~/resources/skills/3p
// skill

3P

Third-party delegation. Mechanical, high-volume code work — find/replace migrations, dead code deletion, boilerplate generation — gets routed to Codex CLI or a local LLM. Claude tokens stay reserved for the work that actually needs a brain.

// what it does

Most of what burns Claude tokens isn't interesting. It's grunt work — migrating a thousand inline styles to a design system, renaming an import path across forty files, deleting an orphaned component. Claude is wildly overqualified for this and on a weekly cap it's the kind of work that makes you run out before Friday.

3P is a router. When a task fits the profile (mechanical, high-volume, no architectural choices, no business logic, no credentials) Claude writes a precise prompt, dispatches it to a cheaper executor, and verifies the result. Default executor is Codex CLI. Next tier down is a local LLM via Ollama. Claude only steps back in if verification fails or the task needed judgment after all.

The fence around the prompt is the whole skill. Every dispatch carries: exact file paths, the transformation, a token map, what to keep as-is, what not to touch, a verification command. Without that fence, a cheap executor will cheerfully rewrite your data fetching while it's "tidying up." With it, you get the same output Claude would have produced — for free, on a budget that scales with the work, not the model.

// how to install
  1. Install Codex CLI: npm install -g @openai/codex@latest then codex login.
  2. (Optional) Install Ollama for the local-LLM tier — ollama.com. Pull a small code model (e.g. ollama pull qwen2.5-coder:7b).
  3. Download the file below — it's a single SKILL.md.
  4. Save it to ~/.claude/skills/3p/SKILL.md. Create the folder if needed.
  5. Restart Claude Code so the skill is picked up.
  6. Trigger: /3p <task> — or let Claude self-trigger when it spots mechanical work.
trigger:/3p
~ download SKILL.md →
// peek inside — the executor matrix 
| Executor    | When to use                                         | Cost            |
|-------------|-----------------------------------------------------|-----------------|
| Codex CLI   | Default. Mechanical code changes.                   | Subscription    |
| Local LLM   | Simple find-replace, linting, classification.       | Free            |
| Claude Haiku| Quick classification / summarization (no code).     | ~$0.001 / call  |

YES — delegate these:
- inline style → design-token migrations (file by file)
- hardcoded value → token-import replacements
- import path renames across many files
- dead code / orphan component deletion (confirmed orphans only)
- stale comment cleanup
- boilerplate page generation from a template
- mechanical test generation for pure functions

NO — keep in Claude:
- architecture decisions
- data model changes
- security-sensitive code (auth, tokens, RLS)
- external API integrations
- novel feature implementation
- anything touching credentials or env vars