Buyer manual · version 1.2 · September 2026
What to type, what happens, when to use it and what it costs. Keep this page open beside Claude Code, Codex or ChatGPT. Everything here works the same in all three editions, and the pass mark is always computed in code, so nothing talks its way through.
The low-cost first look
speed scan
or
quick scan
A full panel is the gold standard, but it is thorough, and thorough uses a lot of your AI plan. Speed Scan is built for everyday checks and for anyone on a smaller plan.
ONE judge takes every expert's seat for your kind of work, reads it once, scores it seat by seat, and hands you the top five fixes in order of importance. No chairman, no re-judging rounds. The same pass-mark maths still runs on every seat.
Use it on drafts, on small changes, on anything you want a fast honest read on. Then send in the full panel before anything important goes live.
> speed scan Speed Scan: one fast read, 5 expert seats. Finds the biggest fixes. Cannot clear. SPEED SCAN 7.84 / 10 Not ready yet: fix the top items first TOP FIX 1 Add a "notify me when it opens" email box to the offer card. Sold readers hit a closed button and leave. Design Taste 8.6 SHIP Compliance 9.5 SHIP Direct-Response 6.6 REJECT ... A first look, not a verdict. Say "send in the full panel" to clear this work.
Usage, measured
about 123k
tokens for a 5-seat sales page
A full panel run
1.75M to 2.3M
tokens across 2 to 3 rounds
Time
about 3 min
one worker instead of 12 to 18
Tested on the same sales page the full panel had judged: every seat's scan score landed within half a point of the full panel's, with the same top fixes. Roughly a fifteenth to a twentieth of the usage.
A Speed Scan can never mark your work CLEARED. That is locked in the code, not left to the AI. The best a scan can say is "looks ready for the full panel". Only the full panel can clear work.
You never need more than a sentence. Finish a piece of work, then type one of these. Run modes and panel sizes combine: "relentless, full panel" or "speed scan this email" both work.
speed scan/perfection
send in the judges
relentlessuntil it passesone judgefull panel| Your situation | Type this | Why |
|---|---|---|
| I just want a quick honest read | speed scan | Every expert, one pass, a fraction of the usage. |
| I'm on a smaller plan and worried about usage | speed scan, then one judge | Scan to find the big fixes, then one judge for real rounds that can clear it. |
| I finished something and want it judged properly | /perfection | The right panel, fixes made for you, up to 3 rounds. |
| It's a sales page or checkout | /perfection | Money pages get the full panel automatically. |
| It's going live to my whole list tomorrow | relentless | More rounds and stronger judges when it counts. |
| This one has to be the best it can be | until it passes | No cap, with safety brakes and your say at every pause. |
| A scan said "looks ready" | send in the full panel | Only the full panel can clear work. |
The very first time you run it, before judging anything, it asks you these in one message.
Two minutes. Your answers are saved in a file called profile.md inside the skill folder.
From then on the skeptical buyer judges become your customer, the voice judges enforce
your voice, and the design judges check your colours. Edit profile.md any time. Delete it and
the next run asks the four questions again.
Got a DESIGN.md or a project brief in your project folder? The judges use it as the bar
and it takes priority over the built-in standard.
You never pick judges by hand. It reads the work, names the type in one line, and sizes the panel. Say "one judge" or "full panel" to change it.
| Size of the work | Panel | Chairman |
|---|---|---|
| One small piece (one email, one banner, one short script) | 1 elite judge | No |
| A standard build (landing page, app feature, deck, long script) | 3 judges | Yes |
| A money page or a full build (sales page, funnel, checkout) | The full panel | Yes |
| Kind of work | The experts it calls |
|---|---|
| Sales page / funnel | Direct-Response MasterONE JUDGE PICK, Scarred Buyer, Funnel Mechanic, Design Taste Judge, Compliance Skeptic |
| Website / landing page | Conversion CriticONE JUDGE PICK, Design Taste Judge, Technical QA Judge, Target Buyer |
| App / product screen | Product Design JudgeONE JUDGE PICK, UX Flow Judge, Technical QA Judge, Microcopy Judge |
| Marketing email | Email Conversion JudgeONE JUDGE PICK, Brand Compliance Judge, Skeptical Subscriber |
| Video script | Retention JudgeONE JUDGE PICK, Voice Judge, Conversion Judge |
| Deck / presentation | Story JudgeONE JUDGE PICK, Slide Design Judge, Room Judge |
| Anything else | Domain Expert (the best person alive at that craft)ONE JUDGE PICK, Ruthless Skeptic, End-User Proxy |
A Speed Scan always uses the whole list for your kind of work. Hearing every expert is the point of it.
8.0+
Every judge's overall score.
7.0+
Every single thing each judge scores, with no weak spot hiding under a good average.
0
Hard fails. An automatic reject like invented proof, fake scarcity or a dead buy button.
| You see | It means |
|---|---|
| Cleared | Every judge passed it. You still get three watch-outs per judge, the things closest to failing, so a pass is never a blank rubber stamp. |
| Not cleared | The round cap was reached. You get a table of what is still wrong, how bad it is, what a 10 looks like, and the effort to fix it. |
| Speed scan | A first look with a score and the top five fixes. Never a verdict. |
| Stalled | Scores stopped genuinely improving, so it stopped wasting your usage and shows you the stuck findings. |
| Disputed | It checks every finding against your actual work before fixing. If a judge got a fact wrong, it disputes it with proof instead of "fixing" something that was fine. |
It never touches your offer. Prices, guarantees, bonuses and access rules are yours alone. If a judge has a concern about them, it is passed to you word for word, never changed.
All three are in your zip. Install help for each one is on your install page.
Flagship
Type /perfection or say send in the judges. Every judge is its own sealed-off worker and fixes are
made for you between rounds. speed scan works here.
Full power
Say send in the judges. Each judge runs as a separate Codex process that cannot see the others.
If Codex looks unsure, say: "read my perfection skill at ~/.agents/skills/perfection/SKILL.md and judge this."
speed scan runs as one process.
Lite
Paste the Lite document into a fresh chat, then paste your work. Judges go one after another and ChatGPT runs
the pass mark in Python. You apply the fixes and paste the new version back. Type speed scan for the lean version.
Start with speed scan to find the big fixes for a fraction of the usage. When you want real rounds on a budget, add one judge. Save the full panel for work that is about to go live.
Not yet. A scan can never clear work by design. Say send in the full panel for the real verdict.
That happens when profile.md is missing from the skill folder, usually after a fresh reinstall. Answer them again,
or copy your old profile.md back in.
Open profile.md in the skill folder and edit it, or delete it and the next run asks the questions fresh.
That is on purpose. The skill never changes your offer terms. Concerns about them come to you word for word, and the decision is yours.
Just reply "install Node for me" and Codex will do it. The Codex edition needs it to run the judges and the pass mark.
Paste your original install prompt again. It downloads the fresh zip and replaces the old copy. Your current version is listed on the install page.
Reply to your welcome email with what your AI said and we will get you sorted. Nothing you do here can break anything.