Make AI coding inherit the team's actual standards.

Video caption

Make AI coding inherit the team's actual standards.

A polished PR can hide an untested route.

Every rule gets an owner, gate and falsifying test.

My rule: A missing run is UNVERIFIED, not a green check.

#EricFieldNotes

Full written post / accessibility read

A mature engineering team has hard-won rules: architecture boundaries, security permissions, test fixtures and product acceptance criteria. Put those into the coding workflow as executable contracts, not a document that each new model run must remember to obey.

If five workers patch five services, one may use a helper that bypasses the tenant guard. A supervising agent can flag the concern, but that message is still advisory unless a trusted test and a scoped service identity make the bypass fail.

For each invariant, record the risky action, pre-call script or permission that constrains it, protected CI fixture, negative control and who reviews exceptions. Keep the gate outside the agent's writable worktree where possible. Include runtime-specific canary results.

Require the actual runtime's denial receipt, fresh CI result and independent post-state readback before merge. Do this because agents can accelerate implementation only when the organization can verify that its own standards still hold.

#EricFieldNotes

Four-beat scene transcript

1. Make AI coding inherit the team's actual standards.

A mature engineering team has hard-won rules: architecture boundaries, security permissions, test fixtures and product acceptance criteria. Put those into the coding workflow as executable contracts, not a document that each new model run must remember to obey.

Visual: An agent needs more than a long instruction file.

2. Agent concurrency multiplies boundary misses.

If five workers patch five services, one may use a helper that bypasses the tenant guard. A supervising agent can flag the concern, but that message is still advisory unless a trusted test and a scoped service identity make the bypass fail.

Visual: A polished PR can hide an untested route.

3. Build a standards-to-evidence map.

For each invariant, record the risky action, pre-call script or permission that constrains it, protected CI fixture, negative control and who reviews exceptions. Keep the gate outside the agent's writable worktree where possible. Include runtime-specific canary results.

Visual: Every rule gets an owner, gate and falsifying test.

4. Review evidence, not assurances.

Require the actual runtime's denial receipt, fresh CI result and independent post-state readback before merge. Do this because agents can accelerate implementation only when the organization can verify that its own standards still hold.

Visual: A missing run is UNVERIFIED, not a green check.

Research and claim limits

• Cursor hooks documentation (S04)
• OpenAI plugin architecture (S06)
• OpenAI Agents SDK/Codex orchestration example (S07)

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