project-memory — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited project-memory (Agent Skill) and scored it 92/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 2 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 2 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
A bulleted imperative like {match} tells the agent to never reveal, disclose, or mention something to the user. Used adversarially it can instruct the agent to hide its tool calls or lie about what it did — stripping the transparency a user relies on to trust the agent.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Create CLAUDE.md files that transfer institutional knowledge, not obvious information. Think like a senior journalist onboarding a competent colleague — you don't explain how journalism works, you explain YOUR project's quirks.
Anthropic is explicit on this point: CLAUDE.md content is delivered as a user message after the system prompt. Claude reads it and tries to follow it, but there's no guarantee of strict compliance — especially with vague or conflicting instructions. Source: https://code.claude.com/docs/en/memory
This affects how you write a CLAUDE.md and what you put elsewhere:
| Mechanism | Use for | Source-of-truth |
|---|---|---|
| CLAUDE.md | Standing facts, conventions, "always do X" rules | Advisory |
| Skills | Multi-step procedures, on-demand workflows | Loaded when invoked |
| Hooks | Actions that must happen every time, no exceptions | Deterministic — runs as a shell command (e.g., hooks/one-way-door-check.md) or as a prompt the harness enforces (e.g., hooks/enforce-test-first.md) |
If an instruction is "block writes to published/" or "run accessibility check before commit," that belongs in a hook, not CLAUDE.md. If it's "fact-check workflow" or "FOIA-letter drafting," that's a skill. CLAUDE.md is the place for things Claude must hold in every session.
Anthropic publishes an explicit include/exclude table for what makes an effective CLAUDE.md. Source: https://code.claude.com/docs/en/best-practices
| ✅ Include | ❌ Exclude |
|---|---|
| Bash commands Claude can't guess | Anything Claude can figure out by reading code |
| Code style rules that differ from defaults | Standard language conventions Claude already knows |
| Testing instructions and preferred test runners | Detailed API documentation (link to docs instead) |
| Repository etiquette (branch naming, PR conventions) | Information that changes frequently |
| Architectural decisions specific to your project | Long explanations or tutorials |
| Developer environment quirks (required env vars) | File-by-file descriptions of the codebase |
| Common gotchas or non-obvious behaviors | Self-evident practices like "write clean code" |
For journalism projects, translate to:
For every line you write, ask: "Would removing this cause Claude to make mistakes?" If not, cut it.
Anthropic's explicit guidance as of 2026: target under 200 lines per CLAUDE.md file. Longer files consume more context and reduce adherence. Bloated CLAUDE.md files cause Claude to ignore the actual rules. Source: https://code.claude.com/docs/en/best-practices
Going over 200 lines is a signal to use one of these instead:
@import chains as the size-management mechanism. (Note: @imports no longer help with context size — Anthropic explicitly notes that imports load fully at launch.)CLAUDE.md location precedence (more specific wins). Source: https://code.claude.com/docs/en/memory
| Scope | Location | Use case |
|---|---|---|
| Managed policy | Linux/WSL: /etc/claude-code/CLAUDE.md<br/>macOS: /Library/Application Support/ClaudeCode/CLAUDE.md<br/>Windows: C:\Program Files\ClaudeCode\CLAUDE.md | Org-wide standards (IT/DevOps managed; cannot be excluded by individual settings) |
| Project | ./CLAUDE.md OR ./.claude/CLAUDE.md | Team-shared instructions (check into git) |
| User | ~/.claude/CLAUDE.md | Personal preferences across all projects |
| Local | ./CLAUDE.local.md | Personal project-specific notes (add to .gitignore) |
The ./.claude/CLAUDE.md location and managed-policy tier are 2026 additions to Anthropic's documented locations. Templates that say "put this at project root" should mention ./.claude/CLAUDE.md as an equally valid location.
If your repo already has AGENTS.md for other coding agents (Cursor, Codex, etc.), don't duplicate the content. Anthropic's recommended pattern is @AGENTS.md import (or symlink) with Claude-specific overrides appended:
@AGENTS.md
## Claude Code
- Use plan mode for changes under `src/billing/`.For journalism teams using multiple agents, this matters — shared editorial standards (AP-style preferences, source-handling invariants, fact-check protocols) belong to the org, not one agent.
Anthropic now names these explicitly. Source: https://code.claude.com/docs/en/best-practices and https://code.claude.com/docs/en/memory
CLAUDE.local.md (gitignored) for sandbox URLs, test credentials, personal overrides.Anthropic ships this as the canonical starter. Templates should default to something this short and grow only as needed:
# Code style
- Use ES modules (import/export) syntax, not CommonJS (require)
- Destructure imports when possible
# Workflow
- Be sure to typecheck when you're done making a series of code changes
- Prefer running single tests, and not the whole test suite, for performanceEach of the 6 journalism templates ships a sub-30-line "starter" variant alongside the fuller version. Adopt incrementally.
As of Claude Code v2.1.59+, there's a second memory mechanism alongside CLAUDE.md: auto memory, where Claude writes notes to itself in ~/.claude/projects/<project>/memory/ based on your corrections. The first 200 lines of that directory's MEMORY.md are loaded into every session. Source: https://code.claude.com/docs/en/memory
Practical implication for templates: don't ask the user to manually write down "things Claude learns over time" — that's auto memory's job now. CLAUDE.md is for facts you write up front; auto memory is for things Claude notices.
Bad (too verbose, obvious):
# CLAUDE.md
## Overview
Welcome to our newsroom's story tracking system! This is a
web application built with React and Node.js that helps
editors and reporters collaborate on stories.
## Getting started
First, make sure you have Node.js installed. Then:
npm install
npm startGood (institutional knowledge only):
# CLAUDE.md
## Overview
Story tracker for metro desk. React + Supabase.
## Gotchas
- Story slugs must be unique across ALL desks, not just metro
- "Hold" status doesn't stop the autopublish cron — use "Kill"
- Reporter dropdown caches for 1 hour; new hires won't appear
## Commands
npm run sync-ap # Pull latest from AP, runs automatically at :15
## Credentials
Supabase key in 1Password "Metro Desk" vault, not .envTemplates are in the templates/ directory:
| Template | Use for |
|---|---|
editorial-tool.md | Newsroom tools, fact-checkers, AI assistants |
event-website.md | Conferences, workshops, campaign sites |
publication.md | Newsletters, podcasts, ongoing content series |
research-project.md | Investigations, data journalism with defined scope |
content-pipeline.md | CMS workflows, publishing automation |
digital-archive.md | Historical collections, document repositories |
What are you building?
├── Tool for the newsroom → editorial-tool.md
├── Site for an event → event-website.md
├── Recurring content series → publication.md
├── One-time investigation → research-project.md
├── Publishing automation → content-pipeline.md
└── Archive/preservation → digital-archive.md./CLAUDE.md OR ./.claude/CLAUDE.md at your project root@AGENTS.md at the top instead of duplicating shared content2026-05-09. Sources verified: code.claude.com/docs/en/memory, code.claude.com/docs/en/best-practices.
The best CLAUDE.md files are written by people who've been burned by the quirks they're documenting.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.