One engineer, one AI, one month.
CreatorStudio — the announcement engine, two Angular apps, the LED sign renderer, three vehicle clients and a ten-document handbook — was built between July 1 and July 31, 2026 by one engineer directing Claude in Claude Code. This page is the measured record of that build: every figure below is computed from the session transcripts, the git history and the test report — nothing is estimated.
The shape of the month
Tokens the models generated per day, and hours the AI was actively working (wall-clock with idle gaps over 10 minutes excluded). The build never paused: even the quietest day moved the product forward.
Generated tokens per day · millions
38.8M tokens of code, tests and documentation generated in total — the peak, Jul 18, was the LED template editor push.
Active build hours per day
~194 hours total — a sustained ~6.3 h/day of AI working time, directed by 1,150 human prompts (~37 a day).
Daily data as a table
| Day (Jul) | Tokens (M) | Active hours |
|---|
Where the compute went
13.0 billion tokens were processed in total. 98.2% of them — 12.76B — were served from the prompt cache at a tenth of the input price; only 38.8M were newly generated. Cost is the API list-price equivalent of the measured usage, per model.
| Model | Responses | Generated | Cache reads | Cost (list-price equivalent) |
|---|
What the AI actually did
Not autocomplete — an agent operating the repository: 19,361 tool calls across the month, reading, editing, running tests and driving a browser against the real apps.
What it bought
The shipped tree, verified — not a prototype. Every deliverable below is in production shape, covered by the same CI gates.
The announcement engine
15.5k lines of TypeScript: journey-state tracking, 33 trigger types + custom rules, the render pipeline, 13-supplier TTS layer with failover, LED sign rendering with FF and Hanover wire protocols.
Two Angular apps
39.8k lines: the CreatorStudio authoring app and the fleet cockpit — plus the 27.5k-line Creator mirror kept parity-tested against the engine.
Vehicle clients & portal
Go, Python and browser players closing the proof-of-play loop; the passenger screen; the documentation portal and this page.
11.7k lines of tests
1,669 cases — engine 738, front-ends 931 — 0 failing, 0 skipped. Engine 93.0% line coverage; parity guards that fail the build if the authoring app and engine drift.
A ten-document handbook
15.7k lines of maintained documentation — architecture, functional spec, security, scaling, test report — drift-checked by CI, with an honest gaps register.
The product story
A 28-slide presenter-ready deck, speaker crib sheet, leaflets and diagrams — all regenerable from the repository.
Efficiency, honestly computed
How these numbers were measured
Sources. Time, tokens and cost come from the 58 Claude Code session transcripts and 124 subagent transcripts of this project — every model response carries exact token usage (input, output, cache read, cache write) and a timestamp. Code size comes from git ls-files line counts of the shipped tree; test and coverage figures from the regenerable TEST-REPORT.md.
Active time is wall-clock across all sessions with idle gaps capped at 10 minutes, measured on the merged timeline so parallel subagents aren't double-counted.
Cost is the API list-price equivalent: the measured tokens priced at each model's published pay-per-token rates (Opus 4.8/Opus 5 at $5/$25 per MTok, Fable 5 at $10/$50, Haiku 4.5 at $1/$5; cache reads at 0.1×, cache writes at 1.25× the input rate). It answers "what would this exact usage have billed" at public pay-per-token rates. The actual out-of-pocket was different — and far smaller: the build ran on a single monthly Claude Code subscription, so the real spend was one subscription month. The $10.4k is quoted because it is the reproducible, scrutiny-proof way to price the compute; the subscription price is a flat-rate bundle and not a cost model.
What's not counted. The engineer's own time — 1,150 prompts of direction, review and product decisions — is the human half of the build and is not in the dollar figure. Line counts are the surviving tree, not lines written: every rewrite, refactor and dead end is inside the $10.4k, not inside the 102k lines.