claude-prompt-architect

A focused Claude Code subagent for designing and reviewing prompts that target the Claude 4.x family — system prompts, agent frontmatter, skill/tool descriptions, and RAG patterns.

Most "make my prompt better" advice is generic or GPT-era. This agent encodes the Claude-specific conventions — system-vs-user split, XML structure, dropping "think step by step," positive framing, prefill deprecation, and cache_control placement — into a single subagent you can dispatch whenever you're writing or fixing a prompt.

What it does

• Designs prompts: separates stable identity (→ system) from volatile task/data (→ user), structures with XML tags, and places cache breakpoints correctly.
• Reviews prompts: returns a verdict + highest-leverage fixes, each tied to the principle behind it.
• Authors agent frontmatter: examples, explicit tool list + model, Key Distinctions / Boundaries, kept lean.

Install

Drop the agent into your Claude Code agents directory:

# user-level (all projects)
mkdir -p ~/.claude/agents
curl -o ~/.claude/agents/prompt-architect.md \
  https://raw.githubusercontent.com/Michael-WhiteCapData/claude-prompt-architect/main/agents/prompt-architect.md

Or project-level: copy agents/prompt-architect.md into your repo's .claude/agents/.

Use it

Just ask Claude Code to design or review a prompt — it will route to the agent:

"Design the system prompt for an email-classification endpoint."

"This agent over-triggers — review the frontmatter and tell me why."

Or invoke explicitly: "Use the prompt-architect agent to review prompts/triage.md."

The checklist it applies

• Stable identity/format → system; variable task/data → user; long docs at the top of the user turn, query after.
• XML tags for structure; don't mix XML and markdown headers inconsistently.
• No "think step by step"; no ALL-CAPS/"MUST" walls; positive framing over negative.
• No volatile content (timestamps, IDs) in cached system prompts; cache_control after large stable blocks.
• No prefill on the last assistant turn — use Structured Outputs.
• Agent frontmatter: 2–3 examples, explicit tools + model, Key Distinctions + Boundaries, under ~100 lines.

License

MIT © Michael Tierney. Contributions and pattern suggestions welcome.