Skip to content

03-model-and-effort.md

Latest commit

 

History

History
109 lines (73 loc) · 8.03 KB

03-model-and-effort.md

File metadata and controls

109 lines (73 loc) · 8.03 KB
Raw

Pick the model and effort

Intermediate · Lesson 3 of 8 · about 15 minutes · Needs: Turn a repeated workflow into a skill · Verified against Claude Code v2.1.292 (stable) on 2026-10-15

Goal

By the end of this lesson you can pick a model and an effort level for a task, set them for a session or for one skill, and tell from a wrong result whether to raise the effort or pick a larger model.

You need

  • The repository where you did lesson 1, with its skills, and your copy of the practice template with Node.js LTS: the check runs from the copy.
  • Usage: light. A larger model costs more per token, and a higher effort usually produces more tokens, so either tends to use more of your plan's usage.

The idea

Two separate choices shape an answer (Choosing a Claude model and effort level in Claude Code):

  • The model decides how capable the answer can be.
  • The effort decides how much work Claude does on each request: how many files it reads, how much it verifies, and how far it goes before it checks back with you (Model configuration).

Anthropic's guidance is to start from the model's default effort for most tasks. When a result is wrong, find out why before you turn a knob:

What happened, or what the task is What to change
Claude lacked context: a vague prompt, a missing tool or skill The context: the prompt, CLAUDE.md, a skill
Claude didn't try hard enough: it skipped a file or didn't run the tests A higher effort
Claude had the context, tried, and was still wrong: a subtle bug, an unfamiliar domain, a design decision A larger model, such as opus
Routine work: edits you can describe precisely, mechanical changes A smaller model, such as sonnet, and a lower effort (Choose an effort level)
Deep planning, then routine work opusplan: Opus in plan mode, then Sonnet to carry out the plan

Where a choice applies decides how you set it. The Models and effort page lists the aliases and levels:

Scope Model Effort
This session /model, then s on a row /effort, then s in the slider
Your default /model <alias> /effort <level>
One turn – ultrathink in the prompt asks for deeper reasoning; the level stays
One skill or subagent model: in its frontmatter effort: in its frontmatter

Worked example

  1. Start claude in your repository. Run /status: it shows the model. Run /effort status: it prints the effort level in force (Commands).

  2. Switch the model for this session only: run /model, move to Sonnet and press s. /status now shows Sonnet, and your default for new sessions is unchanged (Setting your model).

  3. Ask for deeper reasoning on one turn, without changing the effort level. In the practice template:

    ultrathink: which links could linkcheck report as broken although their file exists?

    Claude Code recognizes ultrathink and asks for deeper reasoning on that turn only. Phrases such as "think hard" are ordinary text (Use ultrathink).

  4. Give one task its own effort. five-whys, from lesson 1, traces a cause from evidence, where verification matters. Add a line to its frontmatter in .claude/skills/five-whys/SKILL.md:

    effort: high

    From now on the skill runs at high while it's active, over the session's level. An effort set with the CLAUDE_CODE_EFFORT_LEVEL variable, and any effort cap, still come first (Set the effort level).

Your turn

Give a routine task the smallest model and effort that do the job.

  1. Save this skill as .claude/skills/changelog/SKILL.md:

    ---
name: changelog
description: Adds a one-line entry for the staged changes to CHANGELOG.md. Use when the user asks for a changelog entry.
model: TODO
effort: TODO
disable-model-invocation: true
---

Read the staged changes with `git diff --staged`. Add one list item (`- `) under `## Unreleased` in CHANGELOG.md that says what changed for someone using the tool, creating the file and the heading if they don't exist. Change no other file.

  2. Replace both TODOs, using the first table. In auto mode, use sonnet: auto mode doesn't support Haiku, so a skill set to haiku keeps the session's model there (Eliminate permission prompts with auto mode).

  3. Stage a small change of your own, run /changelog, and approve the edit. If Claude Code says the command is unknown, run /reload-skills first.

  4. Commit the change, CHANGELOG.md and the skill together.

Check

You are done when all of these are true:

  • The committed skill picks a smaller model and a low or medium effort: git show HEAD:.claude/skills/changelog/SKILL.md | grep -E '^(model|effort):' shows both.
  • A committed CHANGELOG.md has a - entry under ## Unreleased: git show HEAD:CHANGELOG.md shows it.
  • Self-check (not tested): for a precise rename, a new feature with a clear scope, and a bug Claude keeps getting wrong with all the context, you can say which model and effort you'd pick, and why.

Run npm run check -- i-3 --dir <your repo> from your practice copy, or npm run check -- i-3 in the copy itself.

If the skill check says the model is larger than the task needs, set model: sonnet: a change you can describe precisely is routine work. If there's no entry, stage a change before you run /changelog, because the skill reads only staged changes.

Watch out

  • /model <alias> and /effort <level> save your choice as the default for new sessions. To change only the current session, press s in the picker or the slider.
  • Switching the model mid-session, and on most models changing the effort, makes the next request read the whole conversation with no cache hits; while the cache is warm, Claude Code asks you to confirm (How Claude Code uses prompt caching).
  • Set with /effort, max lasts for the current session only, and it can overthink: test it before you rely on it (Choose an effort level).

Go further

Sources: Model configuration · Commands · Choosing a Claude model and effort level in Claude Code · Extend Claude with skills · Choose a permission mode · How Claude Code uses prompt caching

← Organize project memory and see what loaded · Intermediate index · Permissions, settings scopes and the sandbox → · Topic: Models, effort and cost · Stuck on this lesson?