Anthropic 1p Prompt Optimizer

Overview

Turn vague “帮我改一下提示词/Make this prompt better” into a repeatable workflow:

1. clarify goals + constraints, 2) restructure the prompt using proven building blocks, 3) add formatting/examples/guardrails, 4) propose a small test plan for iteration.

Workflow (1P-style)

Step 0: Intake (ask only what’s missing)

Request:

• Current prompt (system + user), and where it runs (UI / API / multi-turn).
• Target behavior: success criteria + non-goals.
• Inputs the prompt will receive (and typical edge cases).
• Required output format (plain text / XML / JSON) + strictness.
• Common failures (hallucination, wrong tone, ignored constraints, formatting drift).

Step 1: Diagnose quickly

Check for:

• Ambiguous verbs (e.g., “analyze”, “optimize”) without definitions.
• Missing role/audience/tone context.
• Data mixed with instructions (needs XML delimiting).
• Output format underspecified (needs tags / JSON schema / prefilling).
• No examples for tricky formatting or edge cases.

Step 2: Apply building blocks (pick only what helps)

Preferred building blocks from the tutorial (details in references):

• Clear + direct instructions; explicitly ask for what you want.
• Role prompting (optionally include intended audience).
• Separate variable input from instructions using XML tags.
• Enforce output format via XML/JSON + optional prefilling (“speaking for Claude”).
• Add “precognition” for multi-step reasoning (thinking step-by-step).
• Use few-shot examples to lock tone and formatting.
• Reduce hallucinations with “give an out” + evidence-first extraction.
• For complex prompts, use the 1P complex-prompt element ordering.

Step 3: Ship revisions as artifacts

Always output:

• Revised system prompt (if needed) and revised user prompt.
• Any few-shot examples and XML tag wrappers.
• A short change log (what changed and why; 5–10 bullets max).
• 3–5 test cases with pass/fail criteria.

Reference Files (load as needed)

• references/anthropic-1p-cheatsheet.md: chapter-by-chapter patterns and “why it works”.
• references/complex-prompt-template.md: the 1P complex prompt element list + skeleton.
• references/guardrails-hallucinations.md: “out”, evidence-first, placement, temperature tips.
• references/tool-use-xml.md: function-calling XML formats (<function_calls>, <function_results>).
• references/iteration-checklist.md: prompt chaining + evaluation loop.