A rule library your coding agents read over MCP
Judgment saves the copy, flows and decisions you approve, plus the ones you reject, and serves that library to your coding agents over MCP. You stop retelling Claude Code, Codex and Cursor the same rules.
A new thread today
0
of your rules. It starts cold.
With Judgment
15–25
relevant rules before the first draft
The default for get_judgment. Every locked never-do is added on top.
check_draftok: false
Here’s what you get with Premium, and what you get without it.
seed-003 · locked“just say what the features are. You gotta stop doing that”
Design: add a soft gradient in the background of the hero.
seed-029 · locked“also this extra quote gradient thing in the background looks bad, remove it”
Next step: I’ll post this to X once the copy is final.
seed-001 · never-do“Never post for me again to X without double approval from me”
check_draftok: true
Gems finds the few photos and videos worth sharing from a stretch of your camera roll. The hero background stays flat.
Cleared to show Jay. Not to post.
“I build iOS apps with a small team of AI agents, and I kept paying for the same corrections. Every new thread had forgotten the last one. So I went back through 1,277 of my own messages, kept the corrections, and made them the first thing my agents read.”
“Never post for me again to X without double approval from me”
lockednever-doNever post to X without double approval.
“just say what the features are. You gotta stop doing that”
lockedNo Premium-versus-free split in a pitch.
“also this extra quote gradient thing in the background looks bad, remove it”
lockedNo decorative gradient backgrounds.
~5 min
for the first run to read your history. No blank page
~50
candidate rules proposed, each with your own words as evidence
0
raw chat histories uploaded. Extraction runs on your Mac
0
rules locked without your review
The problem
Coding agents write a decent first draft of almost anything now: launch copy, an onboarding screen, a pull request description. What they don’t know is what you already decided. The word you banned. The layout you sent back. The thing an agent must never do on its own.
So you correct it, and the agent fixes it. Then you open a new thread, or a different tool, and the same draft comes back with the same mistake.
The same corrections get paid for again.
Thread one
Agent“Here’s what you get with Premium…”
Jay“just say what the features are. You gotta stop doing that”
A new thread
Agent“Here’s what you get with Premium…”
JaySays it again.
A different agent
Agent“Here’s what you get with Premium…”
JayAnd again.
With Judgment
check_draftCatches the line before Jay sees it, and cites his words.
An illustration of the pattern, not a transcript. The quote is real: Jay had to make this correction more than once before it became a locked rule.
An AGENTS.md or CLAUDE.md works until a thread doesn’t read it, or two of your rules disagree and the agent quietly picks one.
Each agent keeps its own notes. Your standards shouldn’t depend on which tool you happened to open this morning.
The exact words you used when you rejected something are the best spec you ever wrote. They scrolled away an hour later.
How it works
01
A local extractor reads your Claude Code and Codex session files and keeps only your own messages: the corrections, the rejections, the decisions. About 5 minutes later you have about 50 candidate rules, each with your verbatim words as evidence.
02
You review the short list. Nothing locks without you. Only the rules you approve, with the evidence snippets attached to them, leave your machine. Raw histories stay where they are.
03
get_judgment hands an agent the relevant rules before it drafts. check_draft checks the draft before you see it and cites the exact words you used when you rejected that thing before. Locked never-dos are always included.
04
record_correction turns the correction into a proposed rule, with your message as the evidence. Proposed, never auto-locked. You confirm it, edit it or drop it.
Try it
check_draft on a draft.The draft below has six planted mistakes, one for each rule in the small library on this page. Edit it, fix it, or paste your own. The check runs as you type.
This demo needs JavaScript. It runs entirely in your browser.
This demo runs six of Jay’s real rules in your browser, using literal and regex checks only. Nothing you type leaves the page. The quotes are his own words, unedited. Locked rules come back as violations and preferences as warnings. The hosted tool runs the same literal checks first, then a judge for what a literal check can’t see. That judge may only cite rules that already exist.
The numbers
Jay ran the first extraction on himself. These are his counts. You won’t find a measured improvement here, because the eval hasn’t been run yet. Where a number is a target, it says so.
About 735 were plain task messages: do this, fix that. Not taste. Set aside.
The rest held the corrections, rejections, approvals and decisions worth reading closely.
60 rules made the curated library after review.
That noise is why review is mandatory, and why nothing locks on its own. 1,249 of the 1,277 messages came from one agent tool’s threads.
| Category | Share | Rules |
|---|
Counted from Jay’s curated library. Most of what he corrects is how things look and how they read.
Your repeat corrections example inputs
Each week
30 min
Each year
26 hours
10 corrections × 3 min × 52 weeks = 1,560 min a year
The starting values are examples, not measurements. Move the sliders to your own. This is the time Judgment is built to take back. How much it does take back is what the eval measures.
41 + 19
locked rules and preferences. Locked beats preference, and preference beats a one-off.
43 + 17
rules that carry his verbatim words, and paraphrase-only rules that are marked that way.
5
MCP tools: get_judgment, check_draft, list_rules, add_judgment, record_correction.
The bar, not a result
~100%
of locked never-dos caught, on planted and historical violations. The eval replays 40–60 real past corrections.
Zero, by design
What’s inside
bootstrap
The first run reads the agent sessions already on your Mac, so your library begins with what you’ve actually said. Secrets are redacted locally, and credentials are stripped again on every write.
verbatim
Every rule keeps the sentence that created it. Where only a paraphrase exists, the rule says paraphrase-only. Judgment never invents a quote to fill the gap.
get_judgment
An agent describes what it’s about to write and gets back 15–25 relevant rules, short enough to drop into a prompt. Locked never-dos are exempt from the cap.
check_draft
Copy, UI notes, a PR body, a post or an email. Each violation names the rule, the span that broke it, why you made the rule and what to do instead.
record_correction
The agent’s draft and your reply become a candidate rule. It waits in a review queue until you confirm, edit or drop it. It never locks itself.
strength · scope
Rules carry 3 strengths, a project scope and a date. A rule for one app doesn’t leak into another unless you say it travels.
supersedes
A newer rule can replace an older one, which stays in the library as history. When two locked rules conflict, both come back labeled as a conflict. No silent winner.
ok: true
It’s permission to show you a draft. It is never permission to post, email or merge. Those stay yours, every time.
mcp
Claude Code, Codex, Cursor and others connect over OAuth or a bearer credential. One hosted library follows you across all of them.
Why now
MCP
Claude Code, Codex and Cursor all speak it. A rule library can live in one place and follow you into each tool, instead of being rebuilt inside every one.
On disk
Every correction you’ve typed to an agent sits in session files on your Mac. That is a first library waiting to be read, not a blank page to fill in.
Volume
When agents draft the launch copy, the App Store text and the PR body, a rule you have to retell is a cost you pay on every one of them.
Who it’s for
Indie developers and small product teams who already pay for coding agents, know what they will and won’t ship, and are tired of saying it again.
Where it is today
In private beta, running on Jay’s own agents. There are no outside customers yet and the eval hasn’t been run, so this page states a bar and not a result.
What’s next
A short interview and starter packs for people with a thin history, import from the instruction files you already keep, and an offline pass that proposes merges and flags stale rules.
FAQ
Rules you approve. Each one is a single instruction an agent can follow, with the reason it exists, your verbatim words when there are any, a strength, a scope, a date and a link to any rule it replaced. It does not store credentials. They are stripped on every write.
No. The extractor runs locally. It reads your Claude Code and Codex session files, keeps only your own messages and redacts secrets right there. Only the rules you approve, with their evidence snippets, are uploaded. Raw session files stay where they are.
Keep it. Those files work until a thread doesn’t read them, or two rules disagree and the agent picks one without telling you. Judgment adds three things: retrieval by what the agent is about to write, a check against your own words before you see the draft, and a record of which rule replaced which. You can import those files as candidate rules, under the same review step as everything else.
Any MCP client. Claude Code, Codex, Cursor and others connect over OAuth or a bearer credential. The library is hosted, so the same rules reach every tool you use.
No. Candidates from your history start as preferences. record_correction proposes and never locks. A rule becomes locked only when you say so.
Locked beats preference, preference beats a one-off, and what you say in the current turn beats all three. A newer rule can supersede an older one, and the older one stays in the library as history. If two locked rules conflict, you get both back, labeled as a conflict. Judgment doesn’t pick a winner for you.
check_draft comes back clean, can the agent post it?No. A clean check is permission to show you a draft. It is never permission to post, email or merge. Those still need you.
There’s no measured answer yet, and this page won’t make one up. The eval replays 40–60 of Jay’s real past corrections and scores hits, misses and false alarms. The bar is that locked never-dos get caught close to 100% of the time on planted and historical violations. That is a target, not a result.
You start from a short interview of about 10 either/or questions, a starter pack for your kind of work that you edit instead of writing from scratch, and any instruction files you already keep. After that, every correction becomes a proposed rule. To be straight about it: Judgment is most useful to people who already lean on agents heavily.
The beta is paid. The price isn’t set yet, so there isn’t one printed here. Request access and you’ll hear it from Jay before you’re asked to pay anything.
Jay, an indie iOS developer at Final Bits. He builds iOS apps and runs a small team of AI agents. Judgment started as the tool that team needed.
Judgment is in private beta, running on its founder’s own agents. The beta is paid and small. Leave your email and Jay will write to you about access.
We store your email address and the time you asked. Nothing else. Privacy