JournalDAY 44 / INSTAGRAM

FIELD NOTE / INSTAGRAM

Stop scoring only the first patch.

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

Journal September 25, 2026 · Instagram target November 10, 2026

Stop scoring only the first patch.

Video caption

Stop scoring only the first patch.

Technical fundamentals stay visible.

See whether the plan changes.

Agent use is allowed; the standard is higher.

#EricFieldNotes

Full written post / accessibility read

Show the candidate a clean agent-produced diff. The useful interview starts when it looks complete. Can they explain the system it touched, find the unasked business decision, choose an independent test and describe a handoff a teammate could act on? The interviewer supplies the artifact.

Score the spoken system model, a discriminating test proposal, a release decision and a concise handoff. Write strong-answer examples before seeing candidates. Ask one code-reading follow-up about data contracts or failure handling so polished process language cannot substitute for fundamentals. The interviewer supplies the case; the candidate does not implement it.

Show a fictional producer diff whose tests pass, then disclose a consumer that lags two weeks. Ask how old and new events coexist, what replay would demonstrate and who authorizes the compatibility window. A strong answer may pause release, propose dual-version handling and name a business owner. This is a few minutes of reasoning, not homework.

Give each candidate the same questions, artifact and time; let them refer to the agent output they will see at work. Score the judgments the role requires because typing speed no longer captures the whole job. Compare interviewer scores and candidate feedback before operational use. If a practical build is indispensable, make it small, scoped and paid.

#EricFieldNotes

Four-beat scene transcript

1. Stop scoring only the first patch.

Show the candidate a clean agent-produced diff. The useful interview starts when it looks complete. Can they explain the system it touched, find the unasked business decision, choose an independent test and describe a handoff a teammate could act on? The interviewer supplies the artifact.

Visual: The job continues after the agent says done.

2. Use four observable columns.

Score the spoken system model, a discriminating test proposal, a release decision and a concise handoff. Write strong-answer examples before seeing candidates. Ask one code-reading follow-up about data contracts or failure handling so polished process language cannot substitute for fundamentals. The interviewer supplies the case; the candidate does not implement it.

Visual: Technical fundamentals stay visible.

3. Add one fact late.

Show a fictional producer diff whose tests pass, then disclose a consumer that lags two weeks. Ask how old and new events coexist, what replay would demonstrate and who authorizes the compatibility window. A strong answer may pause release, propose dual-version handling and name a business owner. This is a few minutes of reasoning, not homework.

Visual: See whether the plan changes.

4. Test the work you will assign.

Give each candidate the same questions, artifact and time; let them refer to the agent output they will see at work. Score the judgments the role requires because typing speed no longer captures the whole job. Compare interviewer scores and candidate feedback before operational use. If a practical build is indispensable, make it small, scoped and paid.

Visual: Agent use is allowed; the standard is higher.

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 ↗