talont-org/design-guards

Design Guards Skill

Take a bug pattern audit report and design architectural guards that make each pattern structurally impossible or instantly caught. This is a multi-pattern analysis driven by a pattern report rather than a single bug investigation.

When to Use

• After /autoskillit:audit-bugs produces a bug pattern report
• User says "design guards", "design defenses", or "guard against these patterns"
• User wants architectural immunity for a set of identified bug patterns

Arguments

Path to a bug pattern audit report (typically under {{AUTOSKILLIT_TEMP}}/audit-bugs/). If not specified, use the most recent {{AUTOSKILLIT_TEMP}}/audit-bugs/bug_pattern_audit_*.md.

Critical Constraints

NEVER:

• Modify any source code files
• Propose bandaid fixes or direct-only patches
• Create files outside {{AUTOSKILLIT_TEMP}}/design-guards/ directory
• Run subagents in the background (run_in_background: true is prohibited)

ALWAYS:

• Use subagents for parallel codebase exploration
• Focus on structural immunity over one-off fixes
• Each guard must be concrete and testable
• All output goes under {{AUTOSKILLIT_TEMP}}/design-guards/ (create if needed)
• Final report: {{AUTOSKILLIT_TEMP}}/design-guards/defense_guards_{YYYY-MM-DD_HHMMSS}.md
• Subagents must NOT create their own files - they return findings in their response text only
• Do not change any code

Workflow

Step 1: Parse the Bug Pattern Report

Read the input report and extract each pattern:

• Pattern name and description
• Affected components and sessions
• The root architectural gap identified

Step 2: Investigate Each Pattern in Parallel

For each pattern (or group of related patterns), launch subagents to explore:

Current State

• Find all current instances of the pattern in the codebase (not just the ones that already caused bugs)
• Map which components are vulnerable to this pattern today

Existing Guards

• What tests, contracts, or validation already exists for this area?
• Why did existing guards miss these bugs?

Architectural Solutions

• How do well-designed parts of the codebase prevent similar issues?
• What structural pattern would make this class of bug impossible?
• What's the minimum change that provides maximum coverage?

Step 3: Design Guards for Each Pattern

For each pattern, produce:

1. Detection scan - A concrete grep/search command or pattern that finds latent instances today
2. Contract test - A test that fails if the pattern re-emerges (description of what it asserts, not code)
3. Structural guard - An architectural change that makes the pattern impossible by construction (if applicable)
4. Standard recommendation - Whether this should become an enforced standard in /autoskillit:audit-defense-standards

Prioritize guards by:

• Coverage: How many bug instances would this single guard have prevented?
• Automation: Can it be enforced in CI/pre-commit, or does it require manual audit?
• Cost: How invasive is the change?

Step 4: Write Report

Ensure {{AUTOSKILLIT_TEMP}}/design-guards/ exists (mkdir -p).

Save to: {{AUTOSKILLIT_TEMP}}/design-guards/defense_guards_{YYYY-MM-DD_HHMMSS}.md (relative to the current working directory)

# Defense Guards: {Report Title}

**Date:** {today}
**Source Report:** {path to bug audit report}
**Patterns Addressed:** {count}

## Summary
{Which patterns get guards, expected coverage, recommended standards}

## Guard N: {Pattern Name}

**Addresses Pattern:** {name} ({X} of {Y} historical bugs)

### Current Exposure
{How many latent instances exist in the codebase today}

### Detection Scan
{Runnable grep/search commands}

### Contract Test
{What the test asserts, where it belongs}

### Structural Guard
{Architectural change that prevents by construction}

### Standard Recommendation
{YES/NO - should this become an audit-defense-standards entry?}
{If YES: one-sentence rule statement}

---

## Standards to Add
{List of guards recommended as permanent standards for audit-defense-standards, with rule statements}

Step 5: Terminal Summary

Output: guard count, coverage stats, how many standards recommended, and report location.