Prompt Master

Generate prompts using the Anthropic Prompt Template technique, which structures content to leverage LLM attention patterns (beginning and end receive highest attention).

Procedure

Step 1: Analyze the User's Request

Determine whether the user wants to:

• Create a new prompt from scratch.
• Improve an existing prompt by restructuring it.

If the request is ambiguous, ask the user to clarify the target use case and desired output format before proceeding.

Step 2: Identify the Use Case

Match the user's request to one of these use case types to determine required and optional sections:

Use Case	Required Sections	Optional Sections
Simple Q&A	Task Context, The Ask	Output Formatting
Chatbot	Task Context, Tone, Rules, History, The Ask	Examples
Data Extraction	Task Context, Rules, Examples, The Ask, Output Formatting	Background Data
Document Analysis	Task Context, Background Data, Rules, The Ask, Output Formatting	Examples
Complex Agent	All sections	-

If the request does not fit any use case, default to Complex Agent (all sections) and omit sections that are clearly irrelevant.

Step 3: Read the Template

Read assets/prompt-template.ts and use its structure as the base for the new prompt. Copy the section ordering exactly — do not rearrange sections.

Step 4: Construct Each Section

Build the prompt by filling in each required section in this exact order:

1. Task Context (beginning) — Define the role and primary objective. Place this first; LLMs give highest attention here.
2. Tone Context (optional) — Set the communication style. Include only if the use case requires it.
3. Background Data (middle) — Wrap reference documents in XML tags. See references/xml-tags.md for tag conventions.
4. Rules — Write explicit instructions and edge case handling as a bullet list.
5. Examples (few-shot) — Provide 1-3 input/output examples. Do not exceed 3.
6. Conversation History (if multi-turn) — Include only for chatbot or agent use cases.
7. The Ask (end) — Place the actual user request near the end. LLMs give highest attention here.
8. Thinking Instructions (end) — Add chain-of-thought trigger after The Ask.
9. Output Formatting (end) — Specify the desired output structure last.

For detailed descriptions and templates for each section, read references/section-details.md.

Step 5: Apply Quality Checks

Verify the constructed prompt against these rules:

• Critical content placement: Task Context is at the beginning. The Ask, Thinking Instructions, and Output Formatting are at the end. Confirm no critical instructions are buried in the middle.
• XML tag consistency: Each content type uses exactly one tag name throughout. Tags are not nested more than one level deep. See references/xml-tags.md for conventions.
• Example count: No more than 3 few-shot examples are included.
• Edge case coverage: Rules section addresses at least one edge case (e.g., missing input, ambiguous request, out-of-scope question).
• Output format specified: The prompt ends with an explicit output format instruction.

If any check fails, fix the prompt before presenting it to the user.

Step 6: Present the Prompt

Return the final prompt to the user. If the user requested an improvement to an existing prompt, highlight what changed and why.

Error Handling

• Ambiguous use case: If the user's request does not clearly map to a use case type, ask: "What is the primary task this prompt should accomplish?" Do not guess.
• Missing context: If the user wants Background Data or Examples but has not provided source material, ask for it before constructing those sections.
• Multi-turn vs. single-shot unclear: If the prompt could be either, ask: "Will this prompt be used in a single exchange or a multi-turn conversation?" This determines whether to include History and Conversation History sections.
• Existing prompt is unstructured: When improving an existing prompt, first identify which sections are present (even if unlabeled), then restructure into the correct order. Do not discard content — relocate it to the appropriate section.