SOP — Model Release Prompt Audit

1. Purpose

When Anthropic releases a new Claude model version (e.g., Opus 4.8, Sonnet 4.7), automatically trigger a prompt engineering audit to leverage new capabilities and avoid deprecated patterns. This ensures continuous agent refinement without manual reminders.

2. Trigger

Any of: - New Claude model version announced (check Anthropic changelog, model docs) - CEO or chief.staff notices a model version change in the system - Quarterly review (Jan, Apr, Jul, Oct) regardless of whether a new model dropped

3. Participants

Role	Agent	Action
Audit lead	chief.staff	Reviews model docs, identifies pattern changes, runs council review
Executor	project.manager	Applies approved changes across agent contracts
Quality gate	council-review skill	Validates changes don't regress governance
Approver	CEO	Signs off on batch changes

4. Execution Flow

Step 1: Model documentation review (chief.staff, day 1)

Read the new model's: - Migration guide (breaking changes) - Prompting best practices (new patterns) - What's new (capabilities) - Changelog (deprecated features)

Produce a Model Delta Brief documenting: - Universal pattern changes (apply to all agents) - Role-specific optimizations (by agent type) - Deprecated patterns to remove - New capabilities to leverage

Step 2: Common-pattern scan (chief.staff, day 1-2)

Scan all 21 agent contracts for instances of deprecated patterns. Build change list.

Step 3: Council review (chief.staff, day 2-3)

Run council-review skill on proposed changes. Questions: 1. Do changes preserve governance constraints? 2. Any risk of behavioral regression? 3. Are role-specific optimizations appropriate for model tiers?

Step 4: Apply changes (project.manager, day 3-5)

Apply approved changes. Version-bump all modified contracts. Batch by department (Dev → Exec → Rest) per established pattern.

Step 5: Validate (chief.staff, day 5-7)

Spot-check 3-5 agents in live sessions. Verify: - Agents still follow governance rules - "Ask before assuming" behavior holds - QA validators still report all findings - No silent regressions

5. Output

• Model Delta Brief at /Claude/operations/reports/briefings/model-delta-{model-version}-{date}.md
• Updated agent contracts (version bumped)
• Council review archive
• Completion note in CEO Daily Summary

6. Authority

Chief of Staff owns this process. It runs without CEO initiation — triggered automatically by model release events. CEO approves batch changes but does not need to kick off the audit.

7. Standing principles for all model audits

These carry forward across every model release:

1. Remove anti-laziness scaffolding when the new model handles instructions literally
2. Shift prescriptive → declarative as model reasoning improves
3. Protect validator reporting — newer models may self-filter more aggressively
4. Add explicit scope — newer models interpret more literally
5. Agents present 3-4 approaches before executing non-trivial work (dev agents especially)
6. Agents ask before assuming — if not 100% clear, ask clarifying questions
7. RACI cross-references stay current — update if matrix changes

8. Archive

Previous audit plans preserved at /Claude/plans/ with model version in filename.

---

SOP authored by chief.staff (Eva) · 2026-04-20 · CEO approved