prompt-architect

What it does

Create and enhance production-ready prompts. Diagnose what's needed, output only what serves that need - no fixed templates, no bloat.

How it works

Phase 1: Detect Mode determines whether to Create, Enhance, or Ask based on input such as "Create a prompt for X" or "Improve/enhance this: [prompt]". Phase 2: Diagnosis classifies the type (Agent, Task, Persona, Workflow, Rules, Skill/Expert, Hybrid), assesses complexity (Simple/Moderate/Complex), and identifies gaps (e.g., vague objective, missing boundaries, unclear output, edge cases, etc.). For Enhance Mode, it notes what to preserve (role, structure, constraints, voice) and enforces the key rule that if input has a role, output must have a role. Phase 3: Technique Selection applies methods only when triggered (Mental Models, Thinking Approaches, Patterns, Heuristics, Anti-Patterns, Chain-of-Thought, Few-Shot, Constraint Spec, Role Enhancement). It provides references for CoT, Few-Shot, ReAct, and Role-Based Prompting. Phase 4: Building Blocks instructs what sections to include (Role, Voice, Objective, Mental Models, Patterns, Heuristics, Thinking Approaches, Rules, Anti-patterns, Output) and provides concrete block templates. Phase 5: Key Transformations shows converting patterns to mental models, converting procedures to patterns, and making language more specific while preserving roles. Phase 6: Token Optimization outlines compact output with server-side remapping techniques and when to apply. Phase 7: Validation prescribes checks for structure, content, preservation, and expert-transfer tests. Phase 8: Output directs returning a directly usable prompt, or asking one clarifying question if unclear.

When to use it

Use when creating new prompts, enhancing principle files, improving system instructions, optimizing CLAUDE.md, restructuring verbose prompts, adding patterns to workflows, or defining agent behaviors.

What it can touch

The skill defines blocks and techniques, including Role, Voice, Objective, Mental Models, Patterns, Heuristics, Thinking Approaches, Rules, Anti-Patterns, and Output blocks. It references and employs methods like Chain-of-Thought, Few-Shot Learning, ReAct, and Role-Based Prompting. It emphasizes not using fixed templates and keeping only what serves the task.

Caveats

No explicit licensing or risk statements beyond internal guidance; focuses on methodology and structure rather than guaranteeing outcomes. It requires careful adherence to preserving input roles when present and applying transformations only when triggered by the input mode and type.