Forensic transcript + screenshot pass

Fable 5 Workflow: visible-file reconstruction

This replaces the thin first pass with a screenshot-backed extraction of Theo's CLAUDE.md and skill setup, plus a practical hardware and stack recommendation for running the workflow locally.

Video checked: 2026-07-06 Primary source: YouTube 8GRmLR__OGQ Evidence: frames, OCR, public screenshot, browser QA

Bottom Line

The setup is a routing layer. Global CLAUDE.md teaches model selection and Codex handoffs; small skills tell Claude when and how to call Codex for review, implementation, and computer-use verification; worktrees and artifact directories make parallelism auditable.

What changed in this pass

Added source corroboration, downloaded the public screenshot, copied frame evidence into the report, re-ran OCR over key frames, and split recovered content by confidence.

What is not recoverable

No full file is visible from byte one to EOF. This report reconstructs the visible structure and behavior rather than pretending the screenshots expose complete files.

What to copy from the idea

Copy the architecture: model routing, explicit skill triggers, artifact folders, worktree isolation, and screenshot validation. Do not blindly copy somebody else's exact local rules.

Evidence Gallery

Images are included so the extraction can be audited without re-scrubbing the video.

Visible File Inventory

FileConfidenceVisibilityRecovered role
~/.claude/CLAUDE.mdHighPartial, but multiple frames plus public screenshotGlobal preferences, model-routing table, routing mechanics, wrapper-agent rules, Codex timeout/worktree/budget notes.
~/.claude/skills/codex-review/SKILL.mdHighMostly visible across dense framesIndependent Codex/GPT-5.5 code review for diffs, commits, PR checkouts, and broad implementation audits.
~/.claude/skills/codex-implementation/SKILL.mdMediumTop and workflow visible, command details partly clippedBounded patch-producing agent: pin state, define scope, run codex exec, inspect diff, verify, report.
~/.claude/skills/codex-computer-use/SKILL.mdHighFrontmatter/top section visible repeatedlyRoutes UI/browser/simulator/screenshots/runtime verification to Codex; excludes ordinary code reading/lint/tests.
~/.claude/skills/html-planning/SKILL.mdLowName only in sidebarLikely used to generate shareable HTML plans; contents not opened in the video segment.
~/.claude/skills/use-railway/SKILL.mdLowName only in sidebarLikely deployment/provider helper; contents not opened in the video segment.

CLAUDE.md Reconstruction

Personal preferences

Visible sections cover TypeScript preferences, command discipline, package manager defaults, preferred app stack, concise code style, and a rule to use Codex/GPT-5.5 when computer use helps.

Model scorecard

The visible table ranks GPT-5.5 as low-cost/high-intelligence/lower-taste, Sonnet as mid-cost/mid-intelligence/good-taste, Opus as better taste/intelligence, and Fable as best taste/intelligence for the author.

Routing rules

Bulk mechanical work goes to GPT-5.5; user-facing UI/copy/API work needs a model with stronger taste; reviews prefer Fable or Opus, with GPT-5.5 available as an independent extra pass.

Codex mechanics

GPT-5.5 is reached through Codex CLI commands and through the codex-* skills. For work not covered by a skill, the visible instruction is to call Codex read-only with a self-contained prompt.

Workflow wrapper

Because workflow model parameters only accept Claude models, a thin Claude wrapper writes the Codex prompt, runs Codex through Bash, and returns structured output.

Operational guardrails

The lower video frames add agent labels, long timeout/background polling, worktree isolation for parallel implementation, and the note that Codex work does not count against Claude workflow token budgets.

Skill Skeletons

codex-review

  • Use only when a second-pass reviewer is useful or the model-selection rubric calls for GPT-5.5.
  • Target can be uncommitted changes, a base-branch diff, a commit, PR checkout, implementation, or specific files.
  • Create a temporary artifact directory, write a focused prompt, run a Codex review command, then read the report.
  • Claude must verify important claims against the code before passing findings back to the user.
  • Report confirmed findings separately from unverified Codex suggestions; if Codex finds nothing, say what it inspected.

codex-implementation

  • Use for bounded implementation work when a separate patch-producing worker is helpful.
  • Do not let Codex commit, push, deploy, or edit global config unless the user explicitly asked for that.
  • Pin current git state first, define scope and files to avoid, create a report directory, run codex exec with write access, inspect git status/diff, run cheap verification, then report risks.

codex-computer-use

  • Use for real UI/browser/simulator/screenshot/app-launch/runtime verification outside Claude's direct context.
  • Do not use for ordinary code reading, typechecking, linting, or tests that Claude can run directly.
  • Launching browsers, apps, and simulators for verification is allowed; ask first for disruptive actions such as closing apps, changing system settings, or acting on real accounts/data.
  • The visible edit in the video broadens the trigger from 'ask Claude to have Codex test' to simply 'ask Claude to test' a flow.

Transcript Timeline

TimestampWhat the segment establishes
11:43-12:45He introduces his global CLAUDE.md as the main place where he encoded default app stack, command behavior, and Codex handoff rules.
12:47-15:30He explains workflows vs subagents and why model routing needs explicit local judgment.
18:55-20:27He walks through model-routing mechanics, wrapper agents, labels, timeouts, and the visible skills list.
20:42-22:18He opens codex-review and explains the review workflow, command shapes, and reporting failure modes.
22:45-24:18He shows codex-implementation and codex-computer-use, then edits the computer-use trigger to make it more automatic for UI verification.
38:12-41:59Later workflow examples show worktrees, combining parallel work, using elapsed time/change volume as complexity signal, and keeping human judgment in the loop.

Hardware To Buy

Primary Mac runner

Mac mini M4 Pro, 48GB unified memory, 1TB or 2TB SSD, 10Gb Ethernet

Best fit if you want a quiet always-on macOS box for Codex desktop, browser/computer-use, Xcode, iOS/simulator checks, and Cloudflare/web preview work. Apple's current specs show M4 Pro configurations with 48GB memory and 10Gb Ethernet options.

Cheap parallel Linux runner

GMKtec K8 Plus Ryzen 7 8845HS, upgrade to at least 64GB RAM and 2-4TB NVMe

Good value if the goal is extra worktrees, browser automation, OCR/transcription, and long Codex jobs. The listed K8 Plus has dual 2.5GbE, dual USB4, Oculink, dual M.2, and up to 128GB RAM expansion.

Quiet mini-PC alternative

Beelink SER8 Ryzen 7 8845HS with 32GB/1TB base, upgrade RAM/storage

Comparable 8845HS box with vapor-chamber-style cooling, dual M.2 slots, and a higher stated memory ceiling. Choose this if acoustics and thermals matter more than Oculink.

Local-model workstation

Framework Desktop Ryzen AI Max+ 395, 128GB

Only buy this if you want serious local LLM/vision experiments or many concurrent agents. Framework lists 16 cores/32 threads, Radeon 8060S, 50 TOPS NPU, 128GB LPDDR5x, dual NVMe, Wi-Fi 7, and 5GbE.

Storage and network

4TB external NVMe scratch drive plus 2.5/10GbE switch if moving video/OCR/log corpora

This workflow creates lots of clips, frames, OCR, screenshots, and worktree artifacts. Fast scratch storage reduces drag more than another low-RAM box.

Optional Tech Stack Configuration

  • Keep the Mac as orchestrator: Codex desktop/app, Claude/Fable session, Messages/Browser, Rodney/Showboat validation, Xcode/iOS when needed.
  • Put heavy parallel work on one Linux runner first: Ubuntu 24.04 LTS, Tailscale, GitHub CLI, Codex CLI, Node/Bun/pnpm, ffmpeg, yt-dlp, tesseract, Playwright browsers, wrangler.
  • Use Cloudflare Pages for HTML reports and Cloudflare Tunnel only when a live local preview needs to be shared.
  • Define three local skills rather than copying Theo verbatim: review, implementation, and computer-use. Each should match your safety posture and local commands.
  • Force artifacts: every delegated run gets an artifact directory with prompt.md, report.md, screenshots, and command log. That makes multi-agent work auditable.
  • Worktrees are mandatory for parallel implementation agents; read-only Codex is fine for investigation and OCR/report extraction.
  • For browser validation, keep Rodney/Showboat as the default smoke layer and reserve Playwright for deeper cross-browser or tracing needs.

Extraction Limits

  • The video does not show any full SKILL.md file continuously from first byte to EOF.
  • The public screenshot covers the strongest CLAUDE.md routing section but cuts off before the last wrapper-agent bullets.
  • The sidebar exposes html-planning and use-railway by name only; contents are not recoverable from this video segment.
  • The report therefore gives a high-confidence reconstruction of visible behavior and structure, not a verbatim full-file dump. That is also the safer treatment for third-party screenshot text.

Sources

Local outputs: forensic_extraction.json, Markdown report, screenshot assets, and this HTML page live in the run folder on the Desktop.