A spell gets into the book only after it worked. These four did, on paid client work.
PAGE 01
Quality Gates
Hooks that stop Claude Code before it guesses, builds on stale facts, leaks a key or calls unfinished work done.
- Done gate: no “done” without a real check after the last edit
- Premise check: claims from old notes are verified against live files first
- Git guards: no discarding uncommitted work, no blind amend
- Secret guards: keys never echoed into logs, traces or prompts
- 16 hooks, 6 skills, a canary test for every hook, one-command install and uninstall
PAGE 02
Memory Architecture
A memory your agent reads in the right order, with a guard that flags every note nobody can find again.
- Router read every session, indexes read only when needed
- Index guard: an unlinked note is flagged as it is written
- Decision log: exact words first, interpretation second
- Handoff template, so the next session starts where this one ended
- Weekly hygiene pass for duplicates and stale notes
PAGE 03
Agent Team
Twenty senior specialists, a dispatcher that picks the right one, and a contract for how they hand work over.
- QA, security, DevOps, docs, finance, legal, marketing, growth
- Dispatcher that routes a vague task to the right specialist
- Orchestrator that turns a decision into today's plan
- Handoff contract: agents pass on work, not opinions
- Routing scale: when to answer alone, when to delegate
PAGE 04
AI Influencer Kit
One anchor image, and the same face in every angle, outfit, place and talking head. A photoreal person and a painted mascot.
- One anchor for angles, expressions, outfits, poses, locations
- Double pack: photoreal person and painted mascot
- 70+ tested prompts and a copy-paste file
- Talking heads: front camera, tight 135 mm side, wider three-quarter
- Location lock, so the place does not redraw itself