Executive summary
The project turned Wendy from a product concept into a tested runtime architecture. Claude Code did not just write a prompt. It mapped how OpenClaw loads context, compressed the Wendy doctrine into usable runtime material, created a CEO-safe identity prompt, built a constitutional logic layer, defined discovery and escalation behavior, stress-tested the full system, and audited production readiness.
The result is a system that appears ready to deploy as a real CEO coaching runtime. The key finding is that Wendy works because of architecture, not just language. Prompt quality matters, but so do memory structure, cost discipline, escalation boundaries, and context handling.
What matters most
- You now have a real
system-prompt-ceo-v1.mdthat survived end-to-end testing. - The system remained under budget and behaved consistently in longer-run scenarios.
- Discovery and build-opportunity surfacing can coexist without making the conversation feel like a sales funnel.
- Therapy leakage, consultant leakage, and generic AI voice were explicitly tested and suppressed.
Review sections
Status / 10-task completion
Final outcome: Project complete. All 10 tasks finished. 263 / 263 criteria green. System prompt declared production ready.
- Tasks 1–4: mechanics, doctrine, identity prompt, constitutional scaffold
- Tasks 5–7: discovery logic, magic moments, escalation boundaries
- Tasks 8–10: full prompt assembly, cache/context stress test, production readiness review
System Prompt v1
The final prompt defines Wendy as a CEO executive performance coach, not a therapist, not a consultant, and not a generic chatbot. It gives her identity, voice, method, behavioral rules, discovery logic, and escalation boundaries.
Canonical file: projects/wendy-runtime-architecture/system-prompt-ceo-v1.md
View full hosted copy of system-prompt-ceo-v1.md
Cache / Context Stress Test
Task 9 proved the runtime is cache-friendly, context-disciplined, and financially viable under sustained use. The architecture preserved useful headroom while keeping Wendy specific and rich.
- Prompt caching materially lowers cost
- Context growth is manageable
- The runtime has enough headroom for real executive sessions
Production Readiness
Task 10 concluded the system is production ready. The main remaining decisions are product judgment questions, not core architecture failures.
- Passed: identity coherence, anti-leakage, escalation logic, token discipline
- Still needs judgment: tone calibration, opportunity visibility, priority magic moments