JournalDAY 42 / TIKTOK

FIELD NOTE / TIKTOK

The patch is beautiful. The product choice is wrong.

The complete written thought and the evidence behind it. The video edition will follow its public release.

Journal September 25, 2026 · TikTok target November 8, 2026

The patch is beautiful. The product choice is wrong.

Video caption

The patch is beautiful. The product choice is wrong. The test asks whether the default role can approve. Reject even elegant output when the decision is absent. #EricFieldNotes

Full written post / accessibility read

In this fictional refund tool, an agent builds a crisp dashboard and tests the happy path. A contested refund has no agreed owner, so the code treats an ordinary support role as final authority. The code works. The decision is wrong.

The agent wrote the implementation and test from the same prompt. Both assumed the support role can approve disputes. The test passes while the organization never made that policy choice. More code review alone may not expose the missing decision.

A product owner names who may approve contested refunds and why. A protected fixture attempts the action as ordinary support and expects denial. The agent can then implement freely inside that boundary, and reviewers compare observed behavior with the recorded choice.

Block release until the authority decision and its negative test exist. Do this because code quality cannot compensate for a product rule nobody actually chose.

#EricFieldNotes

Four-beat scene transcript

1. The patch is beautiful. The product choice is wrong.

In this fictional refund tool, an agent builds a crisp dashboard and tests the happy path. A contested refund has no agreed owner, so the code treats an ordinary support role as final authority. The code works. The decision is wrong.

Visual: Agents can implement an unanswered question with confidence.

2. A green test may share the same assumption.

The agent wrote the implementation and test from the same prompt. Both assumed the support role can approve disputes. The test passes while the organization never made that policy choice. More code review alone may not expose the missing decision.

Visual: The test asks whether the default role can approve.

3. Put the policy on one card first.

A product owner names who may approve contested refunds and why. A protected fixture attempts the action as ordinary support and expects denial. The agent can then implement freely inside that boundary, and reviewers compare observed behavior with the recorded choice.

Visual: Owner, alternative, consequential case and veto.

4. Let agents write code, people own choices.

Block release until the authority decision and its negative test exist. Do this because code quality cannot compensate for a product rule nobody actually chose.

Visual: Reject even elegant output when the decision is absent.

Research and claim limits

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.

More notes from the work ↗