omer-faruk-g/HighPer

A Claude Code skill that measures and improves how well Claude Code works: correct results, fewer wasted turns, shorter waits, fewer errors and retries, and more reliable long or unattended runs. It is the sister of MinToken. MinToken lowers token use, HighPer raises quality and reliability. Each works on its own, and when their advice conflicts HighPer shows both sides instead of choosing for you.
- **Measure first, then suggest.** A lever (H1–H72) is suggested only after your own data shows the problem it fixes.
- **No promises in numbers.** No "X% faster", "2×" or "zero mistakes". You measure your own before and after, and the statistics tools tell you how much a small sample can show (rule of three, smallest detectable difference).
- **Nothing changes silently.** Changes to settings.json, hooks, CLAUDE.md, permissions, model or effort are shown as a diff and wait for your explicit approval. HighPer never commits or pushes on its own and never loosens permissions.
- **Labelled claims.** [D] official docs or Anthropic publication · [G] observed in a session · [R] third-party report, unverified · [I] inference, unverified.
- **Python standard library only, no network.** Every script supports --help and --dry-run. Session transcripts are read locally, and reports contain counters and durations only, never message text.
The skill itself (SKILL.md and references) is written in Turkish.
Install
Windows (PowerShell):
Claude Code picks the skill up automatically. It triggers when you say Claude Code is slow, stuck, making mistakes, answering late, or dropping long tasks halfway. You can also call it directly with /highper. Tested with Python 3.13; written for Python 3.7 or later.
Commands
| Command | What it does |
|---|---|
| /highper | General audit plus a short baseline (audit_perf.py, perf_baseline.py) |
| /highper hunt | Inefficiency hunt: at most 10 findings, each with evidence, ordered correctness > reliability > latency > cost |
| /highper plan | Sizes the task (XS/S/M/L) and suggests the verification steps and profile |
| /highper review | Prompt for an independent reviewer: fresh context, read-only, at most 2 rounds |
| /highper measure | Quality metrics, before/after comparison, smallest detectable difference |
| /highper why-slow | When quality or speed suddenly drops, checks the version, model, effort, plugins and the documented incident windows |
| /highper route | Model and effort suggestion from five axes: ambiguity, scope, verifiability, risk, novelty |
| /highper sentinel | Regression sentinel: a fixed mini task set, logged per version, model and effort |
| /highper zero | Zero-Escape profile: lists every gate and its blind spots |
| /highper escape-log | Logs bugs found after delivery and reports the escape rate with an upper bound |
| /highper report | Markdown report: axis → finding → suggestion → trade-off → rollback |
| /highper gradle | Gradle/Fabric settings, version matrix draft, error blocks from logs |
| /highper lint | Lints CLAUDE.md, MEMORY.md and task prompts |
Optional hooks (Zero-Escape gates)
Both hooks do nothing until a project contains its config file. Install them only after checking the schema against your Claude Code version's hooks page. Snippet for ~/.claude/settings.json (replace with the skill path):
/.claude/highper-scope.json is the scope guard. Writes outside allow or inside deny are blocked with exit code 2. It adds a restriction and never approves anything: