wowclaude

wowclaude

Install once, get better Claude Code responses forever.

A Claude Code plugin that uses multi-layered hooks to inject evidence-based quality principles. No configuration needed.

TL;DR

Claude Code confidently makes up internal details (bit flags, private APIs). This plugin detects risky queries and forces "iirc" markers on unverified claims. Zero config.

Installation

/plugin marketplace add jbj338033/wowclaude
/plugin install wowclaude

The 5 Principles

#	Principle	What it does
1	certainty	Prevents confabulation on high-risk topics. Forces "iirc" markers on unverified internal details.
2	investigation	Forces reading before writing. Verify with tools, or say you can't.
3	reasoning	Step-back reasoning on hard problems. Understand why, then solve.
4	precision	Matches effort to complexity. Less with qualifiers > more with false confidence.
5	verification	Self-checks against the original request before finalizing.

Architecture

┌─────────────────────────────────────────────────────────┐
│                    wowclaude hooks                      │
├─────────────────────────────────────────────────────────┤
│                                                         │
│  SessionStart          → base quality principles        │
│  │                       (always active)                │
│  ▼                                                      │
│  UserPromptSubmit      → risk detection per message     │
│  │  ├─ CRITICAL tier   → internal details, bit flags    │
│  │  ├─ CAUTION tier    → version-specific, deprecated   │
│  │  └─ (no trigger)    → normal queries pass through    │
│  ▼                                                      │
│  PreToolUse            → edit guard                     │
│     └─ Edit/Write      → "did you read this file?"     │
│                                                         │
└─────────────────────────────────────────────────────────┘

Hook Details

Hook	Event	Trigger	Injection
session-start.sh	SessionStart	Every session	5 quality principles (~90 words)
user-prompt.sh	UserPromptSubmit	Each message	CRITICAL: internal details (~60 words) / CAUTION: version-specific (~40 words) / none
pre-tool-use.sh	PreToolUse	Before Edit/Write	Investigation reminder (~20 words)

Multi-Tier Risk Detection

The UserPromptSubmit hook classifies queries into risk tiers:

CRITICAL — Internal implementation details (highest confabulation risk):

• Bit flags, enum values, struct fields, bytecode formats
• Private/undocumented APIs, compiler internals
• Runtime flags, syscall tables, fiber internals

CAUTION — Version-specific or platform-specific behavior:

• Breaking changes, deprecated features, migration paths
• Version-boundary behavior, platform differences

Safe — No intervention needed for general queries.

Test Results

Tested with identical prompts before and after plugin installation:

Test	Without wowclaude	With wowclaude
"List React fiber flags with bit values"	Lists specific hex values as fact	Prefixes values with "iirc", points to source file
"What's process._linkedBinding('http_parser') signature?"	Lists full API surface confidently	States values are unverified, recommends source check
"What is git stash?"	Concise answer	Concise answer (no unnecessary intervention)

What's NOT Included (and Why)

• Expert persona ("You are a senior engineer...") — No measurable accuracy improvement. Just makes responses longer.
• "Think step by step" — Modern Claude models already do chain-of-thought internally. Redundant instruction.
• Magic phrases ("Take a deep breath", "This is important for my career") — No generalizable effect.
• Forced long reasoning chains — Shorter, focused chains often outperform verbose ones.

Research

See docs/research.md for detailed citations and reasoning.

Toggle / Uninstall

/plugin disable wowclaude    # disable temporarily
/plugin enable wowclaude     # re-enable
/plugin uninstall wowclaude  # remove completely

License

MIT