Part two
While we are building
What each of the three weeks looks like, who tests and how, how to report a problem, and what we still do slowly.

§ 07 What the three weeks look like
What the three weeks look like
Three weeks, and you are in the software from the first one. Nobody waits for a reveal.
Fig. 4 — Three weeks, drawn
- Week 1: We build and deploy. You: 2 hrs, kickoff + first look.
- Week 2: You use it, we change it. You: daily use, notes as you go.
- Week 3: Hardening and handover. You: sign-off, route decision.
- Week 1You log in
Something deployed
We build the thinnest version that lets one person complete one real case, and we put it somewhere you can reach. It will be ugly and incomplete. That is the point: you react to a real screen instead of a document.
- Week 2Two or three testers
Real cases, real data
Your people run actual work through it. We fix what breaks, add the fields that turn out to matter, and remove the ones nobody used. Most of the shape of the final thing gets decided here.
- Week 3Wider group
Edges and handover prep
The awkward exceptions, permissions, the Audit trailA durable record of who did what, when, and what the system did in response.Full entry in the glossary →, backups, and the documentation. We also measure against the number you named in phase 01.
What stalls a build
- No test data, or data so sanitised the awkward cases vanish.
- Access requests that sit in a queue. Start those in week zero.
- One reviewer with no delegate, on leave in week two.
- Feedback saved up for a formal review meeting at the end.
What counts as delivered, and what happens if you go quiet: Engagement Terms § 04 — Delivery and acceptance.
§ 08 How testing works
How testing works
Who tests
- The people who do the work, not their managers, and not us.
- Two or three of them. Ten testers produce noise and no owner.
- At least one person who is fast at the current process and one who is new to it. They find different problems.
How to test
- Run one real case from start to finish. Clicking around finds cosmetic issues; finishing a case finds the real ones.
- Bring the cases you dread. The duplicate, the one with a missing field, the one that had to be escalated last month.
- When something is wrong, stop and tell us. Working around it silently teaches us the wrong thing.
- Use real timings. If a step took four minutes, say four minutes.
On data
§ 09 How to give feedback
How to give feedback
Feedback is the whole mechanism. Nothing else in the process replaces it, so it is worth doing well. Six habits cover almost everything.
- Screenshot first
- A screenshot of the whole window beats three sentences of description. Include the address bar and any error text. Crops hide the context we need.
- What you did, what happened, what you expected
- Three short lines. This one format removes most of the back and forth.
- One issue per note
- Five things in one message means four of them get lost.
- Say whether it blocks you
- Blocking, annoying or cosmetic. We sequence by that, and you decide the sequence.
- Send it the day you hit it
- A note on Tuesday can be fixed Tuesday. The same note in Friday's summary costs a week.
- Say when something is fine
- Silence reads as approval. It usually is not. A one-line 'this part works now' is genuinely useful.
Note
§ 10 What we need from you
What we need from you
Off-the-shelf software makes you adapt to it. Bespoke software adapts to you, but only if you are in the room while it is being cut. The most valuable thing you have is what your experienced people do when the formal process breaks down, and none of that is written anywhere.
What we need from you
- One named owner who can make decisions without convening a committee.
- An honest description of how the work actually happens, including what your experienced people do when the formal process breaks down.
- Real data, or a realistic stand-in. Fake data hides the awkward cases.
- A reaction every week. Twenty minutes is enough.
What you get back
- Software shaped around your workflow, instead of your workflow bent around a generic tool.
- The vocabulary your team already uses, in the interface.
- The awkward exceptions handled, because we saw them while building.
- The repository, the deployment and the documentation.
§ 11 What speed does not excuse
What speed does not excuse
The fair objection to quick delivery is that it hides risk. It can.
Google's 2025 DORA research found that most technology professionals now use AI at work and believe it improves their productivity, while delivery stability gets worse unless the underlying engineering practice is strong. Speed amplifies whatever practice you already have. So the boring parts stay boring.
- Access and identity
- Roles, permissions and secrets handled properly, not bolted on before launch.
- Data handling
- Classification, retention, deletion, and a clear answer to who can see what.
- Backups and recovery
- Backups that have been restored at least once, and a rollback that works.
- Tests and monitoring
- Tests around the parts that would hurt, plus logging and alerts so failures surface before your users report them.
- Documentation
- Enough that a competent stranger, or your future in-house hire, can take it over.
- One accountable human
- Someone answerable for how the software behaves in production. No tooling replaces that.
All of it sits in the Engagement Terms § 07 — Data protection and security, not in a sales conversation.