model-policy

Model policy

A skill only loads when called, so the policy itself must not live in a skill. This skill copies it into your user-level Claude Code config, where it applies to every session in every project:

File	What it does
~/.claude/rules/model-policy.md	The Pareto ladder, which agent gets which model and effort, the 100k rule. Loaded every session.
~/.claude/agents/Explore.md	Replaces the built-in Explore, which otherwise runs on the session model (Opus).
~/.claude/agents/implementer.md, verifier.md	Haiku high workers for edits and for read-only checks.
~/.claude/settings.json env	CLAUDE_CODE_SUBAGENT_MODEL=haiku: subagents with no model run on Haiku.

Install globally, then run /model-policy:

bunx skills add FixMyBerlin/fixmyskills --skill model-policy -g -a claude-code -y

Needs Claude Code 2.1.296 or newer: older versions ignore autoCompactWindow silently and the agents never compact at 100k.

Steps

1. Run claude --version; stop and say so when it is below 2.1.296.
2. Read assets/model-policy.md and every file in assets/agents/.
3. For each target file that exists and differs from its asset, show the diff and AskUserQuestion: overwrite, or keep theirs. The skill cannot tell a hand edit from an older version of the asset, so it asks on every difference.
4. Copy assets/model-policy.md to ~/.claude/rules/model-policy.md and each assets/agents/*.md to ~/.claude/agents/.
5. In ~/.claude/settings.json, set env.CLAUDE_CODE_SUBAGENT_MODEL to "haiku". Keep every other key; create the file with { "env": { … } } if it is missing. Do not set CLAUDE_CODE_SUBAGENT_MODEL_FORCE — it would also override the per-call model that escalation needs.
6. Report what changed. The rule and agents apply from the next session on.

Updating the ladder

The ladder in assets/model-policy.md comes from the slopalytics chart linked in its "Last checked" line. To refresh it in this repo: open that link (add new Claude models to models=), list the Claude model/effort points on the frontier from cheapest to smartest, update the ladder and the "off the line" list, set "Last checked" to today, and commit. Then everyone re-runs bunx skills update -g -y and /model-policy.

Why each piece

• Rule, not CLAUDE.md: a rule file without paths loads at the start of every session like CLAUDE.md, but stays one file this skill can replace on update.
• Env var: model resolution is the Agent call's model, then the agent file's model, then CLAUDE_CODE_SUBAGENT_MODEL, then the session model. The env var catches every spawn that names no model.
• Explore agent: built-in Explore and Plan ignore the env var. A user-level agent with the same name replaces Explore; Plan stays on the session model on purpose.
• Effort: there is no setting for a default subagent effort. Agent files pin it (effort: frontmatter); for general-purpose the rule tells the orchestrator to pass effort on each call.
• Haiku high, not medium: Haiku medium is also on the Pareto line, but high is a lot smarter for little extra cost, so the agents pin high and the ladder starts there.
• 100K window: Haiku 5.5 has two prices, chosen per request by prompt length: $0.10 / $0.50 per million input / output tokens up to 100K, $0.50 / $2.50 above — for the whole request, with cache reads and writes counted toward the 100K. A subagent resends its context every turn, so once it passes 100K every later turn pays more. autoCompactWindow in an agent's frontmatter only affects that agent, so Haiku compacts near its price step while the Opus session keeps its full window; 100000 is the lowest value Claude Code accepts. Even over 100K, Haiku stays cheaper than switching: on cache reads, which dominate agent turns, it is $0.05 against Sonnet's $0.10 and Opus's $0.20 per million; on fresh input and output 4× and 8× cheaper. On Pro / Max plans usage is not billed per token; the docs do not say whether the price step affects plan limits.

Docs: Subagents, Auto-compact window, Memory and rules, Settings, Haiku long-context pricing.