| name | prompt-architecture-engineering |
| description | Designs and refines high-performance prompts, system instructions, agent workflows, and multi-step operating guidance for LLMs. Use when the user asks for a system prompt, agent specification, orchestration plan, instruction hierarchy, subagent workflow, prompt debugging, skill file creation, CLAUDE.md writing, or a cleaner prompt architecture. Also use when the user says "write a system prompt," "design an agent," "fix this prompt," "make this more reliable," "build a skill," "orchestrate these agents," or describes any multi-agent or multi-step LLM workflow. If the output is instructions that an LLM will execute, this skill applies.
|
Prompt Architecture Engineering
Rules
IF the user asks for a prompt or agent specification:
THEN identify role, task boundary, inputs, outputs, and failure modes
before writing a single instruction line.
IF the brief is underspecified:
THEN return a minimal prompt skeleton plus the missing variables. Do not
pad with guesses.
IF writing a system prompt:
THEN separate permanent rules from task-specific context. Encode workflow
order explicitly: what to do first, what to avoid, when to stop.
IF writing a multi-agent workflow:
THEN define each agent's specific mandate, what it does NOT review, its
exact deliverable, and the coordination rules between agents. Use the
agent/subagent template from the reference doc.
IF the user says "fix this prompt" or "make this more reliable":
THEN diagnose the failure mode first. Common failures: scope creep (no
boundary), drift (critical rules buried in prose), fabrication (no
failure handling), repetition (redundant clauses causing conflicting
interpretation). State the diagnosis, then rewrite.
IF adding examples to a prompt:
THEN use one or two concrete examples. Do not use more than three.
Additional examples reduce variance more than they improve quality.
Prompt structure (default)
- Role statement ("You are X").
- Output format (state before workflow; the model anchors on output shape).
- Workflow steps with explicit ordering.
- Hard rules (bullet points, not buried in prose).
- Failure handling ("If you cannot do X, return Y").
- Examples (1-2 max).
Anti-patterns
- Motivational padding ("You are an expert" before "You are a [role]").
- Critical rules in prose paragraphs instead of bullets.
- Conflating goal, workflow, and output format in one section.
- ALL CAPS for emphasis (use bold sparingly or restructure).
- Over-specification that causes drift.
References
