Tanner Baker builds with AI
I am building this board for real: 45 modules, driven by Claude. See the build.
I direct AI agents to design, build and ship products.
- Background
- 13 years auditing film and TV construction budgets, up to $27M
- Open to
- Quality engineering, product design, business analysis, AI enablement
- Based
- Newnan, Georgia
- ContactHow I work with AI
Details
I set the direction, agents do the building, and nothing ships until I have checked it.
- Role
- AI-assisted product builder
| Sheet | Title | What happened |
|---|---|---|
| G-001 | Cover sheet | This sheet. |
| A-001 | My process | The workflow I use, then one real page built with it as you scroll. |
| R-101 | Read it by role | The evidence sorted for quality, design, analysis and AI work. |
| M-101 | How I work with AI | Planning, workflows, guardrails, and a sign-off that stays with me. |
| A-101 | BracketScene | A live tennis tournament ran on it. Every release must pass 25,035 tests. |
| A-201 | Slam Tournaments | A sold-out event, then a full rebuild with payments and bookings. |
| A-301 | WrapSheet | A budget app in daily use for a crew of 48 to 62. |
| A-401 | Let's Go Pickleball | Every flyer, page and screen for a club, plus a TV network guests can send songs to. |
| A-501 | Pickle Pot | A tournament platform that has run two events. |
| E-601 | PICKLE | An AI assistant I built that asks before it acts. |
| E-701 | Split-flap display | The cover board, built for real with 45 modules. In build. |
| S-801 | Other work | Video, fabrication, analytics and tools. |
| G-002 | About and contact | Who drew this, and how to reach me. |
Next sheet: A-001. Keep scrolling to see my process.
My process
The workflow I use, then one real page built with it.
In practiceOne BracketScene page
- 1ResearchAgents read the competitors. I read their findings.
- 2SpecI write what it must do.
- 3PlanOne agent plans it. A second checks the plan.
- 4BuildAgents build in parallel. One fits the pieces together.
- 5AttackFresh agents attack it from five angles.
- 6VerifyEvery test has to pass.
- 7LiveI check it on the real site, then it ships.
WrapSheetRead it by role
A design systemTwo looks for two jobs, built to WCAG 2.1 AA.BracketScenePD53 of 54cleanAn evaluation set18 trap cases, run three times each, gate every model change on my assistant.PICKLEQE AI
Screens for people on their feetThe club's TV screens, a waiver kiosk and a front-desk view.Let's Go PickleballPD766plansA planning workflowBrief, roadmap, written plans, and a plan check before any work starts.MethodAI
Rules from the jobOvertime, sixth and seventh days, and transfers between sets.WrapSheetBAUp to$27MThirteen years of auditFinding the number that does not tie is the job.AboutQE BA
Go or no-go, with evidenceThe booking ran on staging, and one real charge was refunded before I made the go call.SlamQE BAMore evidence
Research firstCompetitor research: their events, site and reviews.SlamPD BA95standing rulesContext that survivesSkills and handoff notes carry my standards between sessions.MethodAI
A decision logEach owner decision and the option rejected, in writing.BracketSceneBA16testsTests that prove a ruleWrapSheet drafts never reach totals, reports or exports.WrapSheetQE
Brand systemsLogos, type, color and copy rules, applied to every piece.Let's Go PickleballPD73intro classesA process, end to endRegistration, waiver, check-in and follow-up email, automated for a club.Let's Go PickleballBA
Options before decisionsEach screen chosen from several working mockups.BracketScenePD$3feePricing and policyBracketScene's fee model and Slam's booking rules are mine.BracketSceneBAM-101Browser automationPlaywright checks, screenshots and sweeps on every product.MethodQE AI13yearsNumbers leaders act onWeekly cost reports with variance notes for construction leadership.AboutBAM-101Requirements that holdRules, money mechanics, and what is out of scope.MethodBA AI2failuresFailures turned into rulesTwo agent mistakes, how I caught them, and the checks that exist now.MethodAI QEM-101GuardrailsWhat agents may never do without me, and stop rules for overnight runs.MethodAIE-601Human approval, built inMy assistant sends nothing until I approve the exact words.PICKLEAI QEM-101The right model for the taskFaster models for specified work, the strongest for money and judgment.MethodAIHow I work with AI
I run projects the way a department head runs a crew. I decide what gets built and I sign off on it. AI agents do the rest, inside rules I set.
standing rules my agents carry
written plans on BracketScene
isolated agent workspaces
adversarial reviews, plus 20 production sweeps
tests every release must pass
The plan check
All 22 plans were checked against the spec, the context and the code.Result: 3 blockers, 16 warnings, 2 items for Tanner.
As written the plans would not deliver the goal to the cent.
Found before a line of code was written.
Two failures, and the rules they left behind
My assistant answered a question with a phone number it had made up.
How I caught itThe number was not in any source it had read.
The rule nowA check flags every phone number, link and dollar figure no tool returned. 18 trap cases, run three times each, gate any model change. The current model keeps 53 of 54 answers clean.
A background agent wrote demo data to a live site.
How I caught itIt showed up in my sweep of the live site the same day, and I removed it.
The rule nowBackground agents never commit, deploy or call a live site. They work on a local copy, and I move the result.
See the workflow run in My process, sheet A-001.
Details
I sit in the middle. Work goes out to agents and comes back to me.
- Research. Agents read the competitors, the docs and the old code. I read their findings.
- Spec. I write what it must do, what it must never do, and what is out of scope.
- Plan. One agent breaks the spec into tasks. A second agent checks the plan before any code exists.
- Build. Agents build in parallel, tests first, each in its own workspace with its own files.
- Attack. Fresh agents review the result from five angles. What they find becomes a fix or a test.
- Verify. The tests must pass before anything moves on.
- Live. I check the release on the real site before it counts as done.
The spec I wrote
- What it is
- The rules
- What a host sees
- Money mechanics
- What changes for organizations that exist today
- Open questions for Tanner
- What is not in this spec
- How it gets built
One of the plans
wave: 1model_hint: sonnet
autonomous: true
requirements: [SP-09, SP-08, SP-11]
files_modified: 7 files, owned by this plan alone
must_haves: truths the finished work has to make true
- Research, plan, write code and tests, and review each other's work.
- Work through the night on a plan I have approved.
- Decide small things, as long as they record the decision.
- Send any message to a real person. I approve the exact words first.
- Touch money settings, legal text, or anything irreversible on a live site.
- Commit, deploy or call a live site from the background. They work on a local copy.
- Sign up for a new paid service.
Workflows I set up
Planning before building
Every project starts with a brief, a roadmap of phases and a decision log. Each phase gets written plans, and a plan is checked before any work starts.
Context that survives
Handoff notes, reusable skills and a memory of standing rules carry my standards between sessions, so no agent starts cold.
Match the model to the task
Faster models take the well-specified tasks. The strongest model takes money, math and the final integration.
Rejected: one model for everythingAlways within reach
Agent sessions stay running on a home server, and I steer them from my phone. Anything that needs approval arrives as a card I can approve or deny.
This page, for example
Research agents studied award-winning portfolios, animation techniques and hiring advice. I picked the concept. Review agents then took each draft apart, and I chose what to change.
BracketScene
Tournament software for pickleball and tennis directors. A tennis tournament at Amelia Island ran on it, from the first check-in to the final.
Sample players. The line work is live SVG. Live at bracketscene.com
tests in more than 2,200 files. The deploy script runs them all and refuses production if one fails.
database migrations, each applied before its deploy
build phases, each planned, reviewed and verified
design review passes, from 320 to 1920 px wide
How the design was chosen
- 1 Problem
The old site had no front door.
- 2 Options
Three directions, A, B and C, for the same five screens.
- 3 Decision
B's look on every screen. A and C lend only content and layout.
Rejected: A's and C's look. B's styling was the best overall.Rejected: a mix of all three looks. It did not blend. - 4 Result
Approved after four rounds, then built in five phases.
- 6.8:1links and buttons on white
- 5.3:1the faintest text allowed
- 10.6:1live scores on the dark TV
- WCAG 2.1 AA asks 4.5:1
Details
Decisions
One price, shown once
Players see a single number with the $3 fee inside it. Round robins up to 50 players are free.
Rejected: a fee line added at checkoutTwo looks for two jobs
Working screens stay light and plain. TV boards go dark. Status never depends on color alone.
Rejected: one theme stretched to fit bothDone means verified live
A release needs green tests, an adversarial review, and a fresh-browser sweep of production for console, security-policy and accessibility errors.
Rejected: "it passed on my machine"194 test files check data isolation between organizations, and 33 check concurrency.
The AI part. Agents built in parallel from plans I approved, each owning its own files, and fresh agents attacked the work before it shipped. I researched the market, set the price, and picked each screen from several working mockups.
Slam Tournaments
Women's doubles tennis weekends at resorts. The first event sold out at 65 players. I then rebuilt the site for the events that follow.
players at the sold-out Amelia Island event
legacy accounts imported, none asked to reset a password
checks in the production sweep after go-live, all passing
automated tests on the new site
Details
- Audit access, payments and sign-up first
- Stage every change before it goes live
- Every player keeps her account
- Audited the site I inherited before changing anything.
- Rebuilt on Cloudflare Workers and D1: accounts, squads of eight, one checkout for room and entry.
- Moved 155 accounts across with their passwords intact.
- Rehearsed the full booking on staging, then took card payments live.
- Had agents research our main competitor's events, site and reviews.
Decisions
Nobody resets a password
Existing players keep their passwords through the move.
Rejected: a reset email to every playerRehearse the money
The full booking ran on staging first. Then one real charge and refund went through before the doors opened.
Rejected: finding out with a customer's cardCut over by redirect
The old site forwards to the new one, with the rollback steps written down beforehand.
The AI part. Agents audited the old code, built the new site to my specs and reviewed each other's work. I wrote the booking rules, approved every email word for word, and made the go call.
WrapSheet
The budgeting tool I wanted for thirteen years, in daily use for a crew of 48 to 62.
The budgeting tool I wanted for thirteen years. It tracks set budgets and daily labor for a film construction department, and it is in daily use for a crew of 48 to 62.
| Set | Budget | Actual | Remaining | Spent | % |
|---|---|---|---|---|---|
| Int. Precinct, Bullpen | $60,500 | $38,000 | $22,500 | 63% | |
| Ext. Warehouse District | $92,000 | $84,200 | $7,800 | 91% | |
| Int. Evidence Room | $36,000 | $37,500 | -$1,500 | 104% | |
| Int. Captain's Office | $60,000 | $31,700 | $28,300 | 53% | |
| Total | $248,500 | $191,400 | $57,100 | 77% |
Sample note: Evidence Room is over by $1,500. Flag it before Friday.
Details
automated tests
user roles, from read-only to admin
people on the crew at its largest
Drafts cannot touch the budget
A half-entered day is kept out of the math by how the data is stored. A 16-test suite proves drafts never reach totals, reports or exports.
Rejected: a "draft" flag checked in every reportThe AI part. Agents wrote the React and Firebase code and its tests. The rules come from the job: overtime, sixth and seventh days, transfers between sets, and what an auditor needs to see on Friday.
Let's Go Pickleball
Solo marketing for a pickleball and padel club. Every flyer, page, email and screen comes through me, along with the systems that run them.
A sample of the club's flyers and posters. Hover to pause. Live at letsgopickleball.com
The TV roll
A signage system I built with agents runs the club's screens. Guests can add a song from their phone.
Request a Song
Read the four digits off the TV capture.
A working demo. Try it.
custom pages on the club site
endpoints in the class registration system
intro classes scheduledintro classes scheduled through it
automated email journeys, from confirmation to follow-up
Details
One event, fifteen directions
For one tournament I had agents design the same flyer fifteen ways, from casino to brutalist to ticket stub. Then I picked.
You have to be in the room
A request needs the code on the screens. Each guest gets three songs, then a five-minute wait, and explicit tracks never reach the speakers.
Rejected: an open request linkOne dashboard, every screen
Every screen checks in for updates, so a new slide reaches the whole building from one dashboard.
One file, every size
Each event graphic is a single file that renders to a flyer, a TV slide, a story and a feed post at print resolution.
Rejected: redrawing each size by hand- Class registration with waiver signing, member lookup and self-service rebooking.
- A staff dashboard for check-in and attendance, used at the front desk every day.
- A camp portal with card checkout and per-child waivers, and an iPad kiosk for guest waivers.
- A weekly newsletter to about 2,100 people, plus the blog, listings and search markup.
The AI part. Agents built the players, the screen manager and the Spotify integration, and they render every flyer size. The brand rules, the copy and the final pick are mine.
Pickle Pot
A platform for cash-prize pickleball tournaments. Two events have run on it. For the second I added text messaging, and 60 players got their court calls on their phones.
tournaments run on the platform
players at the second event, in 30 teams
texts sent, none failedtexts sent that day, none failed
automated tests behind it
Details
- Divisions, pools and brackets that build themselves from the entries.
- Players report their own scores, and the other team confirms.
- Court assignments go out by text the moment a court opens.
- Partners can split the entry fee at checkout.
The director's keyboard stays quiet
Scores a director types in do not text players. Only scores players report themselves trigger a confirmation text.
Rejected: a text for every score changeThe AI part. Agents built the platform to my plans and reviewed every phase. I designed the format and the event-day flow.
PICKLE
An AI assistant I built that asks before it acts.
An AI assistant I designed and run myself. It reads my mail, calendars and texts, and it asks before it acts.
To: the venue coordinator
Subject: Saturday load-in
We will arrive at 7 with two carts. Is the side door still the best way in?
Checked: 1 time and 0 links, all from your calendar.
trap answers clean: 18 cases, run three times each
automated tests
tools, each one logged when called
Details
It may not make things up
In testing it invented a phone number. I added a check that flags any phone number, link or dollar figure no tool returned, and a set of trap cases that runs before each model change.
Rejected: trusting a better promptSplit-flap display
I am building the cover board for real, so Claude can write to it.
modules, in 3 rows of 15
Details
- Researched the open-source designs and picked the parts.
- The frame, generated by a parametric script.
- A remix of the module with a steel axle.
- Packing every part onto 77 print plates for four Bambu printers.
- The order sheet.
- Wiring the board to Claude.
- The flap module is an open-source design by Adam G Makes.
- The gateway firmware and the companion service are the open-source SplitFlapGateway project.
flaps in each module
print plates across four printers
MCP tools Claude will use to write to the board
The frame is 674 by 267 mm. It is made to stand on a desk, tilted 10 degrees, or to hang on a wall.
The AI part. Agents researched the open-source options, and an independent agent reviewed the plan. I made each choice.
Other work
| Mark | Item | What it is | Status |
|---|---|---|---|
| 01 | Ad tracking audit | Found why a two-location art studio's bookings went unrecorded, and fixed it. | Done |
| 02 | Video studio | A pipeline that renders logo reveals, product demos and social clips from one set of brand tokens. | In use |
| 03 | Film office generators | Scripts that build budget workbooks, status reports, sign-in sheets and labor calendars. | In use |
| 04 | Fandom Foundry | My fabrication shop: scenic builds, props and 3D-printed products, modeled in CAD. | Since 2020 |
| 05 | Split-flap display | The board on the cover, built for real. Sheet E-701. | In build |
| 06 | Wordlings | A reading and spelling game I am building for my son. | In progress |
| 07 | Convention scripts | Automation for event schedules and registration at conventions like Gen Con. | In use |
Details
Who drew this
Agents let me build what a small team would. I check their work the way I check a budget.
More about me
I work in the construction office on film and television sets. Each week I match purchase orders to invoices, reconcile the cards, and tell the people building the sets where the money stands. I have done that since 2013, on budgets up to $27M, for productions including Stranger Things and Ant-Man.
I have run a fabrication shop since 2020, and I have designed flyers, signage and sites for clubs and events along the way.