skill

claude-model-guide

Task-aware Claude model + effort recommendations (Haiku, Sonnet, Opus, Fable) for Claude Code and the Claude apps, with global setup and uninstall.

vamshicreates/claude-model-guidecommunitylistedunreviewed

Use skill

Set up

Skills are installed by placing their folder where your agent loads skills.

Download ZIP from GitHub Open the repository

The download comes from github.com/vamshicreates/claude-model-guide itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 22, 2026

Claude Model Guide 🎛️

Get a one-line model recommendation before Claude starts a new task — Haiku, Sonnet, Opus, or Fable + effort level — based on the work. Switch with /model and /effort and reply go, or reply continue to stay on your current setup.

One-time global setup. You control the switch.

The Claude port of Codex Model Guide. It combines a reusable skill with a one-time installer that adds a marked block to your global Claude Code memory file, ~/.claude/CLAUDE.md, which Claude Code loads into every session.

Install for Claude Code

git clone https://github.com/vamshicreates/claude-model-guide.git
cd claude-model-guide
python3 skills/model-guide/scripts/setup.py install --dry-run
python3 skills/model-guide/scripts/setup.py install

Requires Python 3 and Git. Start a new Claude Code session (or run /clear) after setup. On Windows, use py -3 if python3 isn't available. The installer honors CLAUDE_CONFIG_DIR and also supports --claude-home PATH.

Optional: also install the skill so you can ask for a recommendation on demand with /model-guide:

mkdir -p ~/.claude/skills && cp -r skills/model-guide ~/.claude/skills/

Or just give Claude Code the link:

Read https://github.com/vamshicreates/claude-model-guide, run its setup command to enable global model recommendations, and preserve my existing CLAUDE.md. After setup, tell me how to uninstall it.

Install for the Claude apps (claude.ai, desktop, mobile)

The apps have no CLAUDE.md, so paste the contents of claude-app-preferences.md into your profile preferences in Settings (or a Project's instructions to scope it to one project). Switch models in the model selector.

Behavior

Task	Starting recommendation	Claude Code switch
Focused questions, extraction, small edits	Haiku · low/medium	`/model haiku` · `/effort low`
Multi-file coding, debugging, research synthesis	Sonnet · medium/high	`/model sonnet` · `/effort medium`
Hard reasoning, architecture, exacting deliverables	Opus · high/xhigh	`/model opus` · `/effort high`
Hardest long-running work (if available on your plan)	Fable · xhigh/max	`/model fable` · `/effort xhigh`

Example:

Model recommendation: Sonnet · medium — multi-file bug fix with tests to run.
Switch with /model sonnet and /effort medium, then reply go — or reply continue.

Exact versions depend on the models your plan exposes. These choices are heuristics, not benchmarks. Claude can't always see your current model or effort; when it can't, it gives the recommendation and waits for go or continue. Tasks that clearly fit in one turn skip the recommendation entirely (the sizing guess is shakiest there and the routing tokens are a bigger share of the job). Follow-ups, acknowledgments, and continuations of the same task don't repeat the gate. Subagents, headless (claude -p) runs, and scheduled tasks never stall on it.

Model switching stays manual. Routing costs a few tokens, and savings depend on your workload, model, effort, and plan's usage accounting — no fixed savings are promised. This is an instruction convention, not an enforced runtime interceptor; project CLAUDE.md files and explicit instructions can override it.

Extension: Subagent Routing (Claude Code)

Model Guide picks the model for the main conversation. The subagent-routing skill picks models per step, ahead of time, so nothing stops to ask. The main session stays on one model and hands work to 14 fixed-model subagents.

The rule behind the assignments: match the model to the cost of a quiet failure. Loud, easy-to-spot mistakes can go to a cheap model. Quiet ones that still read fine go to Opus.

Seat	Model · effort	Roles
Cheap workers	Haiku · low	`scout` (find), `extractor` (extract, format), `triager` (label, sort, dedupe), `summarizer` (condense for orientation)
Builders and analysts	Sonnet · medium	`analyst`, `researcher`, `implementer`, `test-writer`, `drafter`; `debugger` runs at high
Planners and checkers	Opus · high	`planner`, `reviewer` (code), `judge` (claims vs sources), `synthesizer` (final answer)

Planners and checkers never get the cheap model. A weak checker fails quietly, and the output still reads fine. Save money on the workers, not on the checking.

Typical flows:

Code: scout → planner → implementer + test-writer → reviewer
Research: researcher + analyst in parallel → judge → synthesizer
Writing: drafter variants → judge (facts) → synthesizer
Bulk data: extractor + triager → analyst → judge

python3 skills/subagent-routing/scripts/setup.py install --dry-run
python3 skills/subagent-routing/scripts/setup.py install

This copies the agents into ~/.claude/agents/ (honors CLAUDE_CONFIG_DIR and --claude-home PATH). It never overwrites an agent you already have under the same name, and uninstall removes only the files it installed. To change a role's model, edit the model: and effort: lines in its file. To have Claude follow the routing rules in every session, also copy the skill: cp -r skills/subagent-routing ~/.claude/skills/. Start a new session afterward. Subagent roles need Claude Code; the Claude apps don't support custom subagents.

Idea credit: a Reddit reply that suggested choosing models per subagent and skipping the recommendation for one-turn tasks.

Disable or inspect

python3 skills/model-guide/scripts/setup.py status
python3 skills/model-guide/scripts/setup.py uninstall

Uninstall removes only the marked Model Guide block and restores the file byte-for-byte. Changes create timestamped backups next to CLAUDE.md. Deleting the skill or this repository alone does not disable the global rule. Start a new session after uninstalling. In the Claude apps, delete the pasted text from your preferences.

Contribute

Test one simple task and one complex task, including manual switching and uninstalling. Run python3 -m unittest discover -s tests -v to check the installer.

References

Community project; not affiliated with Anthropic. Made by Vamshi Creates.