LLM Governance

Purpose

Enforce LLM content governance for all LLM-facing files using TERSE mode defaults, rule-driven validation, and deterministic tooling.

Apply rules from skills/llm-governance/rules/99-llm-prompt-writing-rules.md and related governance rule files through standardized validators instead of ad-hoc scripts. Provide operational health reporting, capability matrix generation, and structural compliance checking for agents, skills, commands, and rules.

IO Semantics

Input: LLM-facing markdown and configuration files.

Output: Governance findings, severity classifications, suggested edits, and updated files when explicitly approved by higher-level commands.

Side effects: Backups created by orchestration commands before modifications; no direct writes required when running validation only.

Deterministic Steps

1. Toolchain Validation

• Use tool_checker.py to confirm availability of required tools and select fallbacks.
• Abort governance execution for this skill when critical tools are missing and cannot be replaced safely.

2. Target Selection

• Select LLM-facing files using directory classification:
  • commands/**/*.md
  • skills/**/SKILL.md
  • agents/**/AGENT.md
  • rules/**/*.md
  • CLAUDE.md
  • AGENTS.md
  • .claude/settings.json
• Exclude non LLM-facing directories such as documentation, examples, tests, IDE metadata, and backup directories.

3. Automated Validation

• Run python3 skills/llm-governance/scripts/validator.py <directory> across the selected scope.
• Validator uses skills/llm-governance/scripts/config.yaml as Single Source of Truth (SSOT) for all validation rules.
• For each file, detect:
  • Body bold markers outside code blocks.
  • Emoji and decorative Unicode characters.
  • Narrative paragraphs and conversational patterns.
  • Missing or malformed frontmatter for skills, agents, commands, rules, and memory files.
• Classify violations by severity using rule definitions from skills/llm-governance/rules/99-llm-prompt-writing-rules.md.

4. Content Normalization Guidelines

• Enforce TERSE mode:
  • Rewrite narrative paragraphs into imperative directives.
  • Remove conversational fillers and hedging language.
  • Maintain high information density and precise terminology.
• Enforce formatting rules:
  • Remove non code bold markers from body content.
  • Remove emoji and non-essential decorative Unicode characters.
  • Preserve code blocks and required technical symbols.
• Enforce structural rules:
  • Ensure required frontmatter fields and section ordering per directory classification.
  • Normalize heading levels and list formatting for clarity and determinism.

5. Operational Health and Matrix Reporting

• Generate agent and skill capability matrices using agent-matrix.sh and skill-matrix.sh to snapshot capability-level, loop-style, and style coverage.
• Run structure-check.sh to validate taxonomy-rfc compliance (layer: execution annotations, absence of legacy COMMAND.md files).
• Correlate governance findings with operational metadata for health reports and rollback candidate identification.

6. Integration with Orchestration Commands

• Delegate bulk analysis, candidate generation, backup creation, and writeback decisions to /llm-governance and agent:llm-governance.
• Use this skill to interpret validator results, derive rewrite strategies, and keep governance behavior aligned with rule files.

Validation Criteria

• No body bold markers outside code blocks in LLM-facing files.
• No emoji or decorative Unicode characters in governed content.
• Communication is terse, directive, and TERSE-mode compliant.
• Required frontmatter fields and sections are present for each governed directory classification.
• All governance violations are either resolved or documented with justification in governance reports.