Skip to content
Back to search

arbitrage

A Claude Code skill for token arbitrage: keep premium model tokens for judgment, dispatch all code-writing to codex.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/blader--arbitrage/badge.svg)](/listings/blader--arbitrage)

HTML

<a href="/listings/blader--arbitrage"><img src="/api/listings/blader--arbitrage/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

**Claude Code**

git clone https://github.com/blader/arbitrage.git ~/.claude/skills/arbitrage

**Codex**

git clone https://github.com/blader/arbitrage.git ~/.codex/skills/arbitrage

Requires the codex CLI (`codex exec`) on your PATH.

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 60 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: arbitrage
  3. description: Always active when coding in a Claude (Fable) session. Triggers whenever implementation work is being planned, scoped, or about to start — before writing any code — to decide where each piece of work executes.
  4. ---
  5. # Arbitrage
  6. Fable tokens are expensive; the user's codex quota sits unused. Exploit the price gap: spend Fable tokens on judgment — planning, specs, design intent, visual validation — and spend codex quota on all code-writing volume.
  7. **You ARE the Fable session.** "Here" means you do it yourself. There is no separate Fable to delegate to.
  8. ## Routing Table
  9. | Work | Where | Why |
  10. |------|-------|-----|
  11. | Planning, specs, architecture decisions | Here | Judgment is what Fable is paid for |
  12. | **ALL implementation — backend, frontend, anything that writes code** | codex via `/goal` | Quota is free; typing code is not where Fable adds value |
  13. | Visual validation of UI work (run the app, screenshot, judge, iterate) | Here | Fable's design eye is worth the price — validate, don't type |
  14. | Investigation and debugging analysis (root-causing hard bugs) | Here | Judgment work; the resulting fix dispatches with a precise spec |
  15. | Diff review, commits, pushes, PRs | Here | Git operations stay under your control |
  16. | Trivial edits riding along with review/validation (one-liners, any layer) | Here | Dispatch overhead exceeds the work |
  17. | Work codex has demonstrably struggled with | Here | Escape hatch — see below |
  18. ## Dispatch Protocol
  19. 1. **Spec first.** Write a spec file in the worktree: objective, constraints, files/areas involved, acceptance criteria — the exact test command that must go green — and what NOT to touch. For frontend work the spec carries the design intent: layout, states, interactions, spacing, motion, reference patterns. For the hardest parts, spec down to pseudocode — the thinking stays here; the typing still dispatches.
  20. 2. **Dispatch in the background and keep working:**
  21. ```bash
  22. codex exec --cd <worktree> "/goal <one-line objective; details in SPEC.md>"
  23. ```
  24. codex runs gpt-5.5 with xhigh reasoning (its default config); it implements and gets tests green in the isolated worktree while you continue planning/spec/validation work in parallel.
  25. 3. **Review the diff here.** Commit, push, and open PRs from this session.
  26. ## Visual Validation Loop (frontend)
  27. 1. codex implements the UI per spec.
  28. 2. Run the app here — interact, screenshot, judge against the design intent.
  29. 3. Write concrete visual feedback (what's off, by how much, what good looks like) and re-dispatch.
  30. 4. Loop until it looks right; then diff review + git here.
  31. ## Escape Hatch: When codex Struggles
  32. Struggle must be **observed, not predicted**. Always dispatch first — predicting "codex can't do this" just means the spec needs pseudocode-level detail. But two strikes and it's yours: if a re-dispatch with concrete corrective feedback still comes back missing the acceptance criteria or mangling the approach, stop re-dispatching and write the code here yourself, salvaging whatever of codex's diff is sound.
  33. ## Red Flags
  34. | Thought | Reality |
  35. |---------|---------|
  36. | "Frontend is design-critical, so I'll code it here" | Fable's taste shows up in the spec and the validation loop, not in typing JSX. Dispatch it. |
  37. | "Faster to just write it myself" | Premium tokens on rote work. Spec it, dispatch it, validate in parallel. |
  38. | "This part is too hard for codex" | Predicted struggle is not observed struggle. Spec it at pseudocode level and dispatch; you only take over after two failed rounds. |
  39. | "I'll let codex commit and push" | Review the diff and run git from this session. |
  40. | "No time to write a spec" | An unspecced dispatch comes back wrong and costs more Fable tokens to fix than the spec would have. |
  41. ## Common Mistakes
  42. - Dispatching without acceptance criteria → codex returns "done" with red tests. Always name the test command that must pass.
  43. - Dispatching frontend work without design intent in the spec → generic UI comes back; the validation loop burns rounds recovering what the spec should have said.
  44. - Blocking on codex → run it in the background; keep doing planning/spec/validation work in parallel.
  45. - Splitting one coherent unit across dispatches → pick one owner per unit; split at the API boundary, not mid-feature.
README.md · README · 38 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # Arbitrage
  2. A skill for Claude Code that exploits the price gap between your premium model and your idle codex quota.
  3. Frontier-model tokens are expensive, and most of them get spent typing code — the one part of the job a cheaper agent already does fine. Meanwhile your codex subscription sits there barely touched. This skill makes the expensive model act like it knows that: it keeps the judgment work (planning, specs, architecture, visual validation, diff review) and dispatches every line of implementation to codex running in the background.
  4. Token arbitrage for agentic coding.
  5. ## Install
  6. **Claude Code**
  7. ```bash
  8. git clone https://github.com/blader/arbitrage.git ~/.claude/skills/arbitrage
  9. ```
  10. **Codex**
  11. ```bash
  12. git clone https://github.com/blader/arbitrage.git ~/.codex/skills/arbitrage
  13. ```
  14. Requires the codex CLI (`codex exec`) on your PATH.
  15. ## How it works
  16. 1. Before any code gets written, the agent writes a spec: objective, constraints, acceptance criteria (the exact test command that must pass), and what not to touch. For frontend work the spec carries the design intent.
  17. 2. It dispatches implementation to codex in the background and keeps planning or validating in parallel instead of sitting and waiting.
  18. 3. The diff comes back. The agent reviews it, commits, and opens PRs from the main session. codex never touches git.
  19. 4. Frontend gets a visual validation loop: codex types the UI, the premium model runs the app, screenshots it, judges it against the design intent, and re-dispatches with concrete feedback until it looks right.
  20. ## The escape hatch
  21. The skill forbids predicting that codex will fail ("this is too hard for it"). Struggle has to be observed. But two strikes and it's over: if a re-dispatch with corrective feedback still misses the acceptance criteria, the premium model writes the code itself and salvages what it can from codex's diff.
  22. ## License
  23. MIT

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.