→ Workflow Design
Workflow DesignIntermediate

Prompt Refinement Loop

Use when a prompt is producing mediocre or inconsistent output and you want a repeatable loop to diagnose and improve it instead of guessing.

Knowledge WorkersExecutives
Save to My LibraryAgent Skills compatible

Prompt Refinement Loop

When AI output is disappointing, the instinct is to rewrite the prompt from scratch. This skill replaces that with a disciplined loop: diagnose the failure, change one thing, and measure whether it helped.

When to use this skill

Use this skill when:

  • A prompt mostly works but produces inconsistent or low-quality output.
  • You are about to reuse a prompt many times and want it reliable first.
  • You want to teach a team a repeatable way to improve prompts.

Inputs needed

  • The current prompt.
  • 2 to 3 real examples of output that disappointed you.
  • A short description of what "good" output looks like.
  • The model or tool you are using.

Process

  1. Write down the success criteria for the output in concrete terms.
  2. Run the current prompt on a fixed test input and save the output.
  3. Diagnose the gap: is it missing context, format, constraints, or examples?
  4. Change exactly one thing per iteration (context, format, constraint, or example).
  5. Re-run on the same test input and compare against the criteria.
  6. Keep changes that improve results; revert changes that do not.
  7. Stop when output meets the criteria on three different inputs.

Prompt or workflow

You are helping me refine a prompt using a controlled loop.

Current prompt:
"""
[PASTE PROMPT]
"""
What good output looks like: [DESCRIBE]
Disappointing examples:
"""
[PASTE 1-3 OUTPUTS]
"""

Do this:
1. CRITERIA: restate success criteria as a short checklist.
2. DIAGNOSIS: classify the main failure as missing context / format / constraint
   / example, with one sentence of evidence.
3. ONE CHANGE: propose a single change that targets the diagnosis.
4. REVISED PROMPT: output the full revised prompt with only that change.
5. TEST PLAN: the fixed input I should re-run it on to check improvement.

Rules:
- Change only one variable so I can tell what worked.
- Do not rewrite the whole prompt unless I ask.

Quality checklist

  • Success criteria are written down before changes are made.
  • Each iteration changes exactly one variable.
  • The same test input is used to compare iterations.
  • Changes that do not help are reverted, not stacked.
  • The final prompt is verified on at least three different inputs.

Common mistakes

  • Changing several things at once, so you cannot tell what helped.
  • Testing on a new input each time, hiding regressions.
  • Declaring success on one lucky output.

Example output

Criteria: output is a 5-bullet summary, no preamble, under 100 words.
Diagnosis: format failure — model adds a preamble paragraph.
One change: add explicit "Respond with only the 5 bullets, no preamble."
Revised prompt: [full prompt with the single added constraint]
Test plan: re-run on the saved earnings-call transcript input.
  • Workflow Audit — when the problem is the workflow, not the prompt.
  • Research Brief Generator — a prompt worth refining with this loop.

Attribution

This skill was created by Vectory and is licensed under CC BY 4.0.

Source: https://vectory.io/skills/prompt-refinement-loop

Attribution: "Prompt Refinement Loop" by Vectory.