Free · private · instant

AGENTS.md checker and CLAUDE.md drift validator

Score one coding-agent instruction file, or compare two files to find commands and guardrails that have drifted apart. Everything runs privately in your browser.

Decide whether you need AGENTS.md, CLAUDE.md, or both Screen six research-defined configuration smells Scan repository instructions for security risks

Step 1

Paste or load the first file

Load .md

# Example project

Next.js dashboard for a small operations team.

## Commands
- `npm run dev`: start locally
- `npm run build`: production build

## Working rules
- Follow the existing component patterns.
- Keep changes scoped to the requested feature.

Compare with another instruction file

Optional: detect one-sided commands and guardrails across AGENTS.md and CLAUDE.md.

Both files are analyzed entirely in your browser and are never uploaded.

Step 2

Instruction health

52/100

Needs context

AGENTS.md52/100

The badge links back to this score summary. It contains no file contents and badge impressions are never counted as users.

Repository purpose

Explains what the project does so the agent can make relevant choices.

Stack and runtime

Names the language, framework, runtime, or package manager.

Runnable commands

Provides copyable setup, build, test, lint, or development commands.

Architecture map

Points to important directories, boundaries, or established patterns.

Validation loop

States which checks must run before work is considered complete.

Change boundaries

Prevents broad rewrites and keeps work aligned with the request.

Safety guardrails

Protects secrets, user changes, migrations, and destructive operations.

Definition of done

Makes the final evidence and handoff explicit.

Focused length

Keeps the root file readable and avoids burying critical rules.

Copy a repair block

Add these missing sections, then replace the placeholders with repository-specific facts.

What gets checked

The facts an unfamiliar agent needs.

A useful root instruction file is not a style guide. It is a compact operating contract: what the repository is, how to navigate it, how to validate work, and which mistakes are expensive.

Executable context

Purpose, stack, architecture, and copyable commands reduce avoidable exploration and command failures.

Repository guardrails

Scope limits, secrets, migrations, and destructive-action rules protect the work already in progress.

Completion evidence

Tests, builds, and a concrete handoff make “done” observable instead of subjective.

AGENTS.md checker FAQ

What the score and drift report mean.

What does the AGENTS.md checker validate?

It looks for repository context, runnable commands, architecture guidance, validation steps, change boundaries, safety guardrails, and a definition of done.

Can it compare AGENTS.md and CLAUDE.md?

Yes. Add a second file to compare copyable commands and explicit guardrails, then review signals that appear in only one file.

Are either of my instruction files uploaded?

No. Scoring and drift detection run entirely in your browser. Repo Agent Kit records only the comparison outcome, never file names or contents.

Does a high score guarantee agent behavior?

No. The score and drift report are practical signals, not a guarantee. Repository-specific instructions still need human review and regular updates.