Primary Takeaway
The real trick is not one magic prompt. It is a routing table plus skills: expensive/tasteful model for orchestration, cheap/strong independent worker for bounded subtasks, and explicit report-back/verification rules so the parent agent does not blindly trust the child agent.
Visible File Inventory
| Global CLAUDE.md | High for shown lines; incomplete file | Personal preferences, TS/tooling defaults, command/package-manager rules, model-routing matrix, rules for delegating GPT-5.5 via Codex. |
|---|---|---|
| skills/codex-review/SKILL.md | High for most visible sections; still partial | Frontmatter, review purpose, workflow, command-shape section, review-prompt guidance, reporting-back rules. |
| skills/codex-implementation/SKILL.md | Partial | Implementation delegation skill; instructs Codex to produce patches/reports, not commit/push/deploy/global-config changes unless asked. |
| skills/codex-computer-use/SKILL.md | Partial top section visible | Routes local app verification, browser automation, simulators, screenshots, app launching, and runtime inspection to Codex. |
| skills/html-planning/SKILL.md | Name visible only | Appears in sidebar; transcript later references HTML plans generated for implementation work. |
| skills/use-railway/SKILL.md | Name visible only | Appears in sidebar; no readable body recovered. |
Shown directlyPartially OCR’dNot complete enough for verbatim reconstruction
Recovered Pattern
- Keep Fable/Claude focused on taste, orchestration, decomposition, and final judgment.
- Use Codex/GPT-5.5 for cheaper independent review, mechanical implementation, codebase investigation, computer-use verification, and long-running UI checks.
- Teach Claude how to call Codex through explicit skills so the expensive model does not rediscover command syntax every time.
- When Claude workflows cannot directly select GPT-5.5, spawn a lightweight Claude wrapper whose job is to write a self-contained Codex prompt, execute it, and return structured output.
- Use labels in workflow UI so it is obvious when a displayed Claude wrapper is really delegating to GPT-5.5 underneath.
- Use worktree isolation for parallel implementation agents to avoid edit collisions.
- Keep human-facing reports honest: verify Codex findings before relaying them, and separate confirmed issues from unverified suggestions.
Timestamped Map
| 00:00-02:45 | Frames Fable 5 as a different operating mode: it shines on long end-to-end work, not just harder single prompts. |
|---|---|
| 02:45-07:30 | Shows volume of Lakebed PR work and argues the model can carry multi-hour implementation/review loops. |
| 07:30-11:30 | Cost-control thesis: run Fable on high effort, avoid ultra-expensive settings, and route token-hungry work to cheaper models/tools. |
| 11:30-12:20 | Global CLAUDE.md walkthrough begins: personal preferences, command rules, package manager rules, and tech stack defaults. |
| 12:20-19:45 | Model-routing policy: GPT-5.5 via Codex for bulk/mechanical/review/investigation/computer-use work; Fable/Opus for taste-sensitive work. |
| 20:20-22:20 | Skill walkthrough: codex-review is shown in the most detail, including workflow, command shapes, review prompt, and reporting-back rules. |
| 22:20-24:10 | More skills: codex-implementation and codex-computer-use are shown partially; he edits the computer-use trigger language live. |
| 24:45-31:00 | Prompt example: review 16 open Lakebed PRs, categorize merge-readiness, and use workflows/subagents to divide the work. |
| 31:00-43:15 | Execution loop: model triages PRs, proposes workstreams, lands/rewrites work, verifies, and produces HTML plans; emphasis is on orchestration and review loops. |
Implementation For Your Stack
- Create a model-routing section in your own AGENTS/CLAUDE instructions.
- Start with three narrow skills: codex-review, codex-implementation, codex-computer-use.
- Put hard boundaries into every delegation skill: no commits, pushes, deploys, global config edits, real-account actions, or destructive work unless explicitly requested.
- Require artifact reports from delegated runs, then verify key claims before presenting them.
- Use isolated worktrees for parallel implementation agents.
- Publish large agent-run plans as durable HTML outputs so they are easy to review on phone or desktop.
OCR Notes
I extracted frames from the 11:15-24:55 portion of the video and ran targeted OCR over the editor pane. The most legible recoveries were the global routing section, most of the visible codex-review skill, and the top sections of codex-implementation and codex-computer-use. Sidebar-only files were recorded as names, not reconstructed bodies.
Brief visible anchors: # Personal Preferences, name: codex-review, name: codex-computer-use, # Codex Implementation.