Tool · New

Adversarial Prompt Tester

Make AI red-team its own answer. Find the 5 worst questions, 3 missing facts, 2 emotional landmines, and the 1 way it could blow up, before you hit send.

1. What this does for you

Ask AI to attack its own answer like a skeptical critic. It finds the holes, missing evidence, vague language, and ways the writing could blow up, so you fix them before you send the email, the client update, the status memo, or the public post.

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 going to attack the answer below, your own answer, like a skeptical, well-prepared critic. Don't be polite. Don't soften. Find every weak spot before I send this to a real reader.

THE ANSWER TO ATTACK:
"""
[PASTE THE AI'S ANSWER YOU ARE ABOUT TO USE: the email, memo, or write-up you were about to send, a client update, a board update, a post-incident statement, anything]
"""

WHAT I NEED FROM YOU (RED-TEAM MODE):

1. THE 5 WORST QUESTIONS A SKEPTIC WOULD ASK
List the five hardest questions a smart, skeptical reader could ask after reading this. That means the questions a frustrated client, a board member, a reporter, your manager, or the person most affected absolutely would ask.

2. THE 3 FACTS THAT ARE MISSING OR UNVERIFIED
What claims in here can't be checked? What numbers don't have a source? What "we are committed to excellence" type phrases would a journalist or an auditor circle in red?

3. THE 2 EMOTIONAL LANDMINES
What word, framing, or assumption could land badly with a reader who has had a different experience than the writer? (e.g. compliance language aimed at your own team, framing that blames the people affected, talking around money or staffing reality.)

4. THE ONE WAY THIS COULD BLOW UP
If this got forwarded to the WRONG person (an angry client, a competitor, a reporter, the board, your boss's boss), what's the one line that comes back to bite us?

5. THE FIX LIST
For each problem above, give me the exact change to make. Be specific. Not "be more concrete." Instead: replace "the project is going great" with "we completed 4 of the 6 milestones and are 2 days behind on the fifth." Replace "we are committed to quality" with "every deliverable gets a second review before it ships, starting Sept 8."

Be ruthless. I'd rather hear it from you now than from a critic later.

Try it now in Claude →

See a full worked example ▾

Red-teaming a generic status update full of buzzwords

You are going to attack the answer below, your own answer, like a skeptical, well-prepared critic. Don't be polite.

THE ANSWER TO ATTACK:
"""
Hi all, I am excited to share that the project is going great this quarter! Through a collaborative, results-driven approach, we are meeting every challenge and delivering real value.
"""

WHAT I NEED (RED-TEAM MODE):
1. THE 5 WORST QUESTIONS A SKEPTIC WOULD ASK
2. THE 3 FACTS THAT ARE MISSING OR UNVERIFIED
3. THE 2 EMOTIONAL LANDMINES
4. THE ONE WAY THIS COULD BLOW UP
5. THE FIX LIST (exact changes)

Be ruthless. I'd rather hear it from you now than from a critic later.

Attacking a vague whole-class family update before it goes home

You are going to attack the answer below, your own answer, like a skeptical, well-prepared critic. Don't be polite.

THE ANSWER TO ATTACK:
"""
Hi families, we are leveraging a holistic, student-centered approach to support every learner this quarter.
There was a minor incident at recess, but we are proactively wrapping supports around all involved with a growth mindset.
Rest assured your child is thriving and safe.
We will circle back with next steps soon.
"""

WHAT I NEED (RED-TEAM MODE):
1. THE 5 WORST QUESTIONS A SKEPTIC WOULD ASK
2. THE 3 FACTS THAT ARE MISSING OR UNVERIFIED
3. THE 2 EMOTIONAL LANDMINES
4. THE ONE WAY THIS COULD BLOW UP
5. THE FIX LIST (exact changes)

Be ruthless.

Attacking a hollow resident notice about a water rate increase

You are going to attack the answer below, your own answer, like a skeptical, well-prepared critic. Don't be polite.

THE ANSWER TO ATTACK:
"""
Dear Residents, as part of our ongoing commitment to service excellence and fiscal responsibility, the City is implementing a strategic rate adjustment next month.
This modest change ensures sustainable infrastructure for future generations and reflects our resident-first values.
We appreciate your partnership and patience.
Together we are building a brighter tomorrow.
"""

WHAT I NEED (RED-TEAM MODE):
1. THE 5 WORST QUESTIONS A SKEPTIC WOULD ASK
2. THE 3 FACTS THAT ARE MISSING OR UNVERIFIED
3. THE 2 EMOTIONAL LANDMINES
4. THE ONE WAY THIS COULD BLOW UP
5. THE FIX LIST (exact changes)

Be ruthless.

Attacking an Owner update that hides a framing delay

You are going to attack the answer below, your own answer, like a skeptical, well-prepared critic. Don't be polite.

THE ANSWER TO ATTACK:
"""
Owner team, great news, the project continues to track well and everyone is fully aligned and executing at a high level.
We hit a few minor coordination items with the framing sub, but nothing that changes the big picture.
We remain committed to delivering a best-in-class result and substantial completion is still our target.
We will keep you posted.
"""

WHAT I NEED (RED-TEAM MODE):
1. THE 5 WORST QUESTIONS A SKEPTIC WOULD ASK
2. THE 3 FACTS THAT ARE MISSING OR UNVERIFIED
3. THE 2 EMOTIONAL LANDMINES
4. THE ONE WAY THIS COULD BLOW UP
5. THE FIX LIST (exact changes)

Be ruthless.

Attacking a price-increase email full of corporate filler

You are going to attack the answer below, your own answer, like a skeptical, well-prepared critic. Don't be polite.

THE ANSWER TO ATTACK:
"""
Hey everyone, due to unprecedented market conditions and our unwavering commitment to quality, we are adjusting our prices effective immediately.
This helps us keep delivering the world-class experience you deserve.
We know you will understand, and we truly value your loyalty during this exciting new chapter.
Thanks for being part of the family.
"""

WHAT I NEED (RED-TEAM MODE):
1. THE 5 WORST QUESTIONS A SKEPTIC WOULD ASK
2. THE 3 FACTS THAT ARE MISSING OR UNVERIFIED
3. THE 2 EMOTIONAL LANDMINES
4. THE ONE WAY THIS COULD BLOW UP
5. THE FIX LIST (exact changes)

Be ruthless.

Attacking a funder report that spins a missed enrollment target

You are going to attack the answer below, your own answer, like a skeptical, well-prepared critic. Don't be polite.

THE ANSWER TO ATTACK:
"""
Dear Friends, thanks to your generous partnership we made incredible strides this year and touched countless lives.
While our enrollment came in slightly under projection, we pivoted strategically and doubled down on impact over quantity.
We are more energized than ever and deeply grateful for your continued belief in our mission.
Together we are changing everything.
"""

WHAT I NEED (RED-TEAM MODE):
1. THE 5 WORST QUESTIONS A SKEPTIC WOULD ASK
2. THE 3 FACTS THAT ARE MISSING OR UNVERIFIED
3. THE 2 EMOTIONAL LANDMINES
4. THE ONE WAY THIS COULD BLOW UP
5. THE FIX LIST (exact changes)

Be ruthless.

Attacking a client note that overpromises on an IRS notice

You are going to attack the answer below, your own answer, like a skeptical, well-prepared critic. Don't be polite.

THE ANSWER TO ATTACK:
"""
Hi Dana, happy to help. Don't worry about the IRS notice, these are totally routine and we will make it go away with no issues.
From a quick look you are definitely fine and should not owe anything, and you could probably deduct most of that anyway.
We will get you the biggest refund possible.
Consider it handled.
"""

WHAT I NEED (RED-TEAM MODE):
1. THE 5 WORST QUESTIONS A SKEPTIC WOULD ASK
2. THE 3 FACTS THAT ARE MISSING OR UNVERIFIED
3. THE 2 EMOTIONAL LANDMINES
4. THE ONE WAY THIS COULD BLOW UP
5. THE FIX LIST (exact changes)

Be ruthless.

Attacking a commission proposal buried in art jargon

You are going to attack the answer below, your own answer, like a skeptical, well-prepared critic. Don't be polite.

THE ANSWER TO ATTACK:
"""
My practice interrogates the liminal spaces between memory and materiality through an embodied, post-disciplinary praxis.
This commission will manifest a site-responsive intervention that destabilizes the viewer's ontological assumptions.
Deliverables and timeline are TBD.
Investment reflects the depth of the conceptual labor involved.
"""

WHAT I NEED (RED-TEAM MODE):
1. THE 5 WORST QUESTIONS A SKEPTIC WOULD ASK
2. THE 3 FACTS THAT ARE MISSING OR UNVERIFIED
3. THE 2 EMOTIONAL LANDMINES
4. THE ONE WAY THIS COULD BLOW UP
5. THE FIX LIST (exact changes)

Be ruthless.

3. What to do with the answer

Take the email, memo, or write-up you were about to send (a client update, a project status memo, a board memo, a post-incident statement) and paste it into the [BRACKETS].
Run the prompt. Read the 5 worst questions out loud. If you can't answer one of them with a specific number, name, or date, you've found the hole.
Look at the missing facts list. Add the actual numbers, dates, and details from your real records. (No specifics? That's the issue. Go get them.)
Take the emotional landmines seriously. Phrases like "going great," "best in class," "tireless dedication," "holistic approach," "data-driven" are the ones critics circle. Replace them with specifics.
Take the "one way this could blow up" warning seriously. Read it as if an angry client, a board member, or a reporter is on the other end. If the warning is real, rewrite that line.
Apply every fix in the Fix List. Then run this tool one more time on the new version. Stop only when AI struggles to find a real problem.

Why we built this

"AI sounds confident" is the most dangerous thing about it. The same model that wrote your draft will, if you ask, attack it harder than any critic. Adversarial Prompting is asking AI to do its own peer review. Use it on every message about a hard topic, every team announcement, every performance write-up, every public statement, especially the ones that feel "done."

Pair this with the Recursive Loop

Recursive Loop polishes. Adversarial Tester breaks. Run Recursive first to get a clean draft. Run Adversarial second to find what the polish hid. The combination is how you go from "good enough to send" to "actually defensible."

Section Recap

Core Principle

AI sounds confident. Adversarial Prompting flips it into peer-review mode and surfaces every weak claim, missing fact, and emotional landmine the polish was hiding.

Best-Fit Example

A project lead is about to send a client an update on a slipping timeline. She runs the draft through Adversarial Prompting. The AI flags "the project is going great" as contradicting the actual dates and warns the "unforeseen circumstances" line could blow up if forwarded to leadership. She replaces the buzzwords with real dates and a recovery plan, and sends something the client actually trusts.

A teacher is about to send a whole-class family update after a recess scuffle. Before hitting send, she runs it through the tester, which throws the questions parents will actually fire back: which child, what happened, was mine hurt, what are you doing about it. It flags that "thriving" and "wrapping supports" say nothing and could read as a cover-up. She rewrites it with the specific facts she can share and one clear next step. The AI stays the intern while she keeps the judgment call on what is appropriate to disclose.

A city administrator drafts a notice about a water rate increase and wants it to land without a packed council meeting. The tester surfaces the five questions residents will ask (how much, why now, where does the money go, who decided, can I appeal) and flags that "strategic rate adjustment" hides the actual number. It warns the buzzwords read as spin and could erode trust. He revises it with the exact dollar figure, the reason, and the public-comment date, keeping it neutral and non-partisan while he owns the final wording.

A superintendent is sending the Owner a monthly update that quietly buries a two-week framing delay. He red-teams it first, and the tool asks the questions the Owner will ask (what is the new completion date, does this affect the loan draw, whose fault is it, what is the recovery plan). It flags that "tracking well" contradicts a known slip and could blow up trust when the schedule finally shows it. He rewrites it with the honest date, the cause, and the recovery plan, and the AI drafts while he makes the call on what to promise.

An owner-operator is about to email customers a 12 percent price increase wrapped in "unprecedented market conditions." She runs it through the tester, which asks what customers will ask (how much, starting when, what am I getting for it, why should I stay) and flags that "exciting new chapter" reads as tone-deaf next to a bigger bill. It warns "you will understand" could trigger churn. She rewrites it warm and direct with the real number, the start date, and one honest reason, keeping her own voice while the AI catches the landmines.

A development lead is sending a funder its year-end report that spins a missed enrollment target as "quality over quantity." Before it goes, the tester asks the program officer's real questions (what was the target, what did you actually hit, why the gap, what changes next year) and flags that "touched countless lives" has no number behind it. It warns the spin could cost the renewal if the funder feels handled. She rewrites it honest about the shortfall with the real figures and a corrective plan, and the AI red-teams while she decides how to frame the ask.

A bookkeeper is about to reassure a client that an IRS notice is "handled" and promise the biggest refund possible. She runs it past the tester first, which flags every overpromise (definitely fine, will make it go away, deduct most of that) as a liability and asks what she has not actually verified. It warns that guaranteeing an outcome she cannot control is the one way this blows up. She rewrites it as calm next steps, marks it general guidance rather than filed advice, and keeps the judgment on what she can commit to while the AI plays the skeptic.

A working artist is about to send a gallery a commission proposal buried in "liminal spaces" and "post-disciplinary praxis." He red-teams it first, and the tool asks the plain questions a director will ask (what are you actually making, what size, when, and for how much) and flags that the jargon hides the fact there is no timeline or price. It warns the statement reads as evasive and could cost him the slot. He rewrites it warm and clear with the piece, the dimensions, the delivery date, and the fee, letting the AI stress-test it while he keeps his own voice.

Key Takeaways

Run this on every message about a hard topic, every team announcement, every performance write-up, and every public statement, especially the ones that already feel done.
If you can't answer one of the 5 worst questions with a specific number, name, or date, you've found the hole.
Replace every 'going great,' 'best in class,' 'data-driven,' or 'holistic approach' phrase with concrete numbers and named next steps.
Take the 'how this could blow up' warning seriously. Read the draft as if an angry client, a board member, or a reporter is on the other end.
Pair it with the Recursive Loop: polish first, attack second. Stop only when AI struggles to find a real problem.

← All Tools