What the checker does, and what it does not claim.

Why this project exists

A repository can accumulate several instruction files as teams switch or combine coding agents. The files often look individually reasonable while disagreeing on package manager, test command, delivery rules, or directory scope. This checker handles that narrow problem. It does not claim that a Markdown file can enforce model behavior.

What is tested today

• Exact @AGENTS.md and @./AGENTS.md imports are recognized in ordinary text, not code examples or comments.
• Existing imports preserve both supplied files unchanged. Adding an import preserves the complete original CLAUDE.md body, including headings, case, indentation, and duplicate lines.
• Package-manager mentions and potential positive/negative rule conflicts are suggestions for manual review, not confirmed conflicts.
• A referenced AGENTS.md that was not supplied blocks file generation; there is no placeholder substitute.
• Generated Markdown is copied and downloaded in the browser.
• Free-checker file contents have no upload route and are never written to the database.

The project's release checks run a production build, rendered-page tests, checker behavior tests, linting, and strict TypeScript validation.

Reproduce the September 12 regression cases

These are static checker tests, not a Claude Code runtime test. Download all three files into one empty folder, inspect them, then run node --test checker-regression-tests.mjs with Node.js 22 or later. No package install, network call, account, or private repository is needed.

• 22 input fixtures and expected outcomes (JSON)
• Checker reference implementation (JavaScript) — mechanically generated from the same pure checker used by this website; release tests verify parity.
• Runnable assertions (JavaScript)

The cases cover fenced examples, inline code, comments, exact filename casing, missing targets, scoped duplicate rules, Windows line endings, standalone migration, and report fences. They deliberately preserve repeated rules for a person to review.

Observed on September 12, 2026: all 22 downloadable cases passed locally. This confirms these fixtures only. It does not establish complete Markdown compatibility, successful Claude loading, or user adoption.

The worked example in the checker

This is a demonstration shipped with the product, not a customer result. The sample AGENTS.md says to use pnpm and not push directly to main. The sample CLAUDE.md says to use npm and to push directly to main after the task.

- Use pnpm install.
- Run pnpm test and pnpm lint.
- Do not push directly to main.

- Use npm install.
- Run npm test.
- Push directly to main after the task.

The checker asks you to review the package-manager mentions and the opposite push rules. It does not assign a health score or decide which command is right. With AGENTS.md selected, it keeps AGENTS.md unchanged and prepends an import to the complete CLAUDE.md body. Resolve the contradictions before saving. Select “Load sample” on the home page to inspect it.

A public issue outcome

On August 2, a technical reply from the project account reviewed a reported lean-ctx instruction-injection bug as both a context-budget and logical-source problem. The maintainer subsequently changed the main branch so Claude Code and Cursor no longer receive a duplicate rules suffix, while other clients receive a bounded suffix.

• The original issue, our reply, the maintainer response, and the implementation remain publicly inspectable.
• At our August 3 review, the issue was open and the combined commit did not have all-green CI. This is a historical observation, not a statement about its current release status.
• The review also identified a follow-up case: the first individual rule alone exceeding the intended limit.

This is evidence that the review method produced an actionable technical observation. It is not a customer testimonial, a traffic claim, or proof that every checker finding is correct.

Change record

1. September 12, 2026Content-preserving conversion and reproducible checks

   Replaced lossy line deduplication with original-text preservation, blocked absent import targets, made report fences safe for Markdown examples, removed the numerical health claim, corrected comparison links, and invalidate generator drafts when inputs change.

2. September 8, 2026Single-source setup as the default

   Start with one AGENTS.md; compare a second file only when it already exists. The current goal is a useful free tool and relevant visitors, without a revenue target.

3. August 15, 2026Command-based starter generator

   Create a local draft from user-provided project commands. Commands are not executed or verified by the generator.

4. August 1, 2026Public checker launched

   Local comparison, findings, generated files, policies, and the repository-audit reservation path.

5. August 2, 2026Real-user measurement added

   Anonymous value-event counts, source attribution, voluntary reservations, and an owner-only operations page. These historical features are not evidence of customers or orders.

6. August 2, 2026Channel quality and operating workflow added

   Tab-session-to-result attribution by channel, daily funnel reporting, lead statuses, retention cleanup, the 18-point checklist, and this build record.

7. August 2, 2026Measurement definitions hardened

   Session-level funnel deduplication, first-touch lead reporting, known-bot filtering, canonical-host redirects, and Asia/Shanghai daily boundaries.

8. August 3, 2026First public technical outcome documented

   A maintainer response and main-branch fix are linked with their verification limits, and content-to-checker CTA measurement was added.

Open items

• Google Search Console and Bing Webmaster Tools are verified; sitemap, crawl, and query changes are reviewed in the owner workflow.
• No payment activation or revenue target is planned in the current visitor-focused work.
• No customer testimonials or outcome claims are published because there is not enough real-user evidence yet.
• Community distribution must be done from an identifiable human or company account, with replies tailored to the discussion.

Corrections and contact

If a checker result, product claim, or instruction-file explanation is wrong, email hello@agentsmdchecker.com. Include a minimal non-sensitive example; do not send repository secrets.