2]
   ...
   

3. Apply fixes after single yes/no confirmation

Mode 2: New Project Setup

When: No .claude/settings.local.json exists, or user says "setup" / "kickoff".

Steps

1. Detect project type from files present:

   Indicator	Type	Preset
   wrangler.jsonc or wrangler.toml	cloudflare-worker	JS/TS + Cloudflare
   vercel.json or next.config.*	vercel-app	JS/TS + Vercel
   astro.config.*	astro	JS/TS + Static Sites
   package.json (no deploy target)	javascript-typescript	JS/TS
   pyproject.toml or setup.py or requirements.txt	python	Python
   Cargo.toml	rust	Rust
   go.mod	go	Go
   Gemfile or Rakefile	ruby	Ruby
   composer.json or wp-config.php	php	PHP
   pom.xml or build.gradle*	java	Java/JVM
   *.sln or *.csproj	dotnet	.NET
   mix.exs	elixir	Elixir
   Package.swift	swift	Swift + macOS
   pubspec.yaml	flutter	Mobile
   Dockerfile or docker-compose.yml	docker	Docker
   fly.toml or railway.json or netlify.toml	hosted-app	Hosting Platforms
   supabase/config.toml	supabase	Hosting + Database
   .claude/agents/ or operational scripts	ops-admin	—
   Empty directory	Ask the user	—

   Types stack (e.g. cloudflare-worker + javascript-typescript).

2. Generate .claude/settings.local.json:

   • Read references/permission-presets.md
   • Always include Universal Base (includes file access for .claude/**, //tmp/**)
   • Add detected language + deployment presets
   • Check if global ~/.claude/settings.local.json has home-relative file access patterns (~/Documents/**, ~/.claude/**). If not, suggest adding them there (NOT in the project file — home paths belong in global settings only)
   • Launch Permission Auditor agent to discover MCP servers and add per-server wildcards
   • Always include WebSearch, WebFetch
   • Always include explicit gh subcommands (workaround for Bash(gh *) bug)
   • Write with // comment groups

3. Generate CLAUDE.md:

   • Read references/templates.md
   • Use project-type-appropriate template

4. Generate .gitignore:

   • Read references/templates.md
   • Always include .claude/settings.local.json, .claude/plans/, .jez/screenshots/, .jez/artifacts/
   • Do NOT gitignore .jez/scripts/ — generated scripts are worth keeping

5. Optionally (ask first): git init + gh repo create

6. Warn: "Project settings.local.json SHADOWS global settings (does not merge). Session restart needed."

Mode 3: Tidy Permissions

When: User says "tidy permissions" or health check found permission issues.

Launch the Permission Auditor agent, then apply its recommended fixes.

Mode 4: Capture Learnings

When: End of session, "capture learnings", "save what we learned".

This runs in the main context (not a sub-agent) because it needs access to the conversation history.

1. Review conversation for discoveries worth preserving
2. Decide placement:

   Applies to all projects?
   ├── YES → ~/.claude/rules/<topic>.md
   └── NO  → Specific to a subdirectory?
       ├── YES → <dir>/CLAUDE.md
       └── NO  → Reference or operational?
           ├── Reference → docs/ or ARCHITECTURE.md
           └── Operational → ./CLAUDE.md (root)
   

3. Draft all changes as diffs in a single batch
4. Apply after single yes/no confirmation

Keep it concise: one line per concept.

Mode 5: Add Preset

When: "add python permissions", "add docker", "add MCP servers".

1. Read the preset from references/permission-presets.md
2. Read existing .claude/settings.local.json
3. Merge without duplicating
4. Remind: session restart required

Mode 6: Restructure Context

When: Root CLAUDE.md over 200 lines, "restructure memory".

1. Launch Context Auditor agent first
2. Based on findings:
   • Split oversized CLAUDE.md into .claude/rules/<topic>.md
   • Extract directory-specific content to sub-directory CLAUDE.md
   • Move reference material to docs/
   • Resolve overlaps
   • Create missing docs for project type
3. Present plan, apply after approval

Size Targets

File	Target	Maximum
Root CLAUDE.md	50-150 lines	200
Sub-directory CLAUDE.md	15-50 lines	80
Rules topic file	20-80 lines	120

Permission Syntax Quick Reference

Pattern	Meaning
Bash(git *)	Preferred — space before * = word boundary
Bash(nvidia-smi)	Exact match, no arguments
WebFetch	Blanket web fetch
WebSearch	Blanket web search
mcp__servername__*	All tools on one MCP server

What Does NOT Work

Pattern	Why
mcp__*	Wildcard doesn't cross __ boundary
mcp__*__*	Still doesn't work
Bash(git:*)	Deprecated colon syntax (works but prefer space)

Important Behaviours

• Not hot-reloaded: settings.local.json edits need session restart
• "Don't ask again" injects at runtime (no restart) using colon format — normal
• Shadows, not merges: Project settings completely replace global
• gh bug: Bash(gh *) sometimes misses subcommands — include explicit Bash(gh issue *) etc.

Autonomy

• Just do it: Detect project type, launch audit agents, discover MCP servers
• Brief confirmation: Write/update files (single batch yes/no)
• Ask first: git init, GitHub repo, delete existing content, major restructures

Reference Files

When	Read
Building permission presets	references/permission-presets.md
Generating CLAUDE.md, .gitignore	references/templates.md
Scoring CLAUDE.md quality	references/quality-criteria.md
Detecting project type + expected docs	references/project-types.md
Setting up commit capture hook	references/commit-hook.md