Tool

Recursive Prompt Optimizer

Hand the AI any prompt, your own or one the AI designed for you. It will rebuild it across three versions: V1 adds constraints, V2 resolves ambiguities, V3 deepens reasoning. Then it tells you which to use.

1. What this does for you

The AI takes a prompt you're already using and improves it three ways: adding constraints, resolving ambiguity, and forcing deeper reasoning. You see all three versions side by side and pick the strongest one. The prompts you reuse most often (client updates, status memos, meeting summaries, proposal sections) improve dramatically after a single pass through this loop.

2. What to type into your AI

Copy the box below. Replace the words in [BRACKETS] with your own info. Then paste it into Claude or ChatGPT.

You are a prompt optimizer. I am going to give you a prompt I want to improve. Optimize it across three explicit versions, then tell me which one to use.

THE PROMPT I WANT TO IMPROVE:
"""
[PASTE THE PROMPT YOU WANT OPTIMIZED: could be one the AI designed for you in the Reverse Prompting exercise, or one you wrote yourself.]
"""

WHAT THE PROMPT IS FOR:
[BRIEF DESCRIPTION OF THE GOAL, e.g. "Drafting the Friday status update I send every client at [ORGANIZATION NAME]." OR "Drafting the announcement email that goes to the whole team."]

OPTIMIZE IT IN THREE VERSIONS:

VERSION 1 (ADD CONSTRAINTS):
Add specific constraints the original prompt is missing: length, audience, tone, must-include elements, format. Show me the V1 prompt with the new constraints called out.

VERSION 2 (RESOLVE AMBIGUITIES):
Read V1 like a fresh reader who has no context. Find every word, phrase, or instruction that could be misinterpreted. Rewrite it to remove all ambiguity. Show me V2 with the fixes called out.

VERSION 3 (DEEPEN REASONING):
Add explicit instructions that force the AI to show its work, self-check its assumptions, and flag uncertainty. Show me V3.

Then tell me: which version should I use, and why?

Try it now in Claude →

See a full worked example ▾

Optimizing a generic client-status-update prompt

You are a prompt optimizer. Improve the prompt below across three explicit versions, then tell me which to use.

THE PROMPT I WANT TO IMPROVE:
"""
Write a status update on this project for the client.
"""

VERSION 1 (ADD CONSTRAINTS): Write a 150-word status update for a busy client who will skim it. Must include one finished deliverable (named), one named risk with where it happens, and one thing we need with a date. Avoid "going well," "on track," "circling back."

VERSION 2 (RESOLVE AMBIGUITIES): "One finished deliverable" means its state (e.g. "the intake form is live and took 12 submissions"), not "great progress." If there was no real progress, say so plainly.

VERSION 3 (DEEPEN REASONING): Before drafting, list three things you'll do to avoid canned status-speak. After drafting, tag any vague sentence with [vague].

WHICH VERSION TO USE: V3, every time.

Sharpening a reused report-card comment prompt across V1/V2/V3

You are a prompt optimizer. I am going to give you a prompt I want to improve. Optimize it across three explicit versions, then tell me which one to use.

THE PROMPT I WANT TO IMPROVE:
"""
Write a report-card comment for this student based on my notes.
"""

WHAT THE PROMPT IS FOR:
End-of-quarter comments for 28 fourth graders. I reuse this every nine weeks and I want it to sound like me, not a form letter.

VERSION 1 (ADD CONSTRAINTS): 3 to 4 sentences, addressed to the family, warm but honest, must name one strength and one next step, no grade-shaming.
VERSION 2 (RESOLVE AMBIGUITIES): define what "my notes" includes (reading level, behavior, one work sample) so a fresh reader does not invent facts.
VERSION 3 (DEEPEN REASONING): make the AI quote the specific note it used for each claim, and flag any comment it could not support.

Then tell me which version to use, and why.

Tuning a recurring council briefing-memo prompt across V1/V2/V3

You are a prompt optimizer. I am going to give you a prompt I want to improve. Optimize it across three explicit versions, then tell me which one to use.

THE PROMPT I WANT TO IMPROVE:
"""
Draft a briefing memo for the city council on this agenda item.
"""

WHAT THE PROMPT IS FOR:
The recurring memo staff send to council before each meeting. I reuse it for every item and I want it neutral and easy to skim.

VERSION 1 (ADD CONSTRAINTS): one page, plain language for a general public audience, neutral and non-partisan tone, must include background, options, fiscal impact, and staff recommendation.
VERSION 2 (RESOLVE AMBIGUITIES): specify that "fiscal impact" needs a dollar figure and a funding source, so no reader guesses where the money comes from.
VERSION 3 (DEEPEN REASONING): make the AI show the tradeoff behind each option and flag any number it is not certain about.

Then tell me which version to use, and why.

Improving a weekly owner progress-report prompt across V1/V2/V3

You are a prompt optimizer. I am going to give you a prompt I want to improve. Optimize it across three explicit versions, then tell me which one to use.

THE PROMPT I WANT TO IMPROVE:
"""
Write this week's progress report for the Owner from my field notes.
"""

WHAT THE PROMPT IS FOR:
The weekly owner update on a 40-unit apartment job. I send it every Friday and I want it clear enough that the Owner does not call with questions.

VERSION 1 (ADD CONSTRAINTS): half a page, addressed to the Owner, plain and factual tone, must include work completed, work planned next week, any schedule slip, and open RFIs.
VERSION 2 (RESOLVE AMBIGUITIES): define what a "schedule slip" is (days behind against the baseline) so the Owner and GC read it the same way.
VERSION 3 (DEEPEN REASONING): make the AI tie each delay to a cause (weather, a Sub, or material) and flag anything my notes do not confirm.

Then tell me which version to use, and why.

Refining a reusable customer-review reply prompt across V1/V2/V3

You are a prompt optimizer. I am going to give you a prompt I want to improve. Optimize it across three explicit versions, then tell me which one to use.

THE PROMPT I WANT TO IMPROVE:
"""
Write a reply to this customer review.
"""

WHAT THE PROMPT IS FOR:
Replies to Google reviews for my coffee shop, both the 5-star ones and the angry ones. I answer a few every week and I want them to sound like a real person.

VERSION 1 (ADD CONSTRAINTS): 2 to 3 sentences, warm and direct, sign off as the owner by first name, thank them, and for a complaint offer one real next step. No coupons unless I say so.
VERSION 2 (RESOLVE AMBIGUITIES): specify how to handle a review with no detail, so the AI does not apologize for a problem the customer never named.
VERSION 3 (DEEPEN REASONING): make the AI point to the exact line it is responding to and flag any review that needs me to reply personally.

Then tell me which version to use, and why.

Strengthening a monthly donor-update prompt across V1/V2/V3

You are a prompt optimizer. I am going to give you a prompt I want to improve. Optimize it across three explicit versions, then tell me which one to use.

THE PROMPT I WANT TO IMPROVE:
"""
Write our monthly donor update from this month's program numbers.
"""

WHAT THE PROMPT IS FOR:
The email update we send about 400 donors each month for our youth mentoring program. I want it honest, including the months we fall short.

VERSION 1 (ADD CONSTRAINTS): about 250 words, warm and plain, must include one number, one real story, and one specific ask. No hype, no "changing the world."
VERSION 2 (RESOLVE AMBIGUITIES): define what counts as "this month's numbers" (youth served, sessions held, waitlist) so no one inflates the count.
VERSION 3 (DEEPEN REASONING): make the AI name where we missed our goal and flag any claim the numbers do not back up.

Then tell me which version to use, and why.

Optimizing a monthly client close-summary prompt across V1/V2/V3

You are a prompt optimizer. I am going to give you a prompt I want to improve. Optimize it across three explicit versions, then tell me which one to use.

THE PROMPT I WANT TO IMPROVE:
"""
Write the month-end close summary for this client from the numbers.
"""

WHAT THE PROMPT IS FOR:
The monthly note I send bookkeeping clients after close. I reuse it for a dozen small businesses and I want plain English they actually read.

VERSION 1 (ADD CONSTRAINTS): under 200 words, addressed to a non-accountant owner, plain and calm tone, must cover revenue vs last month, cash position, and one thing to watch. Label it general guidance, not filed tax advice.
VERSION 2 (RESOLVE AMBIGUITIES): specify whether figures are cash or accrual, so the owner does not misread the cash line.
VERSION 3 (DEEPEN REASONING): make the AI show the month-over-month math and flag any account that looks off so I check it before sending.

Then tell me which version to use, and why.

Polishing a commission-inquiry reply prompt across V1/V2/V3

You are a prompt optimizer. I am going to give you a prompt I want to improve. Optimize it across three explicit versions, then tell me which one to use.

THE PROMPT I WANT TO IMPROVE:
"""
Write a reply to someone asking about a commission.
"""

WHAT THE PROMPT IS FOR:
My first reply when someone emails about a custom piece. I get a few a week and I want it warm without underpricing my work.

VERSION 1 (ADD CONSTRAINTS): 4 to 6 sentences, warm and personal tone, no art jargon, must thank them, ask 2 questions (size and timeline), and give a starting price range.
VERSION 2 (RESOLVE AMBIGUITIES): specify what to do when they do not name a budget, so I do not quote a full price before I know the scope.
VERSION 3 (DEEPEN REASONING): make the AI explain why it set the price range and flag anything it assumed about the project.

Then tell me which version to use, and why.

3. What to do with the answer

Pick a prompt you already reuse often: a client update, a status memo, a meeting summary, a proposal section. Anything generic.
Paste the prompt above into Claude, Gemini, or NotebookLM. Drop your existing prompt in the [BRACKETS], plus one sentence on what it's for.
Run it. The AI returns V1 (constraints), V2 (ambiguity), V3 (reasoning), plus a recommendation.
Compare the three versions side by side. V1 is usually the biggest jump; V2 catches the things only a stranger would notice; V3 is what you keep.
Save V3 to your prompt library. Replace the old version everywhere: shared docs, team folders, templates, onboarding instructions for new staff.
Log what V3 changed (e.g. "added [vague] tags" / "forced a named deliverable instead of vague reassurance"). Those reasoning patterns travel. Apply them to the next prompt you optimize.

Why three explicit versions

Ad-hoc tweaking misses things in predictable ways: you add constraints but skip ambiguity, or you tighten the wording but never force the AI to show its work. The structured V1 → V2 → V3 sequence makes each pass do one job. Constraints first, because most prompts are vague before they're wrong. Ambiguity second, because you stop seeing it in your own prompt after the second read. Reasoning depth last, because that's the layer that catches the sentences the AI was confidently going to make up.

Section Recap

Core Principle

Optimize prompts in three explicit passes (constraints, ambiguity, reasoning depth) instead of tweaking them ad-hoc.

Best-Fit Example

A project lead uses the same status-update prompt on every client project. She runs it through this loop once: V1 caps the length and bans "on track," V2 forces a named deliverable, V3 flags canned reassurance. She replaces the prompt in the team folder, and every update for the rest of the year is sharper for one ten-minute pass.

A fourth-grade teacher reuses the same tired comment prompt every nine weeks and the results read like a form letter. She pastes it into the optimizer with a note about what a good comment must include. V3 comes back forcing the AI to quote the exact note behind each claim and flag anything it cannot support, which stops it from inventing progress a student never made. She saves V3, runs all 28 comments through it, then edits each one for tone before it goes home.

A city clerk drafts the same briefing memo before every meeting, and half of them get sent back for missing a funding source. He runs his memo prompt through the optimizer and V2 nails down that fiscal impact needs a dollar figure and a source. He adopts V2 as the standing template, and the next round of memos reaches council without the usual follow-up questions. He keeps V3 on hand for budget items where the tradeoffs actually matter.

A superintendent writes the Friday owner update from scribbled field notes, and the Owner keeps calling to ask what "behind" means. He drops his report prompt into the optimizer, and V2 defines a schedule slip as days behind the baseline so everyone reads it the same way. He uses V2 for routine weeks and switches to V3 when a delay needs a clear cause tied to weather, a Sub, or materials. The owner's calls stopped, and Fridays got shorter.

A coffee shop owner answers Google reviews between rushes, and her replies started sounding copy-pasted. She feeds her review-response prompt to the optimizer and picks V1, which keeps replies to a few warm sentences signed with her first name and offers one real next step on complaints. For the vague one-line reviews she keeps V3, which points to the exact words and flags the ones she should handle herself. Now a week of replies takes ten minutes instead of an hour.

A development lead sends a monthly email to 400 supporters and worries it reads like hype. She runs her update prompt through the optimizer, and V3 makes the AI name where the program missed its goal and flag any claim the numbers do not support. She adopts V3 so the lean months stay honest, then adds the real story and the ask in her own words. Open rates held and two lapsed donors wrote back that month.

A bookkeeper sends month-end summaries to a dozen small-business clients, and some read the cash figure as if it were profit. He optimizes his close-summary prompt, and V2 forces the note to state whether numbers are cash or accrual, which ends the confused replies. He keeps the "general guidance, not filed tax advice" label from V1 on every send, and switches to V3 at year end when clients want the math shown. He reviews each flagged account himself before anything goes out.

A ceramicist answers commission emails a few times a week and used to either underprice the work or bury people in a wall of text. She runs her inquiry-reply prompt through the optimizer, and V1 gives her a warm four-line reply that asks about size and timeline and states a starting range. She keeps V2 for the emails with no budget so she does not quote before she knows the scope. She saved V1 as her go-to, and her reply time dropped to a couple of minutes.

Key Takeaways

Constraints first: most prompts are vague before they're wrong.
Read V1 like a stranger to catch ambiguity you can no longer see.
V3 should force the AI to show its work and flag uncertainty inline.
Save V3. Every prompt you reuse often deserves one pass through this loop.
The reasoning patterns from V3 travel to the next prompt you optimize.

← All Tools