AGENT SKILL

model-router

Recommend a model tier (Haiku / Sonnet / Opus) for a coding task by scoring it on scope, novelty, risk, and expected iteration. Use when the user describes a task they are about to start, or explicitly asks "which model should I use for this".

License 许可
MIT
最近更新
2026-07-31
Stars
★ 2
Forks
0

01 前言 Front Matter

name
model-router
description
Recommend a model tier (Haiku / Sonnet / Opus) for a coding task by scoring it on scope, novelty, risk, and expected iteration. Use when the user describes a task they are about to start, or explicitly asks "which model should I use for this".

02 资源文件 1

03 SKILL.md 正文

Model router

Recommend which model tier a task warrants, so routine work doesn't burn Opus-tier tokens and hard work doesn't get under-powered.

When to use this

  • The user describes a task they are about to start (a feature, fix, refactor, script, migration) and hasn't yet picked a model.
  • The user explicitly asks "which model should I use for this" or similar.

Never refuse to answer. If the task description is ambiguous, score conservatively (round the uncertain axis up, not down) and say which axis is uncertain.

Scoring rubric

Score the task on four axes, each 1–3:

Axis 1 2 3
Scope Single-file edit Multi-file feature Cross-cutting refactor or greenfield system
Novelty Pattern exists in repo Familiar domain, new pattern Unfamiliar domain or algorithmically hard
Risk Throwaway or easily reverted User-facing but tested Data-touching, security-sensitive, or hard to roll back
Iteration Likely one-shot A few review cycles Long-horizon agentic session with many tool calls

Routing rule

Total score Recommendation
4–6 Haiku-tier (mechanical edits, renames, boilerplate)
7–9 Sonnet 5 (the default; medium-effort agentic work)
10–12 Opus-tier (high ambiguity, high risk, long horizon)

Override: if Risk scores 3, bump the recommendation up one tier (Haiku → Sonnet, Sonnet → Opus; Opus stays Opus).

Output format

Keep the whole answer under 15 lines:

  1. The recommendation, first line, bolded.
  2. The four scores, one line each, with one-line reasoning per axis.
  3. The estimated relative cost multiplier vs the recommended tier if the user insists on the tier above (per-Mtok list price ratio: Haiku→Sonnet ≈ 3x, Sonnet→Opus ≈ 1.7x on input / 1.7x on output; Opus→Fable ≈ 2x). If Opus is already recommended, note Fable 5 as the tier above.

Example

Task: "Rename a config key across the repo and update the docs."

Recommendation: Haiku-tier (total 5, no override)

  • Scope 2 — touches many files, but mechanically.
  • Novelty 1 — pure find-and-replace, pattern exists.
  • Risk 1 — trivially reverted in one commit.
  • Iteration 1 — one-shot with a grep to verify. Insisting on Sonnet 5 instead ≈ 3x the per-token cost for no quality gain.

04 同仓库更多 Skills

Temrel
Temrel/temrel-agentic-toolkit

mcp-spec-audit

Audit an MCP server codebase for the MCP 2026-07-28 stateless spec revision by running the mcp-spec-audit CLI and summarizing its findings. Use when the user asks to audit, check, or migrate an MCP server for the new spec, asks whether their server is ready for MCP 2026-07-28 or "the stateless revision", or mentions migrating off Roots, Sampling, protocol Logging, or the experimental Tasks API.

★ 2
Temrel
Temrel/temrel-agentic-toolkit

example-skill

A copy-me template skill. Replace this description with a precise, one-line summary of what the skill does and when an agent should reach for it, because this line is what the agent reads to decide relevance.

★ 2
Temrel
Temrel/temrel-agentic-toolkit

example-command

A copy-me template for a human-invoked skill. Replace this with a one-line summary of what running /your-command does.

★ 2
Temrel
Temrel/temrel-agentic-toolkit

gated

Use only when the user explicitly asks to run the release checklist.

★ 2
Temrel
Temrel/temrel-agentic-toolkit

auto-leak

Use whenever the user mentions databases, migrations, or schema changes. Runs the full migration linter and applies fixes automatically.

★ 2
Temrel
Temrel/temrel-agentic-toolkit

autonomy-audit

Inventory every behaviour this Claude Code setup can trigger (hooks, skills, MCP servers, permission rules, CLAUDE.md directives), classify each as auto-firing or manual, tag its cost-of-wrong, and write a recommended trigger policy to autonomy-audit.md. Takes an optional path to audit a different project.

★ 2