How I build · Exhibit · updated 8 October 2026
Nine launch stages, one person whose time was scarce, and thirty-six hours to be able to send. Ben set the stages, their order and the one constraint that mattered: his own hours. The AI operating system he built for this search mapped the dependencies at hour zero. Ben made the calls that changed the plan.
Human, then AI, then human. The strategy and the stage order are Ben’s. The dependency analysis, the hole-finding and the time costs are the machine’s. The final rulings are Ben’s again.
The move
Three stages had to be finished the moment an application went out. The other six were needed only when somebody replied, three to seven days later. Sorting every stage by when it is actually needed turned a nine-item blocker list into three.
2 h 25 m of the decision-maker’s time. An application sent without any one of these is worse than one that waits.
About 3 h 40 m to 4 h 40 m, spread across the days between the first send and the first screen. Two exceptions are called out in the holes below.
The chain
At hour zero: a clean draft, measured at two pages. Nothing downstream can start until it closes.
Circular with stage 1; see the first hole. The build is the machine’s. The human part is the domain and one sitemap approval.
The before-state screens stop existing the moment stage 8 runs. Twenty minutes of silent capture now protects the whole case study; see the third hole.
The minimum before launch is a claims check, so the profile never contradicts the resume. The rebuild can follow.
The only stage that runs on other people’s time. It cannot be compressed later, so the asks go out first even though the need comes last.
Split it. Recording an application is a send-time gate and takes minutes. Full scoring is not: hand-pick the first wave and let scoring catch up.
What the plan caught
The resume said a demo video, workflow, collateral and error log were live at the domain. Until they were, the link was a promise that could not be kept, and the site’s content came from the resume, so neither could go first.
Three ways out: ship a thin page carrying all four; or drop the link and add it back later; or keep the link and cut the list to what exists on launch day.
The domain was bought and never pointed. Stage 2 could not start, and the resume link would be dead the moment anyone clicked it.
The cheapest unblock on the board. Do it first.
The tracker’s untouched state, with nothing scored, nothing applied and the go/no-go panel red, is the opening of the case study. Optimising the tracker deletes it for good, and the stage order ran the optimisation first.
Twenty minutes of silent screen capture before anyone touches the tracker. Narration can be re-recorded. The state cannot.
The story bank existed in an earlier working session and never reached the written record, so stage 7 had no source material.
Does not block launch; blocks the week after. Recovering it is the machine’s job.
The scarce resource
Only one person’s hours were priced. The machine’s were not the constraint.
| To be able to send | His time | When | What he is actually doing |
|---|---|---|---|
| Point the domain | 5 min | Night one | One DNS record |
| Capture the before-state | 20 min | Night one | Silent screen recording, no script |
| Resume decisions | 45 min | Day one, morning | Fourteen calls, density first |
| Resume review | 30 min | Day one, afternoon | Read the draft, mark the document |
| Sitemap approval | 30 min | Day one, afternoon | Approve structure, not copy |
| Scoring rulings | 15 min | Day two, morning | Weights and the selection metric |
| Total | 2 h 25 m | By hour 36 | Inside the window with room to spare |
| Before the first reply | His time | Runs on |
|---|---|---|
| Reference asks sent | 30 min | Other people’s calendars, so it starts first |
| Demo recording | 2–3 h | His voice, one sitting |
| LinkedIn claims check, then rebuild | 20 min | After the resume closes |
| Cover letter blocks | 20 min | After the resume closes |
| Interview checklist | 30 min | After the story bank is recovered |
| Total | 3 h 40 m – 4 h 40 m | Spread over seven to nine days |
The handback
The plan ended with the smallest set of decisions that unblocked it, each with its options named. Nothing else was asked.
Two pages tight, or three at the readable layout. This one decision unblocks the resume.
Keep it and ship a thin four-item page, or drop it from the resume and add it back later.
Point the domain and capture the tracker before anyone touches it. Both are his, and both are gone if they wait.
Send them now. They are the one item nobody can compress later.
The rulings
The plan asked
Two pages tight, or three at a readable layout.
Ben ruled
Two pages. The resume was built to that and locked at two.
The plan proposed
A launch split into send time and reply time, with LinkedIn and the demo video left as work in progress.
Ben ruled
Launch is four assets shipping together: the resume, the LinkedIn profile, this site and a short demo video. With any one missing, the other three point at a hole. The launch date became a readiness gate.
The gate left open
A launch date that moved with readiness.
Ben ruled, 8 October 2026
GO LIVE is Thursday 8 October 2026. The three demo videos are recorded that day and the site goes live the same day.