coldHat Lite

Prompts often fail because they are cluttered, repetitive, or internally inconsistent. This skill acts as a linter for your agent instructions, transforming messy rule piles into structured contracts while flagging systemic logic failures. It analyzes system messages, CLAUDE.md files, and AGENTS.md files to ensure your models receive clear, non-conflicting directives.

What it does

• Standardize structure organizes rules into specific Role, Method, Rules, Output, and Locked sections.
• Conflict detection identifies contradicting instructions that cause model hallucination or refusal.
• De-duplication merges redundant rules that waste context window and dilute instruction following.
• Hedge removal converts weak phrases like "try to" into firm imperatives the model can follow.
• Security audit flags API keys, tokens, or passwords accidentally embedded in prompt text.

How it works

1. Input analysis reads every bullet, numbered list item, and imperative sentence as a distinct rule.
2. Classification maps each instruction to its functional part (e.g., Method vs. Output).
3. Logic check scans for contradictions, hedges, missing output formats, and leaked secrets.
4. Correction returns a rewritten, tightened prompt followed by a numbered checklist of specific fixes.

Frameworks & tools

Works with any text-based prompt format including Markdown, system messages, CLAUDE.md, and AGENTS.md.

Why this beats prompting it yourself

Writing a prompt to fix another prompt often introduces new hallucinations or loses original intent. This skill uses a strict sorting methodology and a specific diagnostic syntax to ensure no instruction is silently dropped or wrongly merged.

Use cases

• Agent refactoring when a complex agent stops following its core instructions.
• Instruction linting before deploying a system prompt to a production environment.
• Context optimization to reduce token usage by removing duplicate and wordy directives.
• Security hardening to catch hardcoded secrets before they reach the model.

Known limitations

Does not support prose-only inputs without clear instructions. The Lite version relies on static analysis and does not run live model performance benchmarks.