Builds

Builds.

One build measured end to end, then smaller ones, each with what broke or is still open.

Circlewise · Affiliate networkAug 2026

From spreadsheet exports to questions Claude answers from live data.

11 days from plan to production, 18 → 29 Aug 2026
28 tools built; 10 released to the team on day one
599 tests passing at rollout; 610 the next day

Measured from the project record. Sources on the case page.

Read the full case

More builds

More builds, each with its caveat.

Prospecting engine · affiliate network

7 of 8

brands checked on the first live list were mislabelled, then fixed.

Caveat

Scores not yet calibrated with the client’s sales team; 13 active clients still showed as leads in October; the first real use came back with 5 corrections.

11 Aug 2026: some were the client’s own customers. New rule: no brand is called untouched without positive evidence; brand facts moved into dated data files, enforced by a test.
The full client roster (400 → 915 domains) changed 30 labels and found 17 existing clients on the pitch list.
Shipped 30 Sep with 289 passing tests. First server run: all 93 brands two news sites were promoting got a label from dated evidence; scoring cost $1.16 against a $5 cap.

Dated data files · 25 fetch-policy tests

Edtech · Dana’s medical-vocabulary appa medical-vocabulary app

72 nodes

in 3 n8n workflows: Claude writes each scene, FLUX 2 Pro draws it.

Lesson

In the 200-word stress test the Google Sheets quota broke first (426 requests in a minute against 60), and the orchestrator still reported success.

Scope: about 2,600 images, 13 professions, 4 languages; projected at about $91 for the full set.
Prompt caching cut the Claude prompt cost by 80%.
Delivered 24 Mar 2026, about 3 weeks after the contract, with a 17-page system doc and an Apps Script menu.

n8n · Claude · FLUX 2 Pro · Google Sheets · Apps Script

Method demo · a Warsaw personal trainer

3 looks

of one site, live on the same URLs about 9 hours after starting.

Caveat

A demonstration, not approved by the client.

455 Instagram posts turned into a brand brief.
3 verifiers tried to refute each of its 12 key claims; 1 fell.
A cookie switches between the 3 styles.