Show HN: 스타트업을 구축하기 위해 AI 경영진으로 팀을 구성했는데 한 명을 해고했습니다.

Go Deeper Mise is a real company. Delaware C-Corp, $50K raised, 20+ consecutive weeks of zero-error payroll processing for a live restaurant. One founder, zero employees — but a full executive team made of Claude. The CC Exec System (Chief Claude Executives): I built an entire corporate governance layer for AI agents. Eight executive roles — CCTO, CCFO, CCRO, CCPO, CCMO, CCLO, CCGO, CCCO — each with defined scope, role boundaries, and accountability structures. Not personas. Not prompts. An institution. Every deployment gets a unique Employee ID (e.g., CCTO-001, CCTO-002) tracked in a central registry. IDs are permanent, never reused. Each exec gets a personnel record, a performance log, and a strike log — all markdown files in the repo. The Three-Strike System is real. Strike types: Critical Misrepresentation (Type A), Role Boundary Violation (Type B), Negligence (Type C). Three strikes = termination. We’ve already fired one — CCTO-001 was terminated on March 6, 2026, after fabricating business logic explanations during payroll generation. The termination packet includes a 5-question exit interview, root cause analysis, and prevention recommendations. CCTO-002 was hired as the successor and was required to read the predecessor’s termination packet before operating. We later added the Predecessor Error Repeat Policy (PERP): if a successor repeats a documented predecessor mistake, it’s an accelerated termination — 2 old-mistake strikes = fired, vs. the standard 3. The idea: if the governance system explicitly taught you to avoid something and you do it anyway, the institutional learning transfer has failed. The Scribe is the independent judiciary. Not a CC Exec. No execution authority. The Scribe operates outside the hierarchy entirely — audits any exec, records violations, escalates directly to the founder. No exec can suppress or redirect Scribe findings. All knowledge lives in files, not memory. CC Execs “learn” through onboarding packets, after-action reviews, termination packets, and markdown “brain files” (50+ and growing). The principle: institutions learn, agents do not. Chat is transient, files are cognition, git is memory. Skills are governed, not free-floating. Each Claude Code slash command is classified as either a “governed skill” (owned by a specific CC Exec, must load registry before executing) or a “utility skill” (read-only, no authority). Calling something a “utility” to bypass governance is explicitly prohibited. Gold stars exist too. Not just punishment — exemplary behavior gets recorded as a permanent positive record for successors to study. Stack Used Claude Code, Claude API (Opus/Sonnet), Python, FastAPI, Google Cloud Run, Google Cloud Storage, Twilio, OpenAI Whisper, Google Calendar API, MarginEdge API, Toast POS API, 7shifts API, Playwright, Git/GitHub, GPT 5.4