Claude model picker
Skill gabimoncha/skills/skills/essentials/claude-model-picker
Use before spawning subagents, workflow dispatch or ultrawork, in order to choose the model and reasoning effort.From its SKILL.md
npx -y skills add gabimoncha/skills --skill claude-model-pickerAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 12 stars12 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
5.4 KB, as published. Nobody here has run it
Picking the right models for workflows and subagents
Rankings on a 1–10 scale, higher = better. Cost reflects what I actually pay (OpenAI has really generous limits), not list price. Intelligence is how hard a problem you can hand the model unsupervised. Taste covers UI/UX, code quality, API design, and copy.
| model | cost | intelligence | taste |
|---|---|---|---|
| gpt-5.6-sol | 7 | 8 | 6 |
| gpt-5.6-luna | 9 | 6 | 5 |
| gpt-5.3-codex-spark | 6 | 4 | 3 |
| sonnet-5 | 3 | 5 | 7 |
| opus-4.8 | 2 | 7 | 7 |
| fable-5 | 1 | 9 | 9 |
How to apply
- These are defaults, not limits. You have standing permission to override them: if a cheaper model's output doesn't meet the bar, rerun the work with a smarter model without asking. Judge the output, not the price tag. Escalating costs less than shipping mediocre work.
- Cost is a tie-breaker only; when axes conflict for anything that ships, intelligence > taste > cost.
- Mechanical preprocessing: gpt-5.6-luna low.
- Bounded read-only exploration: gpt-5.6-luna medium.
- Completed exploration summaries: gpt-5.3-codex-spark high when supported; otherwise gpt-5.6-luna high.
- Clear-spec implementation: gpt-5.6-luna high.
- Judgment or an independent review: gpt-5.6-sol low.
- Ambiguous multi-subsystem implementation: gpt-5.6-sol medium.
- Security and architecture: gpt-5.6-sol high.
- Only genuinely hardest cases: gpt-5.6-sol max when the dispatch surface supports it.
- Never use Haiku.
- Never use Terra.
- Anything user-facing (UI, copy, API design) needs taste ≥ 7: fable-5, opus-4.8, or sonnet-5.
- Reviews of plans or implementations can use fable-5 for a native Claude perspective or gpt-5.6-sol low for an independent Codex perspective.
- Device QA with Argent or agent-device: gpt-5.6-luna low, via
/codex:rescue --model gpt-5.6-luna --effort low.
Mechanics
Claude models (sonnet-5, opus-4.8, fable-5) run via the model parameter on Agent and Workflow calls.
From the main session, prefer the openai/codex-plugin-cc plugin's slash commands — they own session tracking, background jobs, and structured review output:
/codex:rescue— delegate active debugging, multi-file refactoring, or implementation loops. The only command with model selection:--model <gpt-model>(shorthand:spark→ gpt-5.3-codex-spark) and--effort <none|minimal|low|medium|high|xhigh>./codex:review— non-destructive, read-only code quality review. Supports--base <ref>for branch analysis and--wait|--backgroundfor execution mode. No model/effort flags — it uses Codex's default model./codex:adversarial-review— skeptical design review pressure-testing tradeoffs, auth, and reliability. Append focus text at the end to steer it. Same flags as/codex:review./codex:status//codex:result//codex:cancel— check, fetch, or abort jobs launched with--background.
When delegating through the Agent tool or a workflow, skip the plugin and run the Codex CLI directly through a thin Claude wrapper — model on Agent/agent() only accepts Claude models, so the wrapper is a cheap sonnet that shells out:
Agent(subagent_type: "general-purpose", model: "sonnet", run_in_background: true, prompt: `
You are a thin execution wrapper. Run exactly this command via Bash, then return the
final Codex message, tokens used, and exit code. Do no other work; no retries with
different flags — return errors verbatim.
codex exec -m gpt-5.6-sol -c model_reasoning_effort="high" "<self-contained task>" </dev/null
`)
In workflows, use the same wrapper prompt via agent(prompt, {label: 'gpt-5.6-sol:<task>', model: 'sonnet', effort: 'low', schema: {...}}) — put the GPT model in the label so the TUI shows what's actually thinking; the model column will show the sonnet wrapper.
- Effort values:
minimal|low|medium|high|xhigh. Add-s read-onlyfor review/diagnosis-only tasks; the default sandbox is write-capable. - The Codex prompt must be self-contained — Codex sees none of the conversation context.
- ALWAYS keep the
</dev/null: with a kept-open non-TTY stdin (e.g. a backgrounded Bash call),codex execblocks forever on "Reading additional input from stdin...". - Long runs: prefer one foreground Bash call with a raised timeout (max 10 min);
run_in_background+ polling is safe only with the stdin redirect in place. - NEVER spawn
codex:codex-rescueas a named teammate (name:param) — it is a one-shot forwarder, and a named spawn idles forever without starting work. Unnamed background spawns of it still work if the plugin path is preferred. - Do NOT pass a GPT model string to
model:on Agent/agent() — without an LLM gateway it dies immediately ("model may not exist"). The wrapper is the only direct-dispatch path, and it keeps Codex-plan pricing (a gateway would bill OpenAI API per-token rates, invalidating the cost table).
For closed-loop quality assurance, keep the review gate on via /codex:setup --enable-review-gate: a stop hook automatically challenges Claude's outputs with Codex before finalizing, so broken code or weak design assumptions never reach the main session unvetted.