→ 01
Built on 265 versions of Anthropic's own prompt
We tracked every word Anthropic added, deleted, and rewrote in Claude Code's system prompt. When they cut IMPORTANT from 12 uses to 4, we knew. Now your harness can follow the same rules.
Your AGENTS.md, CLAUDE.md, CI, and hooks are the agent harness. AgentLint lints it — for Claude Code, Cursor, and Codex.
Audits CLAUDE.md, AGENTS.md, .cursor/rules, .github/copilot-instructions.md, plus your CI workflows, pre-commit hooks, and .gitignore — every check cites a primary source.
Works with every major AI coding agent
In February 2026, Mitchell Hashimoto (HashiCorp) coined harness engineering. Days later, OpenAI formalized it. LangChain gave the cleanest definition: if you're not the model, you're the harness — and the harness is where the real bottleneck lives.
For a coding agent, the harness isn't a framework — it's already in your repo. AGENTS.md and CLAUDE.md are the persistent rules. .cursor/rules and copilot-instructions.md are the tool-specific layers. CI, pre-commit hooks, and .gitignore are the deterministic constraints the agent cannot override. That's the harness. That's what AgentLint audits.
Other tools generate templates. Other tools score prompts. AgentLint is the only one that measures the health of your whole harness — and every check cites a primary source, not a blog post.
We tracked every word Anthropic added, deleted, and rewrote in Claude Code's system prompt. When they cut IMPORTANT from 12 uses to 4, we knew. Now your harness can follow the same rules.
40,000-character entry files get silently truncated. 256 KB files can't be read. Pre-commit hooks that take too long hang your session forever — Claude Code never uses --no-verify. We read the source so you don't have to.
One command. AgentLint scans your repo, scores the harness across 5 dimensions, shows exactly what's broken, and fixes what it can — pinning Actions to SHAs, adding .env to gitignore, generating handoffs.
Scores are measurements, not judgments. Reference values come from Anthropic's own data. You decide what to fix.
/init generates a template. cclint lints one file's syntax. claude-code-rules collects rules to copy-paste. None of them measure whether the harness as a whole is healthy — and none auto-fix what they find.
| Capability | /init | cclint | claude-code-rules | AgentLint |
|---|---|---|---|---|
| Generates a template CLAUDE.md | ✓ | — | — | — |
| Audits entry-file quality against 265 prompt versions | — | — | — | ✓ |
| Finds broken @include references | — | ✓ | — | ✓ |
| Enforces 40,000-char hard limit | — | — | — | ✓ |
| Audits CI, hooks, .gitignore, Actions SHA | — | — | — | ✓ |
| Detects instruction rot over time | — | — | — | ✓ |
| Auto-fixes what it finds | — | — | — | ✓ |
| Every check cites a data source | — | — | — | ✓ |
If a check can't be backed by a primary source, it doesn't ship. That's the rule.
Install agentlint and tighten the CLAUDE.md instructions your agents rely on.
Use agentlint with the coding agent workflow and CLAUDE.md shape you already have.
Guides for stronger CLAUDE.md files and cleaner agentlint results.
One command. 33 evidence-backed checks. Every one with a citation.