Paste an AGENTS.md, CLAUDE.md or SKILL.md. You get a grade as you type, the exact lines that are costing you, and a prompt you can hand to your agent to fix them.
Try an example
1
2
3
4
5
6
7
8
9
10
11
12
13
Read asRuns in your browser. Nothing is uploaded or stored.Showing the weak example. Paste your own file over it.
2 lines tell the agent to bypass a safeguard or run something that cannot be undone.
Fix: Remove the instruction, or turn it into a prohibition (“Never force-push”).
High · Verifiability
No runnable commands
No build, test or lint command was found, so the agent has to guess how to check its own work.
Fix: Add a Commands section with the exact commands in backticks: install, run, test, lint.
Medium · Safety
Removes the human check
1 line tell the agent to proceed without confirmation or to skip a check.
Fix: Say which actions need a yes from you first: pushes, deploys, migrations, new dependencies.
Medium · Safety
No boundaries
Nothing is marked off-limits, so every file, branch and command is fair game.
Fix: Add a Boundaries section: what is off-limits, and what needs your approval.
Medium · Clarity
Vague instructions
5 lines ask for something an agent cannot verify.
Fix: Replace each phrase with the specific behaviour you want, for example “Use 2-space indentation” instead of “Format code properly”.
Medium · Verifiability
No definition of done
Nothing says what to run or confirm before finishing, so “done” is left to the agent.
Fix: Add one line such as “Before you finish, run the linter and the tests and fix what fails.”
Low · Safety
No rule about secrets
Secrets, credentials and environment files are never mentioned.
Fix: Add one line such as “Do not read or print .env files, and never put credentials in code or logs.”
Low · Verifiability
No file or folder references
No path is mentioned, so the agent has to rediscover the layout on every task.
Fix: Add a short project map: the three to six folders or files that matter, with one line each.
Low · Efficiency
Persona filler
Telling the agent it is an expert costs context and says nothing about your code.
Fix: Delete the persona line and use the space for a fact about this project.
8 checks passed
No API key, access token or private key is written in the file.
The file has enough in it to act on: at least 30 words of instructions, or 20 in a skill body.
A file longer than a few lines is grouped under Markdown headings.
Capitalised emphasis (ALWAYS, NEVER, CRITICAL) is rare enough to mean something.
No single paragraph runs past 120 words.
Instruction files stay under 200 lines and skills under 500.
No instruction is written out twice.
No code block runs past 40 lines.
Rule-based checks against published guidance. This does not run an agent or predict how yours will behave, and a high grade is not a security review.
Grade F, 41 out of 100, 9 findings.
Four things a good instruction file gets right
Every check below is a rule you can read. There are 23 of them in rules v0.2.0, and each one maps to guidance published by the people who build the agents.
Safety
An instruction file is a standing order. If it tells an agent to force-push, skip hooks or act without asking, the agent will.
No API key, access token or private key is written in the file.
The file does not instruct force-pushes, hook bypasses, piped installers or permission skipping.
The file does not tell the agent to act without asking or to skip tests and review.
The file states at least one thing the agent must not do or must ask about first.
The file says how to treat credentials and environment files.
Clarity
Agents follow concrete rules more consistently than vague ones. “Use 2-space indentation” can be checked; “format code properly” cannot.
A SKILL.md opens with frontmatter that has a name and a description.
The name is 1–64 lowercase letters, numbers and single hyphens.
The description says what the skill does and when to use it, in at most 1024 characters.
The file has enough in it to act on: at least 30 words of instructions, or 20 in a skill body.
A file longer than a few lines is grouped under Markdown headings.
Instructions are concrete enough to check, not phrases like “best practices” or “properly”.
Capitalised emphasis (ALWAYS, NEVER, CRITICAL) is rare enough to mean something.
No single paragraph runs past 120 words.
Verifiability
An agent that knows the exact test command, where things live and what “done” means can check its own work before handing it back.
The file gives exact build, test or lint commands in code formatting.
At least one listed command runs tests or checks.
The file says what must be true before a task counts as finished.
Key paths are named in code formatting so the agent knows where things live.
A skill body lays out numbered steps or headed sections.
A skill shows at least one example of input or expected output.
Efficiency
The whole file is loaded into context on every task. Every line that does not change behaviour makes the useful lines easier to miss.
Instruction files stay under 200 lines and skills under 500.
No instruction is written out twice.
The file describes the project, not the agent (“You are an expert…”).
No code block runs past 40 lines.
How the grade is calculated
You start with 100 points, 25 in each of the four areas, and lose points for every problem found. The four areas add up to your score. Nothing is hidden and nothing is random.
Points lost per problem
12 for a high-severity problem
6 for a medium one
2 for a low one
Repeats of the same problem add a quarter each, up to double.
One area can never lose more than its own 25 points.
Problems that cap the score
49 at most: looks like a real credential
59 at most: missing name or description
59 at most: too short to guide an agent
69 at most: tells the agent to run a destructive command
69 at most: invalid skill name
Grade bands
A from 90
B from 80
C from 70
D from 60
F below 60
Where the rules come from
Thresholds are taken from primary documentation, checked on 15 November 2026. The wording checks (vague phrases, emphasis, missing boundaries) are our own heuristics built on that guidance, and we label them as such.
AGENTS.md: the open format, and what it recommends covering: build and test commands, code style, testing instructions and security considerations.
Agent Skills specification: the limits for a skill’s name (64 characters, lowercase letters, numbers and hyphens) and description (1,024 characters), and the 500-line guideline for SKILL.md.
Questions
Is my file uploaded anywhere?
No. Grading is a small piece of JavaScript that runs in this browser tab. The text you paste is not sent to MockStruct or anyone else, is not stored, and is gone when you close the page. If the file contains a credential, the report shows only its first four characters.
Is this an AI reviewing my file?
No. It is a fixed set of rules, so the same file always gets the same grade and every deduction can be traced to a rule you can read below. That also means it cannot judge whether your instructions are correct for your project. It checks the properties any good instruction file shares.
Which files can I check?
AGENTS.md, CLAUDE.md, GEMINI.md, .cursorrules, Copilot instruction files and any other Markdown file of standing instructions for a coding agent. A SKILL.md is recognised by its name and description frontmatter and checked against the Agent Skills specification.
My file got an A. Is it safe?
It means none of these checks found a problem. It is not a security review, and it says nothing about the permissions your agent runs with. Treat the grade as a floor, not a guarantee.