Official References: Best Practices · Review · Worktrees · Automations
Why an operating manual matters
Most teams can get good one-off outputs from Codex. Fewer teams can keep quality stable for weeks.
The gap is rarely model capability. The gap is missing operating rhythm.
This manual is the “how we run” layer between prompts and production.
Operating model in one table
| Time horizon | Goal | Main artifact | Owner |
|---|---|---|---|
| Daily | keep lane progress reliable | evidence snapshot | lane owner |
| Weekly | ship reviewable scope safely | merge packet | review/release owner |
| Release window | protect production | release decision record | final decision owner |
| Incident window | recover predictably | rollback + incident note | on-call + escalation owner |
If one row is missing, reliability drops fast.
Lane contract (non-negotiable)
Each lane handoff must include five fields:
- goal status
- evidence
- diff scope
- risk note
- next owner
No exceptions for “small” changes if teamwork is involved.
Daily cadence (15-minute structure)
Start-of-day
- confirm lane owner
- confirm top blocker
- confirm today’s verification target
Mid-day checkpoint
- attach fresh command output
- confirm scope drift has not occurred
- flag blocked status early
End-of-day snapshot
- done/partial/blocked
- command outputs
- changed directories
- residual risks
- next owner
Short cadence beats long status reports.
Weekly merge pipeline
- scope freeze
- lane execution with daily snapshots
- verifier lane rerun
- branch-vs-base review
- merge decision record
If work is “almost ready” every Friday, scope freeze is probably weak.
Artifact bundle you should standardize
For each merge candidate, keep this bundle:
- implementation summary
- verification summary
- merge packet
- rollback note
- release communication draft
When incidents happen, this bundle determines recovery speed.
Quality metrics that actually help
Track a small set consistently:
- review lead time
- rework rate after review
- blocker age
- rollback frequency
- evidence freshness compliance
Avoid vanity metrics like prompt count or token usage without outcome context.
Escalation policy
When a blocker repeats across two loops:
- mark
blocked - attach evidence
- assign new owner
- set deadline
- rerun verification before unblocking
Unowned blockers are more dangerous than failing checks.
Incident-time mode switch
If production risk materializes:
- pause feature lane merges
- shift to containment lane
- execute rollback trigger policy
- document decisions in real time
- resume normal lanes only after stabilization
Don’t invent incident workflow during the incident.
Quarterly calibration routine
Every quarter, review:
- which gates catch real defects
- which artifacts nobody uses
- where handoff quality still degrades
- which checks are too slow for daily cadence
Then simplify ruthlessly. An operating manual should evolve with evidence.
Quick checklist
Before weekly merge:
- lane contracts complete
- verifier lane pass captured
- merge packet ready
Before production release:
- release decision record complete
- rollback owner reachable
- communication package prepared
Codex gives you acceleration. The operating manual gives you repeatability.