macOS app · runs 100% locally · launching soon

Know in 10 seconds if your prompts are good enough.

Prompt Janitor scans every AGENTS.md and CLAUDE.md on your Mac, grades them A–F against the industry's own standards, and flags what's rotting before your agents trip on it.

Diagnosis free forever Runs fully offline macOS Sonoma 14+ Early-bird pricing locked: $19

What a clean prompt file gives you back

Prompt health isn't cosmetic. Every defect taxes tokens, turns, and your own time, on every single run.

Token economy
+36.8k tokens

Stop paying the defect tax

In our first controlled runs, one missing example cost an average of 36,759 extra tokens per task. Same task, same agent, one defect apart. Every prompt defect is a recurring bill: it charges you again on every run, in every repo. Finding and fixing them is the cheapest optimization you haven't done.

Human involvement

Less babysitting after the run

In the same runs, the fixed prompt left 0.4 fewer major issues per task for a human to clean up. The point of an agent is that you review once, not line by line.

Fewer round-trips

Agents converge faster

Ambiguous context makes agents wander: 0.8 extra turns per task with the defective prompt. Clear roles, examples, and output contracts get to done sooner.

Trends & analytics

Watch the waste shrink

Health history and per-project trends ship in the app. Next up: an impact view that translates every fixed defect into tokens and turns saved.

Impact view on the roadmap
Chat analysis

Learn from your real sessions

Mining your actual agent transcripts to show where a prompt file cost you money, and what changed after the fix. Grounded in your usage, not our lab.

On the roadmap

Benchmark figures come from our first N=5 controlled runs and are not yet statistically significant. Read the methodology.

Evidence, not vibes

We measure what bad prompts actually cost.

We run controlled benchmarks: the same coding task, the same agent, one prompt defect apart. Then we count the damage.

+36k
tokens burned per task
with one defective prompt
+0.8
extra agent turns
to finish the same task
−0.4
major review issues
after the prompt was fixed

Early numbers from our first controlled runs (N=5), not yet statistically significant, and we say so. The full powered benchmark runs next, and we're publishing everything, methodology included.

Read the methodology →

Visibility you've never had

Until now there was no single place to see how good your prompts actually are. Prompt Janitor is that layer.

Auto-scan every 6hA–F gradesBest-practice rulesSource-cited issuesLocal-firstCustom rulesRegression alertsWeekly digest
A Health scores

Every prompt, graded A to F

One health score per file, rolled up per project. Watch grades rise and fall as you edit, so you never have to guess whether a prompt is actually any good.

Karpathy Anthropic

Catches what's rotting

Stale model names, contradictory instructions, missing examples, walls of text. Every issue is cited to its source (Anthropic, OpenAI, or the practitioners who wrote the playbook) and comes with a suggested fix.

Your rule Standards

Your standards, enforced

Start from trusted rule packs, then write your own in plain English. “Never name a specific model version.” Done, and checked on every scan, every file.

Calm by design

Quiet by default

Scans in the background on your schedule. A glance from the menu bar, a calm weekly digest, and alerts only when something regresses. Never naggy.

Scan. Grade. Treat.

Diagnosis is free forever. Treatment is what you pay for.

Scan

Point it at your projects. It finds every prompt file (CLAUDE.md, AGENTS.md, .cursorrules) and rescans on a schedule.

Grade

Each file gets an A–F health grade against source-cited standards from Anthropic, OpenAI, and the practitioners who wrote the playbook.

Treat

Pro rewrites the weak parts with AI: apply with a backup, one-click undo, and an optional git branch so changes stay reviewable.

Custom rules, plain English

House rules live right alongside the built-ins: type the intent, pick a severity, done.

No hard-coded model namesKarpathy
Always mention the deploy commandCustom
Must define a clear roleAnthropic

Grouped by project & source

Filter your whole prompt estate by repo, file type, or which guidebook flagged the issue.

AnthropicOpenAIKarpathyCustom
Aweb-app6 files
Dapi-worker2 critical

Built for everyone with a CLAUDE.md

If you write prompts for agents, you need to know they're holding up.

Solo developers

Keep your one CLAUDE.md sharp without thinking about it.

Eng teams & leads

One view of every repo's prompt health across the org.

AI engineers

Hold your agent prompts to a measurable standard.

Agencies

Audit client prompt files and prove the quality.

Indie hackers

Ship faster knowing your prompts won't quietly rot.

Anyone, really

Got a prompt file on disk? Now you can see how good it is.

Why we're building this
“Prompt files are infrastructure.
Nobody inspects them.”

Your agents read these files on every single run, yet there's no linter, no review, no grade. We think diagnosis should be free, for everyone, forever. Treatment is what you pay for.

Read the manifesto →

Diagnosis free. Treatment paid.

We grade your prompt files against the industry's own standards: free, unlimited, on your machine, with your compute. Grading against YOUR standards, and fixing anything: that's Pro.

Free · Diagnosis
$0
Free forever · no scan caps
  • Unlimited scanning: scheduling, watch mode, notifications, history & trends. No scan caps, ever
  • All deterministic fact rules, with source-cited findings, never hidden or blurred
  • Built-in 25-standard AI catalog evaluation, free when you bring compute (local Ollama or your own API key)
  • Standards updates keep flowing to free users

Launching soon: the waitlist gets the download first. No payment, no account.

Pro · Treatment
$19$30
Early-bird pricing · one-time

Everything in Free, plus:

  • AI rewrites + Apply fix / Auto-fix: backup, undo, optional git branch
  • Custom rules in plain English: your standards, enforced on every scan
  • Starter template packs: A-grade CLAUDE.md, AGENTS.md & .cursorrules per stack
  • Bonus: Prompt-File Field Guide (PDF): the 25 standards, explained
  • 12 months of feature updates

$19 is the launch sale price; it goes to $30 afterwards. Waitlist members lock in the sale price.
This pricing isn't definitive and may still change before launch.
One-time purchase: perpetual license + 12 months of updates · $29/yr optional renewal, never required.

30-day grade-up guarantee. If your prompt-file health doesn't go up a full letter grade within 30 days, full refund.

Frequently asked questions

Privacy, compatibility, and how grading actually works.

Soon. Prompt Janitor is in pre-launch: join the waitlist and you'll get the download link the moment it ships, with the launch sale price ($19 instead of $30) locked in. Pricing isn't definitive yet and may change before launch.
You get one confirmation email right away, and one email when the app launches. That's the whole campaign: no drip sequences, and you can unsubscribe any time.
Prompt and agent-instruction files in the folders you choose: AGENTS.md, CLAUDE.md, .cursorrules and more.
No. Scanning and grading run entirely on your Mac. Nothing is uploaded. Prompt Janitor is local-first by design.
Built-in rule packs distilled from Anthropic and OpenAI guidance and respected practitioners, plus any custom rules you write yourself.
Yes, in plain English, with a severity level. They're checked on every scan, right next to the built-ins.
On a schedule you set: hourly, every 6 hours, daily, on file save (watch mode), or manually whenever you click.
Diagnosis is free, forever: scanning, scheduling, watch mode, notifications, history & trends (no scan caps, ever), plus every deterministic fact rule with cited findings. The built-in 25-standard AI catalog evaluation is free too when you bring your own compute (local Ollama or your API key), and findings are never hidden or blurred. Treatment is Pro: grading against your standards, AI rewrites, and applying fixes.
No. Pro is a one-time purchase: perpetual license + 12 months of feature updates. After that there's a $29/yr optional renewal for continued updates, never required to keep using the app.
Checkout runs through Polar, our merchant of record. Right after purchase, Polar emails you a license key. Paste it into Settings → License and Pro unlocks. Keys are verified offline; no account needed.
If your prompt-file health doesn't go up a full letter grade within 30 days, full refund.
Only when you apply a fix, and it snapshots a backup first, with one-click undo and an optional dedicated git branch so changes stay reviewable.
It's pre-launch. Join the waitlist and you'll be scanning within a minute of the launch email. No account, no sign-up.