AntiVibe - Code Learning & Audit Framework

Purpose

AntiVibe generates learning-focused explanations or architectural audits of any code — AI-generated, legacy, or otherwise. It helps developers understand:

• What the code does (functionality)
• Why it was written this way (design decisions)
• When to use these patterns (context)
• What alternatives exist (broader knowledge)

Works on any codebase — you don't need recent git history or AI-authored files.

When to Use

Use AntiVibe when:

1. Manual invocation: User types /antivibe or "deep dive"
2. Post-task learning: After a feature/phase completes, user wants to learn from it
3. Legacy codebases: User wants to understand existing code they didn't write
4. Proactive: User says "explain what AI wrote", "walk me through", "audit this", or points at a file/directory

What AntiVibe Produces

Output saved to deep-dive/ folder as markdown:

deep-dive/
├── auth-system-2026-01-15.md
├── api-layer-2026-01-15.md
└── database-models-2026-01-15.md

The exact sections depend on the output mode (see Output Mode):

Configuration

Known Concepts (Skip List)

Concepts listed here will not be explained in full — the explainer will only note that they were used and in what context. Edit this list to match your current knowledge.

known_concepts:
  - async/await
  - React hooks
  - REST APIs

Output Mode

Controls how much detail is generated per run. Default is compact to keep token costs low.

output_mode: compact

Override inline in your request:

• "/antivibe full", "full deep dive", "include resources" → full mode
• Default: compact

Default Skill Level

Sets the explanation depth when no level is specified in the request. Options: junior, mid, senior. Default: mid.

default_level: mid

Level can also be specified inline in the request:

• "explain for a junior", "I'm new to this" → junior
• "I know the basics", "mid level" → mid
• "senior mode", "skip the basics", "just the trade-offs" → senior

Workflow

Step 0: Apply User Configuration

Before analyzing, read the configuration above:

• Load the known_concepts skip list. Any concept in this list will be acknowledged in one sentence instead of fully explained.
• Detect the skill level: check the user's request first (inline phrases take priority), then fall back to default_level. Apply this level consistently throughout the entire output.
• If level = senior, route to agents/auditor.md instead of continuing this workflow.

Step 1: Identify Code to Analyze

Use the first applicable mode:

1. Explicit — User named specific files, a directory, or a module in their request → use those directly. No git needed. Example: "explain src/auth/" or "walk me through api/routes.py".

2. Recent — No explicit target given, project is a git repo, and git diff HEAD has output → use those changed files (current behavior for post-AI-task learning).

3. Scan — No explicit target, no usable git diff (legacy project, no recent changes, or not a git repo) → ask the user: "Which file, directory, or module would you like to analyze?" Do not attempt to guess.

The code does not need to be AI-generated. AntiVibe analyzes any code.

Step 2: Analyze Code Structure

For each file:

• Identify main purpose and responsibilities
• Note key functions, classes, modules
• Identify design patterns used (factory, singleton, observer, etc.)
• Find any complex logic or algorithms

Step 3: Explain Concepts

For each concept/pattern found:

• What: Plain-language explanation
• Why: Why this approach was chosen over alternatives
• When: When to use this pattern (with context)
• Alternatives: Other approaches and trade-offs
• Prerequisites: 2–4 foundational concepts the developer must understand first (e.g., "To understand JWT, you need: HTTP request/response, Base64 encoding, cryptographic signing")

Step 4: Find External Resources

Only run this step in full mode. Skip entirely in compact mode.

Search for and include:

• Official documentation for libraries/frameworks used
• Quality tutorials or blog posts
• Video resources (if available)
• Related concepts for further learning

Step 5: Generate Output

Create markdown file in deep-dive/ folder:

• Name format: [component]-[timestamp].md
• Detect output mode from the request or output_mode config (default: compact)
• Compact mode: Use the compact template. No line-by-line, no resources, no Next Steps. Max 5 files — if more are in scope, summarize extras in one line each and offer to go deeper.
• Full mode: Use the full template from templates/deep-dive.md. Include all sections. No 5-file limit — analyze every file in scope; for very large inputs, split the output across multiple deep-dive files rather than truncating.
• Make it educational, not just descriptive

Auto-Trigger Configuration

AntiVibe can be configured to auto-trigger via hooks:

• SubagentStop: After a Task completes a feature
• Stop: At session end

To enable auto-trigger, configure hooks in your project (see hooks/hooks.json).

Principles

1. Why over what - Always explain design decisions
2. Context matters - Explain when/why to use patterns
3. Curated resources - Quality links, not random Google results
4. Phase-aware - Group by implementation phase
5. Learning path - Suggest next steps for deeper study
6. Concept mapping - Connect code to underlying CS concepts

Dependencies

Optional scripts in scripts/ folder:

• capture-phase.sh - Detect implementation phase boundaries
• analyze-code.sh - Parse code structure
• find-resources.sh - Search for external resources
• generate-deep-dive.sh - Create markdown output

These are helpers - you can also do everything via direct code analysis.

Examples

Input: "Explain the auth system Claude wrote" (recent AI code) → Mode: Recent (git diff). Output: deep-dive/auth-system-2026-01-15.md

Input: "Walk me through src/payments/" (explicit target — legacy codebase) → Mode: Explicit. Analyzes files in that directory directly, no git needed.

Input: "Deep dive" (no target, legacy project with no recent changes) → Mode: Scan. Asks: "Which file or module would you like to analyze?"

Input: "Audit this, just the trade-offs" (senior mode) → Routes to agents/auditor.md. Produces architectural audit, not an explanation.