Show, Don't Tell
Anyone can draw a process diagram. Below are three real artifacts from my own work — one for each step of build, check, measure. Every build includes the review record and the run-cost memo; the ledger comes with measured operation.
Check · A real review round
The Review That Checked This Website
Before this site went live in October 2026, I had it reviewed by an AI model from a different company than the one I build with. The reviewer was told to try to prove the pages wrong, to check every claim against its source, and — just as important — not to invent problems to look useful.
The first round found five problems that mattered. None was a typo. Each was the kind of thing a business owner would have read as a promise.
Found · 1
The site said "Most clients start with the audit." The audit was brand new, so there was no client history behind "most."
What changed
Rewritten as a recommendation: "I recommend starting with the audit."
Found · 2
"Every build independently checked" sat next to a list of past projects, so it read as a claim about all of them.
What changed
Made it a commitment about new work: "independent review included in every build."
Found · 3
The homepage implied a monthly results readout came with every build. It is a separate monthly service.
What changed
Tied the readout to that service, with its price and minimum stated.
Found · 4
The booking page every button pointed to still described a service I no longer offer.
What changed
Booking page description rewritten, then re-checked as a visitor sees it.
Found · 5
The audit page carried five terms — payment timing, when the credit window starts, NDA, time commitment, scope — that had never been agreed.
What changed
Each term put to a yes/no decision, approved, recorded, and made consistent on every page.
Then a second round, on the corrected text
Fixes are new writing, and new writing has new mistakes. So the same reviewer checked that each fix had actually been made and looked for anything the fixes broke. Verdict: ready, no new problems. It also caught an error in my own notes — I had miscounted some past reviews — which didn't change anything on the site but did change the record. That's the point: the check is only worth something if it can come back with bad news.
Measure · A real ledger entry
An Experiment That Failed — On the Record
I run my own businesses on a ledger: every change gets a hypothesis, a number, a date to read it, and a rule for what happens if it fails. Here is one that failed, from the launch of two of my books in July 2026.
The rule and the experiment
- The rule, written July 9 — the day before launch
- No advertising for the book until it has at least 10 ratings, averaging 4.2 or better, with at least 7 from strangers. Get there with free review copies, not ads.
- The experiment
- Free review copies of two of my books offered through a review-copy platform, from July 10.
- Ledger target
- 15 or more verified reviews per book (each had 3).
Graded July 23, 13 days in
FAILED
- Book one: 1 of 25 reader slots claimed, 0 reviews.
- Book two: 0 of 25 claimed, 0 reviews.
The reviews each book already had came from my own network, not the platform. Rather than wait for the campaign to run out on August 21, I dropped the platform early. The ads stayed off, as the rule required, and effort moved to readers I can reach directly.
Why show a failure? Because a measurement that can only ever say "it worked" isn't a measurement. On a client project, this is what the monthly keep / change / kill call is built on.
Build · A real run-cost memo
What It Costs to Keep Running
An AI system has a monthly bill, and you should know it before you sign off. This is the cost section from the handover notes for the tax-litigation assessment tool I built for a law firm, as written at handover.
Hosting
Static hosting on a free tier.
Server proxy (keeps the AI key off the browser)
About $0–5/month at low-to-moderate use; scales to zero when idle.
AI model usage
Pay per use: about $0.003 per 1,000 input tokens and $0.015 per 1,000 output tokens at the time.
One typical assessment
About $0.05–$0.15.
These are estimates made at handover, not a bill. Hosting and AI usage are billed to the client directly by the providers, so the real numbers are always visible.
On Your Project
The same three artifacts, about your system instead of mine.
The review record
What the independent reviewer found, and what I changed. Included with every build and in the audit.
The run-cost memo
What it costs each month to operate, and what would change that. Included with every build.
The ledger
If you choose measured operation: success and stop rules written before launch, graded each month. See measured operation.