Nicholas Dunzelman

AI coding deployment workshop ยท checked May 19, 2026

A workshop for turning one AI coding run into team adoption evidence.

This is the enablement layer of the OpenAI Codex proof packet: a cleanroom, public-safe session plan for scoping a real repo task, running an AI coding pass, classifying failures, setting rollout gates, and handing the team a receipt they can inspect after the demo ends.

Output

What a team gets

A bounded task, a live run, a failure map, and a manager-readable receipt. The point is not to make AI coding look magical. The point is to decide which task patterns are ready to repeat, which ones need better harnesses, and which ones should stay human-owned for now.

The session is built for the adoption layer: engineering workflow, review quality, tool boundaries, customer feedback, and proof artifacts.

90 minutes

The run

  1. 0-10 Scope card

    Name the repo, task owner, acceptance criteria, disallowed actions, data boundary, rollback path, and verifier.

  2. 10-25 Context load

    Read only the files, commands, dependencies, and product behavior needed to act correctly.

  3. 25-50 Live AI coding run

    Run the agent visibly enough to capture what it inspected, changed, checked, guessed, or refused.

  4. 50-65 Failure map

    Separate bad context, weak verifiers, missing env, permission limits, unclear requirements, and model/tool mismatch.

  5. 65-80 Rollout gate

    Classify the task pattern as ready, needs work, or fail before it becomes a team habit.

  6. 80-90 Receipt and handoff

    Package the diff, checks, proof, risks, replay instructions, and next owner.

Receipts

Artifacts produced

Scope card The task boundary and verifier.
Run receipt What changed, what ran, and what passed.
Failure map Where the workflow, harness, context, or agent broke down.
Rollout checklist The gate before repeating this pattern on a team.
Handoff note A manager-readable summary with proof links and open risks.
Adoption call Ready, needs work, or fail.

Rubric

Readiness gates

Ready The task is explainable, inspectable, reversible, and covered by a verifier that matches the claim.

The receipt makes the outcome legible to an engineer and a manager.

Needs work The task is valuable, but the team lacks a test, fixture, preview, permission rule, or review convention.

The agent can help, but the workflow cannot yet compare runs honestly.

Fail The task exposes secrets, has unknowable acceptance criteria, uses cosmetic proof, or gives the agent an authority it should not have.

A failed run would be hard to detect or roll back.

Triage

Failure map

Route

Where this fits

The submitted OpenAI application already has a build packet. This page adds the workshop/enablement proof that a Codex deployment lane needs: customer-facing structure, technical content shape, product feedback loops, safety boundaries, and repeatable adoption evidence.

Proof brief Deployment memo Proof Ledger inspector Cleanroom repo

Boundary

Cleanroom line

This page is independent from OpenAI. It uses OpenAI and Codex only to name public target-role context. It does not use OpenAI logos, product UI, private systems, screenshots, copied workshop material, or affiliation language.

Sources: AI Deployment Engineer - Codex, OpenAI Interview Guide, and OpenAI Brand.