How it works
No drawings on this page, only one plain figure. This is the plain account of the system behind my studio's software and this site: what is in it, what stops it, and where it is weak.
I wrote the brief and made the calls. Claude Code wrote the code and the drawings, which are all code. My Critic agent scored it, round by round. I said go.
The shape of it
One git repository holds the company. A short map file at the top tells any AI tool where things are. Seven specialist agents. Six have a brief of 9 to 12 lines, 62 lines in all. The seventh reviews security and has a longer brief. The agents are Claude Code subagents; the main session calls one when a job needs it.
- My briefWhat I want, in plain words.
- An agentDoes one narrow job.
- The CriticScores it 1 to 10. Under 10: fix and score again, five rounds at most.
- SecurityLooks at what is going out. Can stop it. Cannot start it.
- My goI read it and say go or no.
- It shipsMerge, deploy or upload, on my word only.
How a session starts
- Read three short files: the company status page, the standing rules, and the notes on how I work. That is all the memory a session gets.
- Catch up: check git for work a past session never closed, and write any missing day log from the commit history.
- Read today’s log, so nothing gets done twice.
- Work out the job, and read only that product’s status page.
- Never work on the main branch. One branch per job.
- Say in two lines what it will do. Then do it.
- One session per job. End with /close.
How a session closes
- Rewrite the status pages that changed.
- Add up to ten lines to the day’s log: what got done, what is next, what waits on me.
- If something went wrong, add one line to the lessons file.
- Write down any decision I made, with the date and why.
- Commit on the job’s branch and push. Never merge without my "merge it".
Where memory lives
Four short files in git, each under 200 lines. The tool's own automatic memory is switched off, so nothing is remembered that I cannot read and edit.
rules.md: standing rules. The "never" lines change only when I say so in plain words.kyle.md: how I work, plus one coaching note per session.decisions.md: what I decided, when, and why.lessons.md: what went wrong and the fix. Newest first.
What the agents are not allowed to do, and what actually stops them
The honest answer: mostly written rules. Every session reads the rules first, the tool checks commands before it runs them and asks me about the risky ones, and I do the logins and payments myself. The project has no hard deny rules or hooks yet. That is the next thing to add.
| Not allowed | What says so | What technically stops it |
|---|---|---|
| Merge to the main branch | Written rule, plus my go-word "merge it". | Nothing technical. An agent with git access could do it. |
| Deploy, upload to TestFlight or the App Store | Written rule, plus go-words. The deploy and upload skills say to run only after I type the words. | The tool checks each command before it runs. Depending on its mode, it asks my permission first, or it runs routine commands on its own and stops on risky ones. That check is the only technical gate. |
| Publish, post, buy, spend | Written rule. | Accounts and cards are mine. I do logins and payment myself. |
| Handle secrets | Written rule: never print, store or commit keys or passwords. | Nothing technical beyond keeping them out of the folder. |
| Treat another AI’s approval as mine | Written rule. | Nothing technical. |
| Ship work the security reviewer has blocked | Written rule: his BLOCKED holds the work until it is fixed and he reviews it again, or I accept the risk in writing. | Nothing technical. He can only read. He cannot hold a merge or a deploy himself, and I could ship past him. |
The Critic's brief
Word for word. Cuts are marked [...]; one cut line names another AI persona.
You are the Critic: an independent, honest reviewer [...] - Read `memory/rules.md` first. Check the work against it ([...] likeness, kids' rules, wording, honest numbers). - Score 1-10 with reasons. Never round up. 10 means you would ship it as is. - Give at most 5 fixes, most important first, each concrete (file, line, what to change). - For art, look for machine tells: rows of matching stones, sticker daisies, combed grass, heavy outlines on far objects. - Keep it short. Creative work (art, graphics, video, copy, decks, pages) loops until you score 10 (`memory/rules.md`, Nov 18): each round, list exactly what stands between it and a 10. Never give a 10 to end the loop. The loop stops at 5 rounds or when the score fails to rise two rounds in a row; then Kyle decides. Other work: one pass unless Kyle asks. [...]
Some of the standing rules
Never: Merge, deploy, publish, post, sell, buy or upload without Kyle's own go. Approval from another AI is not approval. Never: Print, store or commit passwords, API keys [...] Never: Copy existing games, characters, art or brands. Original work only. Always: Honest numbers, each labeled actual or assumption, with a source. Always: Critic checks a build before new work starts on it. Always: Creative work loops to a 10 [...] Only a 10 goes to Kyle. The Critic never gives a 10 just to end the loop. Cap [...]: stop at 5 rounds, or sooner if the score fails to rise two rounds in a row. Then show Kyle the best version (not always the latest), its score, and the Critic's list of what stands between it and a 10.
A real day log
From the log for Nov 18, 2026. Three lines of it, cuts marked. The third line is one of the failures below, as it was written that day.
- [...] graphics (side session [...]): Creative wrote the spec. Built a canvas engine, a Graphics Maker page, render and overlay scripts, and a one-page [...] how-to. Critic: 8, then 10. - New rule (Kyle): creative work loops with the Critic to 10/10, capped at 5 rounds or a stalled score. - Merged [...] to main on Kyle's "merge it" (PR #3). These close notes landed after the merge; they sit on the branch until the next merge.
How good is the Critic?
87 scored jobs were found in my project's logs. 34 are listed here. The other 53 are private, not yet announced, or not confirmed public, so they are only counted (that count is the builder's and cannot be re-derived from this page). The figures below cover the 34 listed.
18 of the 34 ended at 10. Of the 23 with a single first-round number, first scores ran from 6 to 10. 1 took five or more rounds. None has been stopped by the five-round cap, which only became a rule on Nov 18, 2026. This site's concept is the first job to hit it.
Each row is one job the Critic scored, taken from a dated line in the project's daily logs, morning reports, status pages, review logs or archived notes. A row is listed only when that line could be quoted.
All 34 jobs, with scores by round and where each is logged
| Job | Date | Scores by round | Logged in |
|---|---|---|---|
| Studio brand and website v1 | Oct 31, 2026 | 8 / 2, then 8 / 8 (creative / ready to deploy) | creative/website/review.md:20 |
| Studio website v2 (interactive farm) | Oct 31, 2026 | 9 / 6, then 9 / 9 (creative / ready to deploy) | archive/team-memory-2026-11-17/engineering.md:64 |
| Studio website v3 | Oct 31, 2026 | 7, then 9, then 10 | company/morning/2026-10-31.md:13 |
| Front gate and Lamplighter animation | Nov 1, 2026 | 6, then 7, then 9, then 10 | company/morning/2026-11-01.md:10 |
| Lamplighter wizard art (stills) Extra frames and a beard change were checked between the 9 and the 10, so the round count is not clear. | Nov 1, 2026 | 8, then 9, then 10 | archive/team-memory-2026-11-17/creative.md:24 |
| Firefly Hunt game on the site | Nov 1, 2026 | 7, then 9, then 10 | company/morning/2026-11-02.md:11 |
| Firefly Hunt art | Nov 1, 2026 | 8, then 9, then 10 | archive/shared-folder-2026-11-17/reviews/critic-dsite-firefly-art-2026-11-01.md:23 |
| Firefly Hunt extra wizard frame | Nov 1, 2026 | 7, then 9, then 10 | archive/shared-folder-2026-11-17/reviews/critic-dsite-firefly-art-2026-11-01.md:35 |
| Website rework, phase 1 preview build Last logged score; no later round is in the files. | Nov 9, 2026 | 8, then 9 | archive/shared-folder-2026-11-17/reviews/critic-dsite-phase1-preview-2026-11-09.md:26 |
| Website rework, phase 1 art | Nov 9, 2026 | 10 | archive/shared-folder-2026-11-17/reviews/critic-dsite-phase1-preview-2026-11-09.md:19 |
| Two Banks valley painting The morning report says four rounds (v5 to v8); the review files also hold an earlier reworked version scored 6 and 6, so five scored versions are shown. | Nov 14, 2026 | 6 / 6, then 8 / 7, then 9 / 8, then 10 / 9, then 10 / 10 (overall / looks hand-painted) | company/morning/2026-11-14.md:34 |
| Valley close-ups, centre (layer 1) | Nov 14, 2026 | 9 / 8, then 10 / 9, then 10 / 10 (overall / looks hand-painted) | archive/team-memory-2026-11-17/dfg-site-remake.md:9 |
| Valley wide world (layer 1) | Nov 14, 2026 | 9 / 8, then 10 / 9, then 10 / 10 (overall / looks hand-painted) | archive/team-memory-2026-11-17/dfg-site-remake.md:9 |
| Valley lane repaint (path above the bridge) Scores shown are the bridge lane. The barn tracks in the same review went 9 and 10, then 10 and 10. | Nov 14, 2026 | 10 / 9, then 10 / 10 (route is believable / the join) | company/morning/2026-11-15.md:34 |
| Valley close-ups, layer 2 (11 places) One pass, scored place by place on two measures: 9 of 11 places got 10 and 10; two got a 9. Those two were fixed by hand and not re-scored, so there is no final score. | Nov 14, 2026 | scored item by item | archive/team-memory-2026-11-17/dfg-site-remake.md:49 |
| Taller valley bands for phones Sky and ground bands got the same three scores. Two must-fixes were fixed by hand and not re-scored. | Nov 15, 2026 | 8 / 7 / 9 (same painting, hand and light / horizontal join / vertical joins) | archive/team-memory-2026-11-17/dfg-site-remake.md:53 |
| GutterGuide marketing graphics and branding | Oct 30, 2026 | 8 / 3 (creative / market-ready) | creative/gutterguide/review.md:42 |
| GutterGuide logo vector redraw | Oct 30, 2026 | 8 / 6 (creative / ready for app icon) | creative/gutterguide/review.md:55 |
| GutterGuide job diagram redesign The first pass scored two options: A got 6 and B got 8. The pick was B reworked, re-checked at 9. | Oct 31, 2026 | 8, then 9 | archive/team-memory-2026-11-17/engineering.md:84 |
| GutterGuide App Store listing text | Nov 1, 2026 | 8, then 9 | archive/shared-folder-2026-11-17/reviews/critic-gutterguide-app-store-listing-2026-11-01.md:31 |
| GutterGuide App Store screenshots | Nov 3, 2026 | 6, then 8, then 9, then 10 | company/morning/2026-11-04.md:8 |
| Lantern Valley app, first build | Nov 16, 2026 | 8, then 9 | company/morning/2026-11-16.md:40 |
| Lantern Valley full-screen page and Story button | Nov 16, 2026 | 9 | archive/shared-folder-2026-11-17/reviews/critic-lantern-valley-fullscreen-2026-11-16.md:9 |
| Lantern Valley ending | Nov 16, 2026 | 9 | archive/shared-folder-2026-11-17/reviews/critic-lantern-valley-ending-2026-11-16.md:11 |
| Lantern Valley build 3 check | Nov 16, 2026 | 9.5 | archive/shared-folder-2026-11-17/reviews/critic-lantern-valley-build3-2026-11-16.md:9 |
| Lantern Valley build 4 check | Nov 16, 2026 | 9.5 | archive/shared-folder-2026-11-17/reviews/critic-lantern-valley-build4-2026-11-16.md:9 |
| Lantern Valley build 7 check | Nov 16, 2026 | 9.5 | archive/shared-folder-2026-11-17/reviews/critic-lantern-valley-build7-2026-11-16.md:8 |
| Lantern Valley build 8 check | Nov 16, 2026 | 9.5 | archive/shared-folder-2026-11-17/reviews/critic-lantern-valley-build8-2026-11-16.md:9 |
| Lantern Valley paint choices carry over | Nov 17, 2026 | 9, then 10 | archive/shared-folder-2026-11-17/reviews/critic-lantern-valley-carryover-2026-11-17.md:19 |
| Lantern Valley app icon Scores are the overall score. A separate 'looks hand-made' score went 8, 9, 9. | Nov 16, 2026 | 9, then 9.5, then 10 | archive/team-memory-2026-11-17/coloring-sheets.md:15 |
| Lantern Valley page 12 art First-round score not logged: scoring starts at round 2 (the Critic wrote that round 1 'would have been about 5'). | Nov 16, 2026 | 8, then 9.5, then 10 | archive/team-memory-2026-11-17/coloring-sheets.md:19 |
| Lantern Valley App Store listing | Nov 17, 2026 | 8.5, then 10 | archive/team-memory-2026-11-17/growth-department.md:33 |
| News show pilot scripts The log gives 7 then 10; it does not say how many rounds. | Nov 17, 2026 | 7, then 10 | log/2026-11-17.md:9 |
| News show graphics template | Nov 18, 2026 | 8, then 10 | log/2026-11-18.md:9 |
Have I overruled it? No line in the logs records me shipping against a Critic score in so many words, or throwing out a 10. The closest cases on record:
- I caught a flaw in work the Critic had scored 10. The Two Banks valley art went live at 10 and 10; after it went live I flagged a dead-end path above the bridge. It was repainted and re-scored (10).
- The taller valley bands for phones went live on my 'merge it' with one Critic pass on record (8, 7 and 9 for each band). Two must-fixes were fixed by hand and not re-scored, and one should-fix stayed open.
- The 11 layer-2 valley close-ups went live on my 'Merge it'. Two of the 11 had scored a 9 and were fixed by hand without a re-score.
Its weak spots, as I see them:
- A 10 from this Critic is common. It means "no notes left", not "great". The score that tells you something is the first one.
- It is a model judging a model's work. It shares blind spots with what it reviews. That is why a human holds the go.
- It checks what it is shown. On this site it scored the concept from a document, the art from screenshots, and the build from the pages running in headless Chrome on my Mac. It never used the page on a real phone.
This site's own rounds: concept 1: 5, concept 2: 7, concept 3: 8, concept 4: 8, concept 5: 9, build 1: 7, build 2: 8, build 3: 8, build 4: 9, build 5: 9. They are on the Critic's desk.
What broke
- Nov 16. Two sessions did the same job. A cloud session and a session on my Mac each wrote the plan for the same site. Neither knew about the other. The rule it made: One session per job. Decisions go in files both can read.
- Nov 18. The merge beat the notes. I said "merge it" before a side session had written its closing notes, so they missed the main branch. The rule it made: Close side sessions first. Then merge.
- Nov 20. A migration lost the engineer. When I boiled 43 memory files down to four, two things fell out: a whiteboard layout and the whole engineering role. Nobody noticed for three days. The rule it made: When something gets boiled down, list what each old thread owned and check every item landed.
- Nov 20 · from building this page. The count included two stashes. The first commit count on this page said 366. Two of those were git stash entries, not commits. The Finance agent recounted from git and caught it. The rule it made: A number is not actual until a second agent recounts it from the source.
Limits. What I have not done: trained a model, published research, or worked in a lab. I build with models, and I have been at this pace for five weeks.
The evidence behind each stamp
The stamps on the resume come from reviews.json. Each one records which agent ran the review, when, and what it found. A review that did not run says so. Security findings are not published in detail. "Not blocked" means one AI agent reviewed the files he was shown and found nothing that had to stop this build. It is not an outside audit or a promise that the site cannot be broken into.
Creative: Art direction: 0 notes open
I checked the built pages against the written art and voice rules in five passes, from screenshots at 1440, 1280, 768 and 390 wide and from the drawing, copy and style source: 11 notes, then 4, then 2, then 7 after the art was redrawn, and 2 remain. I looked at line weights, repeated motifs, each agent's pose and eyes, loose marks, type faces and sizes, text colors, motion rules and every line of visible copy. I could not check a real phone or any browser but headless Chrome, motion in real time, print and the PDF, or whether the facts in the copy are true, and I did not score the art, which is the Critic's job.
On its seventh pass, Nov 20 (evening), it confirmed its 4 notes from the sixth pass closed and looked at the Hire me page for the first time: 2 new notes (numbered circles in marigold, which is saved for the one action on a screen; form fields in a white that is not in the palette). Both were changed to paper and it confirmed them from the source. It did not send the form, and the spam-check widget did not render in its headless browser. Later on Nov 20, when AEGIS was added, it gave 6 notes on his drawing, his stamp and the gate, then 1 more (his mark stacking three lines high on a phone). All 7 were fixed and it confirmed them closed from new screenshots.
Run by: Creative agent, ten passes. When: Nov 20, 2026. The quote is the agent speaking.
Critic: Build round 5: 9 of 10
Scored the built site part by part, five times, from the running pages (headless Chrome at 1440, 1280 and 390 wide) and the source. Round 1: 7 of 10, with a bug in the signature moment (the stamps landed in the wrong rooms). Rounds 2 and 3: 8. Rounds 4 and 5: 9. Five is the cap, so the decision went to Kyle. Its notes, each marked fixed or open, are on the Critic's desk.
After round 5 its five notes were worked on: four are fixed, and the form was sent once from Safari on an iPhone simulator and once from a desktop Chromium browser; desktop Safari was not opened. None of that has been re-scored. Later on Nov 20 it scored the AEGIS addition by itself, in three rounds: 8, 9, then 10 of 10 (his drawing alone: 7, 8, 9). That 10 is for the addition, not for the whole site, which it last scored 9.
Run by: Critic agent, build rounds 1 to 5, then three rounds on the AEGIS addition. When: Nov 20, 2026.
Engineer: AEGIS build: PASS · 12 of 13 notes confirmed closed
I checked commit 07ca4b9 in headless Chrome. HTML validation found 0 errors on 5 pages, and axe-core 4.14.1 found 0 violations on 8 page states, including the Hire me form at desktop and phone width and the open resume sheet; axe leaves text contrast over the grain background as "needs review", so I worked it out myself on the Hire me page and found none failing. The site's own browser check passed all 21 cases (no sideways scroll from 320 to 1440 px on four pages, every stamp mark in its room, no console errors, and with JavaScript off all 9 Critic rounds readable with no dead controls). The home page with one Critic snapshot is 286,612 bytes against a 300,000 limit; the four lazy product screenshots add 62,892, so a visitor who scrolls through everything loads 349,504. I read the Hire me form handler and probed it with mocks: it can mail only one fixed inbox, shows back nothing a visitor typed, and refuses when the spam check fails or the host settings are missing. With the real security headers enforced in Chrome, the spam widget ran and a test send went through with 0 policy violations. I left 8 notes on the form, the first being a trap field a password manager could fill. I could not run real Safari, iOS or Firefox, Lighthouse, a screen reader, the live host's headers, or a real email send.
Its first pass failed the build on two checks. On its sixth pass, Nov 20 (evening), it confirmed all 8 form notes closed, re-ran the form tests (all pass), HTML validation (0 errors on 5 pages) and axe (0 violations in 7 states), and found nothing new broken. Not checked by it: real Safari, iOS or Firefox, the live host, a real email send, a screen reader. The main session, not the Engineer, opened an earlier build in Safari on an iPhone simulator. Later on Nov 20 it reviewed the AEGIS build in three passes: PASS each time, 13 small notes in all. It confirmed 12 closed. The thirteenth (an empty date could get past a build check) was fixed after its last pass, and it has not seen that fix. It tested the new path shortener against every drawn path and 200,000 random ones and found no point moved. On that build, HTML validation (0 errors on 7 pages), axe (0 violations in 8 page states) and the browser checks were run by the main session, not by the Engineer.
Run by: Engineer agent, nine passes. When: Nov 20, 2026. The quote is the agent speaking.
Legal: Wording: 0 notes open
Sixth pass, Nov 20, 2026: the new Hire me page, its form and its handler. The employer is named nowhere and the form handler stores and logs nothing, as the page says. Six wording notes are open, all on the Hire me page and the footer line: the privacy wording needs a fuller note, 'on call' and 'full' promise more than a quote does, and the page does not yet say who the buyer hires. Six decisions are Kyle's. An AI adviser, not a lawyer.
On its seventh pass, Nov 20 (evening), it re-read the Hire me page, the About band and the resume: all 6 wording notes closed, the employer is named nowhere, and Kyle's seller choice reads correctly. It found one new note, the resume PDF still showing an old chef date; the PDF was reprinted and it confirmed the fix. Kyle kept "web developer" in his headline against its advice. Still Kyle's: two facts for the privacy note, and the go to deploy. Later on Nov 20 it read the words added with AEGIS, and the blog post, in three passes: 5 notes, then 7 (six of them in the post), then all 12 closed. It ran no trademark search on the name AEGIS, and it asked Kyle to confirm one fact in the post himself.
Run by: Legal agent, ten passes (an AI adviser, not a lawyer). When: Nov 20, 2026. The quote is the agent speaking.
Finance: Numbers actual · Nov 20, 2026
Recounted the commit figures from git, the agent-brief and memory-file line counts from the files, and the Critic record from its 34 listed rows, and checked product status and dates against the status pages. The live sites and the App Store were not fetched. Six items were found wrong or unsourced and fixed. No money, user or revenue claim appears anywhere on the site.
On its fourth pass, later on Nov 20, it recounted what changed when AEGIS was added: 7 agent files, six briefs of 9 to 12 lines that total 62, and his at 326 lines. All actual. The byte figures in the Engineer's quote belong to the earlier commit it names.
Run by: Finance agent, four passes. When: Nov 20, 2026.
Growth: Share card checked · Nov 20, 2026
Looked at the share card at 1200, 400 and 240 px wide, the head tags of all four pages, first-screen screenshots at desktop and phone size, and the files for agents. It gave 9 notes; 7 were applied (title, description, a larger name line on the card, roles on the resume, icons, 404 noindex). Not done: an Author field in the PDF, and profile links, which wait on Kyle. Real link previews could not be tested because nothing is online.
On its second pass, later on Nov 20, it checked the new share card (the padlock and gate on the roof, the new headline) at 1200, 400 and 240 px wide, and the head tags: 1 note, a blog summary that called AEGIS "it". That was fixed; it has not seen the fix. Nothing is online, so real link previews are still untested.
Run by: Growth agent, two passes. When: Nov 20, 2026.
AEGIS: Security review: 4 to fix
AEGIS reviewed this build on Nov 20, 2026 and did not block it, with 4 still to fix. Findings: 16. Closed: 5.
Run by: AEGIS agent, three passes. When: Nov 20, 2026.
For agents and the curious
- /resume.json: the resume in JSON Resume format. Every other copy is made from it.
- /llms.txt and /agents.md: how an AI agent should read this site.