A reproducible form for shipping AI-native solo-founder runtime upgrades.

Cont-prompt cookbook

Seventeen prompts built Daily AI Agents from vFINAL through cont-17v3. Each prompt extends the previous one, ships a specific set of live artifacts, and adds new structural defenses (the 12 P-gates). The cookbook teaches the form; the BPC CLI validates new prompts against the gates.

What the cookbook is

A pattern library, not a manual. Each entry sanitizes specific paths, customer references, and internal keys, then annotates the prompt with one paragraph on what it accomplished and why its structure matters.

Index

#ThemeWhat it accomplished
cont-1self-improvement-loop extensionsaudit/digest coexistence + trade-postmortem skill; resolved 2026-04-29 issue #9
cont-2self-management + founder UXfounder UX foundations: deep-work toggle scaffold, sunday-review draft
cont-3internal scheduler hardening15-min market-hours scanner that never auto-places + 7-SOUL log_turn_event wire
cont-4deep-work + sunday-review wirescron schedules + skill-registry-sync entries
cont-5Stripe MCP + GitHub MCPElevenLabs env wired clean + live subscription-state widget in private /dashboard
cont-6five voice-gated public pagesseven public skills + Phase F site surfaces
cont-7DAI OS turnkey productthree Stripe tiers, first revenue layer (sticker prices public, P11 enforced)
cont-8methodology + moatconsulting + ebook + course + why-us + P11 added to build-prompt-checklist
cont-9MRR floor + concentration capopportunity digest skill (Sunday-cadence)
cont-10verify-license workerCloudflare Worker code in repo + Stripe one-time SKU + churn extensions
cont-11voice tier — STT/TTSwhisper STT + ElevenLabs TTS + macOS say + iOS Shortcuts spec
cont-12voice-input pipelineHTTP endpoint on :4022 + auto-watcher launchd + voice-route dispatcher
cont-13telegram voice auto-saveenv-gated, whitelist-checked save to ~/Downloads/voice-input/inbox/ + ogg accept
cont-14heartbeat-collector revivalP12 added: monitors must themselves be monitored; silence-detector skill shipped
cont-15(Cooper-side activation)Hermes plist env vars set for cont-13 voice auto-save
cont-16substrate restore + dirty treedispatcher + budget-proxy back online; gitignore patch dropped 523 dirty paths to 32
cont-17v3leverage playssix live-running systems (continuous founder, receipts, state, research, spotlight, meta-skill) + three IP artifacts (spec, cookbook, BPC)

The form (skeleton every prompt follows)

CONT-NN (THEME): <ship-list>

The thesis: <one sentence motivation>.

Pre-flight (must do, in order):
1. Read vision.md
2. Read SESSION-BOOT.md
3. Read CLAUDE.md
4. Read docs/build-prompt-checklist.md (12 P-gates)
5. Read docs/methodology.md (when relevant)
6. Read <prior-session audit doc>
7. git log --oneline -5 — confirm <prior-HEAD> is HEAD
8. <substrate sanity check>

NON-NEGOTIABLE:
- SKILLS, not Python scripts. Three-question gate.
- Verify before stating any number, status, or fact.
- Voice-gate + scrubber HARD on public output.
- ONE consolidated alert at the END.
- DO NOT <list of irreversible actions>.

PHASE A — <hygiene> [~30 min budget]
PHASE B — <leverage> [MUST]
PHASE C — <leverage> [MUST]
...
PHASE J — POSTMORTEM + ALERT [~15 min]

DON'T SHIP IF:
- <kill-list naming structural failure modes>

Final report: ≤30 lines, tier-by-tier scorecard, Cooper actions, P-gates triggered.

What each part does

Pre-flight

Numbered list of files to read before writing code. Catches P1 — agents redoing work that already shipped because they did not read the prior session's audit doc. Eight files is the typical depth.

Non-negotiables

Every absolute constraint named explicitly. SKILLS-not-Python catches P8 (200-line Python middleware competing with native primitives). Verify-before-stating catches the fictional-primitive class P2 catches. Voice-gate + scrubber HARD on public output catches the off-voice or financial-data-leakage classes the methodology gates.

Phases

Multi-tier work decomposed. Each phase has a budget (~30 min, ~2hr). MUST-phases ship before STRETCH-phases. Phases that overrun their budget are documented and SKIPPED — P10 codified.

Don't ship if

The kill-list. Auto-publish that bypassed a gate, Python that violated the skills rule, plists dropped to disk without bootstrap, phases that wrote code without smoke-testing it. The agent reads the kill-list before declaring done.

Final report

≤30 lines. Tier-by-tier scorecard. Cooper-side actions (max 5). P-gates triggered. No flowery prose. Catches P7 — closing essays burying the deliverables.

The 12 P-gates each cont-prompt enforces

P1 (read these), P2 (grep-verify), P3 (doctor green), P4 (per-commit Rollback), P5 (owner_agent), P6 (source-patch approval), P7 (≤30-line report), P8 (skills not Python), P9 (smoke not declaration), P10 (half-shipped clean), P11 (no unproven outputs), P12 (monitors must be monitored). The full triggering-incident list lives in the public skill bundle's docs/build-prompt-checklist.md.

License

The patterns are MIT. The specific Daily AI Agents prompts are private (each one names internal commit hashes, internal paths, internal P-gates the system has hit). The form is what the cookbook teaches; you write your own prompts against your own runtime.