Effort tuner

Skill claude-hangar/claude-hangar/core/skills/effort-tuner

Production-grade configuration management for Claude Code. Hooks, agents, skills, multi-project orchestration.

Install

npx -y skills add claude-hangar/claude-hangar --skill effort-tuner

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Analyzes the current task and recommends the optimal Claude Code effort level (low | medium | high | xhigh | max) based on complexity heuristics. Helps avoid burning xhigh/max on trivial tasks or shipping undercooked analysis at low/medium. Use when: "effort tuner", "what effort", "effort level", "which effort should I use", "effort recommendation".

SKILL.md

2.5 KB, as published. Nobody here has run it

/effort-tuner

Recommends an effort level for the current task. Runs cheap — uses heuristics, not expensive reasoning — so it can sit in front of expensive agent calls without tax.

Effort ladder (Claude Code v2.1.111+)

Level	Token budget	Use case
`low`	small	Known-good mechanical edits, formatting
`medium`	moderate	Focused bug fix, localized refactor, single-file change
`high`	generous	Multi-file features, code review, architecture sketches
`xhigh`	larger still	Deep analysis, cross-cutting refactor, complex planning
`max`	full budget	Opus-scale reasoning: novel design, security audit, race-condition hunt

Decision heuristics

Score the task on these dimensions and sum:

Breadth — files touched: 1 (+0), 2–5 (+1), 6+ (+2)
Novelty — pattern exists in repo (+0) vs. new pattern (+1) vs. novel domain (+2)
Risk — style/doc (+0), logic (+1), security/data (+2)
Reversibility — local change (+0), PR-visible (+1), deployed/shared (+2)
Dependency depth — self-contained (+0), one module (+1), cross-module (+2)

Mapping

Sum	Recommendation
0–2	`low` or `medium`
3–5	`high`
6–8	`xhigh`
9–10	`max`

Output template

Effort recommendation: <level>
Task signals:
  - Breadth: <N files> (+<score>)
  - Novelty: <note> (+<score>)
  - Risk: <note> (+<score>)
  - Reversibility: <note> (+<score>)
  - Dependency depth: <note> (+<score>)
Total: <sum>/10
Rationale: <1 sentence>
If unsure: prefer one step down; you can always re-run with higher effort.

Anti-patterns flagged

max on trivial — recommend downgrade + explain token cost
low on security — reject and force ≥high
xhigh on single-file edit without dependencies — flag as possibly wasteful

effortLevel in settings.json — the session default
/effort slider (v2.1.111) — interactive switch for the current turn
Per-agent effort: frontmatter — agent-level default