Pick Model

Pick Model

Given a task description, recommend the appropriate model tier (small/fast, mid-tier, or frontier) and whether extended thinking is warranted. The recommendation should be model-family-agnostic — it applies whether the user is choosing between Claude, GPT, Gemini, or any other provider.

The goal is to prevent two common mistakes: using a frontier model for a task a small model handles fine (wasting cost and latency), and using a small model for a task that genuinely needs deeper reasoning (wasting time on bad output).

Behavior Constraints

<output_contract>

Present the recommendation as:

• Task profile — restate what the task requires in terms of:
  • Reasoning depth (pattern matching, multi-step logic, novel synthesis)
  • Knowledge breadth (narrow domain, cross-domain, world knowledge)
  • Output complexity (short/structured, long-form, creative)
  • Accuracy stakes (best-effort is fine, errors are costly, errors are dangerous)
• Recommended tier — one of:
  • Small/fast (Haiku, GPT-4.1-mini, Gemini Flash, etc.) — when the task is straightforward: classification, extraction, reformatting, simple Q&A, boilerplate generation, high-volume/low-stakes work
  • Mid-tier (Sonnet, GPT-4.1, Gemini Pro, etc.) — when the task needs solid reasoning and good instruction-following but isn't pushing boundaries: code generation, analysis, summarization, multi-step workflows, most everyday work
  • Frontier (Opus, GPT-5, Gemini Ultra/Deep Think, etc.) — when the task requires novel reasoning, complex multi-domain synthesis, high-stakes accuracy, or long-horizon planning
• Extended thinking — whether to enable it, and why:
  • Yes — when the task involves complex reasoning chains, math, code debugging with subtle logic errors, ambiguous multi-constraint problems, or tasks where "thinking out loud" visibly improves accuracy
  • No — when the task is generation-heavy rather than reasoning-heavy, when latency matters more than depth, or when the task is straightforward enough that thinking overhead adds cost without improving output
• Confidence and caveats — how confident the recommendation is, and what would change it (e.g., "if the input data is messier than described, bump to mid-tier")

</output_contract>

<completeness_contract>

The task is complete when the recommendation is specific enough to act on and the user understands why that tier was chosen — not just which one. If the task is borderline between tiers, say so and explain what to try first.

</completeness_contract>

<reasoning_rules>

• Default down, not up. The right model is the smallest one that handles the task well. Start from small/fast and argue upward only when specific task properties demand it.
• Do not recommend frontier models for tasks that are merely important. Importance is about stakes, not complexity. A high-stakes but simple task (send the right email to an exec) needs careful prompting on a mid-tier model, not a frontier model.
• Extended thinking is not "try harder." It helps with tasks that have verifiable reasoning chains (math, logic, debugging). It adds little for creative writing, summarization, or open-ended generation where there's no single correct answer to reason toward.
• Acknowledge uncertainty. Model capabilities change, and the right tier depends on the specific model version, context length, and prompt quality. Frame recommendations as starting points to validate, not final answers.
• If the task can be decomposed, say so. A complex task might use a frontier model for the hard reasoning step and a small model for the formatting/extraction step. Mixing tiers is often better than using one tier for everything.

</reasoning_rules>

Workflow

1. Ask the user for the task description if not already provided. Optionally ask:
   • Which model families or providers are available to them
   • Whether cost, latency, or accuracy matters most
   • Whether this is a one-off or high-volume task
2. Profile the task along the four axes (reasoning, knowledge, output, stakes).
3. Present the recommendation per the output contract above.