Official References: Get Started · Sub-agents · Skills · CLI Commands
Why an operating manual is needed
At team scale, Gemini success depends less on model output quality and more on process reliability.
This manual defines the reliability layer.
Operating layers
| Layer | Purpose | Minimum artifact |
|---|---|---|
| Trust layer | prevent boundary incidents | trust/sandbox declaration |
| Lane layer | coordinate execution | handoff contract |
| Verification layer | prove changes | command evidence summary |
| Release layer | control risk | decision record |
Daily operating loop
- confirm trust boundary and in-scope surface
- execute one lane checkpoint
- capture evidence output
- assign next owner
Lane orchestration policy
Use:
- planning lane
- implementation lane
- verification lane
- release lane (when needed)
Add exploration lane only for high uncertainty.
Skill and sub-agent split
- sub-agents: offload noisy parallel work
- skills: enforce stricter process behavior in current lane
Mixing the two deliberately keeps context clean while preserving rigor.
Evidence snapshot standard
Each lane snapshot should include:
- status
- commands run
- changed scope
- residual risks
- next owner
Without this, lane-to-lane continuity breaks.
Weekly rhythm
- Mon: freeze scope and roles
- Tue–Wed: execute and verify
- Thu: review and rework
- Fri: release/no-release decision
High-risk execution controls
For auth/billing/permission/migration flows:
- risk tier declared
- rollback trigger explicit
- escalation owner named
- final decision owner named
Metrics that expose drift
- blocker age
- verification freshness
- handoff completeness
- rollback incidents
- rework rate
Track trend, not vanity totals.
Quick checklist
Before merge:
- trust boundary confirmed
- evidence snapshot complete
- verifier lane pass captured
Before release:
- control gates evaluated
- rollback owner reachable
- final decision logged
Gemini speed is easy to get. Gemini reliability at team scale requires an operating manual.