3.2 KiB

Raw Blame History

Tool Patterns Catalog

Patterns for building robust, self-correcting Claude Code tools (skills, agents, commands).

Pattern Summary

Pattern	Purpose	When to Use
Audit	Smart implicit audits	Modifying existing tools, vague requests
Verification	Mandatory completion checks	Before claiming any work complete
Brainstorming	Design decision refinement	Design decisions, new concepts
Reversion Protection	Prevent behavior regression	All tools with defined behaviors
Decisions Record	Capture audit decisions	During any audit workflow
Deterministic Scripting	Shift mechanical work to scripts	Planning/design, identifying automation
Adaptive Verbosity	Scale output to complexity	Skills with variable complexity (simple to complex scenarios)
Role-Workflow	Multi-workflow orchestration	Plugins with 2+ subagent workflows needing synthesis

Quick Decision Guide

Starting a new tool?

Brainstorm the design first (if non-trivial)
If orchestrating 2+ subagent workflows with synthesis → use Role-Workflow pattern
Identify scripting opportunities (mechanical, repeatable tasks)
Define invariants and golden examples (reversion protection)
Build with verification hooks

Modifying an existing tool?

Check for prior decisions (.decisions/)
Run implicit audit before changes
Verify against golden examples
Record any new decisions made

Reviewing a tool?

Run audit workflow
Record decisions made
Update golden examples if behavior changes were intentional

During planning/design?

For each component, ask: "Is this mechanical, repeatable, and unambiguous?"
If yes → recommend a script (see deterministic-scripting pattern)
Choose language based on project context

Integration Points

These patterns work together:

Brainstorming ──→ Design decisions captured
        │
        ↓
    Planning ──→ Script opportunities identified
        │        Multi-workflow orchestration? → Role-Workflow pattern
        ↓
    Building ──→ Invariants + golden examples defined
        │        Scripts implement mechanical tasks
        │        Role-Workflow: Task files + dispatch structure
        ↓
   Modifying ──→ Audit runs implicitly
        │        │
        │        ↓
        │   Decisions recorded
        │        │
        ↓        ↓
 Verification ──→ Golden examples checked
                 Scripts provide deterministic evidence
                 Role-Workflow: Synthesis step verifies completeness

Model Guidance

Pattern	Recommended Model
Audit execution - structure	Haiku
Audit execution - content/quality	Opus
Decision recording	Haiku
Brainstorming	Opus
Verification	Haiku
Golden example comparison	Opus
Role-Workflow - research	Haiku
Role-Workflow - processing	Sonnet
Role-Workflow - judgment	Opus
Cleanup	Haiku

3.2 KiB Raw Blame History