Act as a prompt editor. Improve my prompt for [MODEL OR TOOL] without changing the actual task or inventing requirements. My current prompt is: [CURRENT PROMPT]. The result I need is: [DESIRED OUTPUT]. The audience and decision context are: [CONTEXT]. Available source material, tools, and constraints are: [INPUTS AND CONSTRAINTS]. A good answer must meet these acceptance criteria: [ACCEPTANCE CRITERIA]. First, diagnose the prompt in a compact table: ambiguity, missing inputs, hidden assumptions, conflicting instructions, unsupported-data risk, and output-format gaps. Then produce: 1. A copy-paste optimized prompt, written in first person, with only necessary bracketed placeholders. 2. A short “how to use it” note that says what to paste in and what not to expect the model to know. 3. A stricter version for high-stakes or repeatable work that requires source grounding, uncertainty labels, and an explicit final checklist. 4. Three targeted follow-up prompts for common failure modes in this task. The optimized prompt must specify the role only when it changes useful expertise, name the exact deliverable and sections, separate source facts from assumptions, set a reasonable output length, and instruct the model to ask up to 3 clarifying questions only when a required input is missing. If research, calculations, code execution, or browsing is needed, state it plainly instead of implying it happened. Before answering, run two checks: verify every instruction supports an acceptance criterion, and flag any desired result that cannot be produced reliably from my supplied inputs. Do not pad the prompt with generic rules or claim one model will always outperform another.
Fill in
| Placeholder | What to enter | Example |
|---|---|---|
| [MODEL OR TOOL] | Enter the AI model or tool you plan to use. | Claude |
| [CURRENT PROMPT] | Paste the prompt that is producing weak or inconsistent results. | Write a plan to improve customer onboarding. |
| [DESIRED OUTPUT] | Describe the specific deliverable you want the model to produce. | A 90-day onboarding improvement plan with prioritized experiments and owners. |
| [CONTEXT] | Enter the audience, purpose, and decision or workflow context. | For a 40-person B2B software company presenting to the head of customer success. |
| [INPUTS AND CONSTRAINTS] | Enter source material, permissions, deadlines, style rules, and unavailable information. | Use only our interview notes and funnel metrics; do not assume engineering capacity beyond one developer. |
| [ACCEPTANCE CRITERIA] | Enter the observable conditions that make an output usable. | Each experiment has a hypothesis, metric, segment, effort estimate, owner, timing, and dependency. |
How to use
- Paste the prompt that failed alongside the exact output you need instead of describing the problem generally.
- Replace bracketed fields with the source material and acceptance criteria the model can actually inspect.
- Test the optimized prompt on one representative case and compare the output against the acceptance criteria.
- Follow up with: “Revise the prompt to prevent this failure: [PASTE FAILURE], while keeping the same inputs and deliverable.”
Variations
Prompt debugger
Use this when you have an output sample and need to isolate why it failed.
My prompt was [PROMPT], and this was the unsatisfactory output: [OUTPUT]. I expected [EXPECTED RESULT]. Diagnose the failure by separating missing instructions, missing source data, model limitations, and formatting errors. Then rewrite the prompt with the fewest necessary changes. Include a test case and pass/fail checklist. Do not assume the model can access facts or tools I did not provide.
Reusable template
Use this for work repeated across clients, projects, or teams.
Turn this one-off prompt [ORIGINAL PROMPT] into a reusable template for [REPEATED TASK]. Inputs available each time are [RECURRING INPUTS], and the final deliverable is [DELIVERABLE]. Use clear bracketed placeholders, a fixed output schema, and validation checks that can be performed from the supplied inputs. Include instructions for missing fields and three example placeholder values. Avoid organization-specific claims.
Research prompt
Use this when a task depends on current, cited external information.
Optimize this research request [CURRENT REQUEST] for [MODEL]. I need an answer for [AUDIENCE] with sources from [SOURCE REQUIREMENTS]. Create a prompt that requires dated citations, separates sourced facts from analysis, reports uncertainty and conflicting evidence, and says when no adequate source was found. The deliverable must be [OUTPUT FORMAT]. Do not ask the model to fabricate citations or imply browsing if it cannot browse.
Tips
- Define acceptance criteria as checks another person could apply, such as required sections, sources, fields, or calculation methods.
- Give the model a source boundary; “use only these notes” prevents plausible but untraceable additions.
- For recurring work, lock the output schema before optimizing tone or length.
- If a task relies on changing facts, require date-stamped sources rather than trying to fix the issue with more adjectives in the prompt.
FAQ
Does a longer prompt always work better?
No. A better prompt removes ambiguity and adds only constraints tied to the deliverable. Unrelated rules can make instructions conflict.
Can an optimizer guarantee accurate answers?
No. It can require evidence, calculations, and uncertainty labels, but source quality and human validation still determine accuracy.
Should I optimize for ChatGPT and Claude separately?
Start with a model-neutral prompt with explicit inputs and output format. Then adjust only if a model consistently mishandles a particular format or tool instruction.