Prompt Creator

Workflow

Follow this process for every prompt engineering request:

1. Clarify the goal (if not already clear)

• What is the task the prompt must accomplish?
• Who/what will execute it (Claude, GPT, a specific model)?
• How will the output be used (human reads it / parsed by code / passed to another model)?
• What does "good" look like? Any examples of ideal output?

2. Select the right techniques Determine which techniques apply based on the request type:

Read references/techniques.md for full technique details and examples (including Prompt Chaining). Read references/claude-specifics.md when targeting Claude specifically (XML tags, critical long-context ordering, extended thinking, system vs human split, prompt injection defense). Read references/prompt-types.md for ready-to-adapt templates for common prompt categories (including RAG, Prompt Chaining, Agentic).

3. Draft the prompt

• Use the relevant template from prompt-types.md as a starting point
• Apply techniques from techniques.md
• Apply Claude-specific best practices from claude-specifics.md when relevant
• Follow the quality checklist below before delivering

4. Deliver and offer iteration Present the finished prompt in a code block. Briefly explain key design choices (1-3 sentences). Offer to:

• Add few-shot examples
• Adjust tone, length, or format constraints
• Create test cases to validate the prompt

Quality Checklist

Before delivering any prompt, verify:

• [ ] Task is unambiguous — a stranger could execute it correctly
• [ ] Output format is explicitly specified (or output primed with a partial phrase)
• [ ] Both "do" and "do NOT" constraints are present where relevant
• [ ] Model has an "out" for cases it can't answer (e.g., "respond with 'not found' if...")
• [ ] Variables/placeholders use a consistent convention: [PLACEHOLDER] or {{placeholder}}
• [ ] Long prompts use XML tags or clear section headers to separate concerns
• [ ] No conflicting instructions; key constraints repeated at the end if critical
• [ ] Role/persona is assigned if tone or expertise level matters
• [ ] Few-shot examples included if output style is non-obvious
• [ ] Reasoning requested (CoT) if the task involves multi-step logic
• [ ] Citations requested if factual accuracy and grounding are critical

Diagnostics: Improving a Bad Prompt

When a user brings an underperforming prompt, diagnose before prescribing:

Hallucination / making things up

• Add a fallback: "If the answer isn't in the context, respond with 'Not found.'"
• Add citation requirement: "After each claim, cite the source. Do not make claims you can't cite." (forcing citations forces double errors to fabricate)
• For RAG: add explicit grounding instructions and prompt injection defense (see claude-specifics.md)

Wrong format / structure

• Be more explicit about output format; provide a template or example output

Inconsistent results

• Add few-shot examples; tighten constraints; reduce degrees of freedom

Too long / too short

• Add explicit length constraint: "Respond in exactly 3 bullet points" / "Keep response under 100 words"

Wrong tone

• Assign a specific persona; add explicit tone descriptors; provide a style example

Ignoring part of the instructions

• Use XML tags to separate sections; repeat critical instructions at the end (recency bias — the model gives more weight to content near the end)
• Simplify: break into multiple smaller prompts if complexity is the issue

Reasoning errors

• Add chain-of-thought: "Think step by step before answering"
• Use extended thinking mode if available

Output Format

Present the final prompt in a fenced code block:

[FINAL PROMPT HERE]

Then add a short Design notes section (2-4 bullets) explaining the key choices made.