claude-code-boost

Make Claude Sonnet perform like a frontier model. One skill, zero extra spend on big-model tokens.

Frontier models (Fable, GPT-5.x) aren't magic. They just don't skip steps. Research on inference-time scaffolding shows smaller models close most of the gap when you force three habits:

1. Define the pass/fail check before solving. A test command, a schema diff, verifiable claims.
2. Plan in a separate pass before touching code. Planner and actor are different jobs.
3. Have fresh eyes try to break the result. An adversarial reviewer in a clean context, not the author grading itself.

And one anti-pattern to avoid at all costs: self-reflection without a verifier makes models worse. GPT-4 dropped from 95.5% to 89.0% on GSM8K when it looped on its own judgment with no oracle. So boost only ever loops on an objective gate.

What's inside

skill/SKILL.md     # Claude Code skill: drop into ~/.claude/skills/boost/
scripts/boost.sh   # headless version for cron jobs and CI

Install

mkdir -p ~/.claude/skills/boost
curl -fsSL https://raw.githubusercontent.com/agenticluke/claude-code-boost/main/skill/SKILL.md -o ~/.claude/skills/boost/SKILL.md
curl -fsSL https://raw.githubusercontent.com/agenticluke/claude-code-boost/main/scripts/boost.sh -o ~/scripts/boost.sh && chmod +x ~/scripts/boost.sh

Use

In Claude Code, just say boost this on any hard task, or let the skill auto-fire on multi-file changes and stubborn bugs.

Headless:

boost.sh "Fix the failing auth flow in this repo" ~/my-project

The scaffold runs: gate first → plan → act → run the gate (max 5 loops) → adversarial reviewer must re-run the gate and fail to refute → verdict.

If the worker never reaches a passing gate, the script fails closed and says so. No silent green lights.

Does it actually work?

Head-to-head on planted-bug code with real test suites as the judge:

Run	Plain Sonnet	Boosted Sonnet
3 hidden bugs, 5 tests	4/5 (missed the caller-mutation bug)	5/5
Pricing edge cases	mixed across runs	6/6, reviewer confirmed

Stress testing also caught two failure modes in our own first version (a worker that silently did nothing, and a reviewer that almost approved an empty diff). Both are fixed and now guarded against: the verdict requires proof the check actually ran.

Research this is built on

• Planner-Actor-Rater scaffolds measurably lift Sonnet-class models: arxiv.org/html/2606.08529v1
• Role-split pipelines let an 8B model beat a 33B model: arxiv.org/abs/2604.11465
• Why reflection loops need an oracle: The Harness Is Not the Model

MIT licensed. Built and battle-tested with Claude Code.