claude-model-selection

Claude-Model-Selection

A Claude Code plugin that helps you spend fewer tokens. At the start of every chat it looks at what you asked for and recommends the cheapest Claude model (Haiku, Sonnet, Opus or Fable) and effort level that will still do the job well. It also shows how much of your plan you have used.

🧭 Claude-Model-Selection
Task: design architecture for a live-money trading bot
Recommended: Opus · effort high  (current: Sonnet)
Reason: system design with real-money risk
Usage: 5h 12% (resets in 3h 40m) · week 49% (resets in 3d 4h) · context 58k tokens

It replies in your language.

Install

In Claude Code, paste these two commands one at a time (Claude Code runs one slash command per message).

1. Add the marketplace:

   /plugin marketplace add Svelez97/claude-model-selection
   

2. Install the plugin:

   /plugin install claude-model-selection@claude-model-selection
   

Or paste this single message into Claude and it installs everything for you:

Install the plugin claude-model-selection from GitHub: Svelez97/claude-model-selection

Or, from a terminal, in one line:

claude plugin marketplace add Svelez97/claude-model-selection && claude plugin install claude-model-selection@claude-model-selection

Start a new chat afterwards so it loads.

What it does

When the first real message of a chat arrives (a bare "hi" is ignored), it:

1. Reads your usage: the 5-hour limit, weekly limit and context size.

2. Classifies the task from your message alone. It does not read files, so the check stays cheap.

   Task	Model	Effort
   Quick questions, formatting, simple commands	Haiku	n/a (Haiku has no effort setting)
   Small edits, scripts, clear bugs, explanations	Sonnet	low / medium
   Multi-file features, refactors, unclear bugs, data analysis	Sonnet	medium / high
   Architecture, hard bugs, security, real-money or high-impact decisions	Opus	high / xhigh
   Very large, long-running projects Claude runs mostly on its own, where mistakes are very costly	Fable	high / xhigh

   It always picks one effort level. Two more levels are kept for rare cases, and it explains why when it suggests them:

   • max: deepest reasoning, for extremely hard problems that already failed at xhigh.
   • ultracode: a Claude Code setting that plans a whole multi-step workflow per task; for large builds only.

   Fable uses the most of your limit and needs a paid plan, so it is recommended only for the last row. If you do not have it, you get Opus instead.

3. Adjusts to your plan: if you are close to your 5-hour or weekly limit, it recommends one tier lower (Fable → Opus → Sonnet → Haiku), avoids max and ultracode, and tells you when the limit resets.

4. Avoids pointless switches: switching models mid-chat throws away the prompt cache, and the new model re-reads the whole conversation. On borderline tasks it keeps your current model. In a long chat it offers a summary so you can open a new chat on the right model.

5. Suggests savings only when they apply:

   • compact the chat or start a new one when the context gets large;
   • turn off connectors the task does not need (only with your OK);
   • hand large mechanical sub-tasks to a cheaper Haiku or Sonnet subagent (only with your OK).

6. Waits for you when your setting does not match. It compares your current model and effort with its recommendation, and stops before starting the task if they differ:

   • Lower than recommended: the result may fall short. Switch up, or reply "continue" to approve working with the lower setting.
   • Higher than recommended: you could save usage. Switch down, or reply "continue" to keep your setting.

   After switching, send any message such as "done" and it carries out your original request. If model and effort already match, it goes straight to the task. The current effort can only be read in the Claude desktop app; elsewhere only the model is compared.

If the task changes a lot mid-chat, it gives a one-line recommendation instead of the full block.

Subagent cost guard

When Claude hands part of the work to a subagent:

• Same model as your session, or a cheaper one: it runs without asking.
• A more expensive model (Haiku < Sonnet < Opus < Fable): you are asked to approve it first.

A PreToolUse hook enforces this, so it holds even if Claude forgets the rule. It also catches custom agents whose definition sets a more expensive model. If the session model cannot be determined, only Haiku subagents run without asking.

To skip it for a chat, say "skip selection" in your first message.

Notes and limits

• Claude cannot switch its own model. You change it in the model picker, or with /model in the terminal. The plugin only recommends.
• Live usage numbers need the Claude desktop app. It reads them with a tool that only the desktop app (Code tab) provides. Elsewhere (terminal, IDE, web) it asks you to run /usage and /context and still recommends a model and effort.
• The subagent guard and the per-message reminder need bash. macOS and Linux have it, and on Windows it comes with Git for Windows (Git Bash), which Claude Code normally uses. Without bash they do not run; the start-of-chat reminder still works. It also cannot ask you in bypass-permissions mode or in non-interactive runs (claude -p); there, an approval request blocks the subagent instead.
• The thresholds (80 % of the 5-hour limit, 85 % of the week, 150k context tokens) are rules of thumb. To change them, edit skills/claude-model-selection/SKILL.md.

How it works

• skills/claude-model-selection/SKILL.md: the skill with the selection rules.
• hooks/hooks.json:
  • a SessionStart hook that tells Claude to run the skill once at the start of each new chat;
  • a UserPromptSubmit hook (hooks/remind-until-run.sh) that repeats that reminder with each of your messages until the skill has run in the chat, so it triggers reliably;
  • a PreToolUse hook on subagent launches that runs hooks/check-subagent-model.sh, the cost guard.

Uninstall

/plugin uninstall claude-model-selection@claude-model-selection

License

MIT

En español

Plugin para Claude Code que, al iniciar cada chat, recomienda el modelo más barato (Haiku, Sonnet, Opus o Fable) y el nivel de esfuerzo adecuados para tu tarea. También muestra tu uso del límite de 5 horas, el semanal y el tamaño del contexto. Si tu modelo o esfuerzo no coincide con lo recomendado (sea menor o mayor), se detiene antes de empezar: cámbialo y escribe cualquier mensaje (por ejemplo "listo"), o responde "continuar" para aprobar seguir con tu configuración actual. Si Claude quiere lanzar un subagente con un modelo más caro que el de tu sesión, primero te pide autorización; con el mismo modelo o uno más barato lo hace sin preguntar. Para instalarlo, pega los dos comandos de arriba uno por uno (Claude Code ejecuta un solo comando / por mensaje), o pega en Claude el mensaje de una sola línea de arriba y él lo instala por ti. Responde en tu idioma.