Skill: generate-guards

Trigger: "generate guards" · "add guards" · "create project guards" · "what guards should I add" · "guards from my errors"

Why this matters

The 4 generic guards that ship with Clankbrain (NULL_BEFORE_ACCESS, HARDCODED_URL, RAW_ERROR_EXPOSURE, CONSOLE_LOG_LEFT_IN) are starting points. The ones that actually prevent your mistakes are derived from your own error history. Every entry in error-lookup.md and regret.md is a candidate guard — a mistake the codebase could have caught automatically.

This skill converts your error history into automated prevention.

Steps

Step 1 — Read existing guards

• Read .claude/rules/guard-patterns.md
• List guard IDs already present — don't propose duplicates
• Note which file types and patterns are already covered

Step 2 — Mine error history for guard candidates

From error-lookup.md (highest priority — already proved costly):

• Read .claude/memory/error-lookup.md
• For each known error: could a grep pattern have caught this before it ran?
• If yes → strong candidate; mark it as [from error history]

From regret.md (rejected approaches that might slip back in):

• Read .claude/memory/tasks/regret.md
• For each rejected approach: is there a code pattern that signals this approach being used?
• If greppable → candidate; mark it as [from regret log]

From lessons.md (recurring patterns):

• Read tasks/lessons.md
• Look for lessons mentioning a specific code pattern to avoid
• If a pattern appears in 2+ lessons → strong candidate

From the stack (general best practices for detected tech):

• Check what languages are in the project (Glob for .java, .js, .py, etc.)
• Apply stack-specific candidates: SQL concatenation (Java/Python), async without await (JS), missing null checks (Java), etc.
• Use Grep to confirm the pattern actually exists in this codebase before proposing it

Step 3 — Propose guards

For each candidate, format it as:

## GUARD_ID
- **Check**: [one sentence: what this guards against]
- **How to scan**: Grep for `[exact regex pattern]` in [file scope, e.g. "*.java", "*.js"]
- **Files**: [specific files or file types]
- **Why**: [what breaks if this pattern appears — reference the error or lesson that caused it]

Show all candidates. For each one:

• Mark source: [from error history], [from regret log], [from lessons], or [best practice]
• State what specific past mistake it would have caught, if known

Prioritize error-history and regret-log guards first — they have proven cost.

Step 4 — Confirm before writing

Do NOT write to guard-patterns.md without confirmation.

Ask: "Add these [N] guards to guard-patterns.md? Say yes to add all, or list which ones to keep."

Step 5 — Write confirmed guards

Append each confirmed guard to .claude/rules/guard-patterns.md with a blank line between guards.

Then report: "[N] guards added. Run Guard Check to scan the codebase for existing violations now."

Closing message

"Guards added to guard-patterns.md. These run automatically on Guard Check and Pre-Ship Check. Re-run this skill after every 5 sessions — your error history grows, and so should your guards."