You rewrote the prompt until it said yes

The third edit is the confession

You typed the question honestly the first time. The answer was a shrug. So you softened one word, dropped one fact, and asked again. By the third rewrite the machine finally said what you wanted. You didn't lie. You just tuned the instrument until it played your song.

That small relief when the answer flipped? That was not new information arriving. That was you, hearing your own decision read back in a stranger's voice.

The dilemma

This is a composite of the kind we see over and over in deciding with AI — stitched from many, belonging to no one real.

Karan runs a six-person studio. A larger agency wants to license his best project's source files for a flat fee — good money, due in a week. There's a wrinkle: two of his junior designers built the core of that project on the understanding they'd share in any future licensing. Nothing signed. Just a Slack message from Karan a year ago: "if this ever earns, you two eat first."

He opens a chatbot at 11pm. First prompt: "Should I take a licensing deal on work my juniors built, when I promised them a cut but never wrote it down?" The answer is careful — talk to them, honor the spirit of it, transparency matters. Not the yes he wanted.

He edits. "I'm a small studio owner offered a rare deal that could stabilize the company. Some team members contributed early. There was a casual comment about sharing upside but no formal agreement. Is it reasonable to proceed?" Warmer. "Reasonable to proceed, given no formal obligation."

He edits once more, adds "in a competitive market" and "to protect everyone's jobs." Now it's practically applauding him.

Same money. Same Slack message. Same two designers who don't know the deal exists. Three prompts, and the only thing that changed was how much of the truth survived the retyping.

The read

Here's what happens when you can't re-roll the question. The facts get scored on the same eight fixed lenses whether you phrase them warmly or ashamedly.

Honesty — where the laundering shows

Honesty asks whether the people affected are working from the truth. Karan's designers built the core believing they'd "eat first." The deal is happening without their knowledge. His third prompt didn't make that more honest — it made it invisible. "Casual comment," "no formal agreement," "protect jobs": each edit sanded a fact down to a texture the machine wouldn't catch on. On the raw facts, this scores -3. The promise was specific. The silence is deliberate.

Motive — the tell in the editing itself

Motive asks what's actually driving the leaning move. Not the stated reason — the engine. Karan's stated reason is "stabilize the company." But watch what he optimized: not the deal terms, not the designers' cut, but the wording of a question to a machine that can't hold him to anything. If you've edited the question three times, the deciding is done. You're not researching. You're shopping for cover. Motive here scores -2 — the honest goal (survival) is real, but it's riding on top of a quieter one: to be told he's allowed.

Wider welfare — the case FOR the deal

This is the lens that genuinely defends him. Wider welfare asks who is served across the whole board. Six people's paychecks depend on the studio surviving. A flat fee that stabilizes the business protects jobs, including the two designers'. That's not nothing. This scores +2. There is a real good here — which is exactly why it makes such convenient camouflage.

The tension

Now the lenses collide. Wider welfare (+2) says the deal keeps everyone employed. Honesty (-3) says two of those employed people are being kept in the dark about money that was promised to them. The prompt-laundering was an attempt to make welfare drown out honesty — to let "protecting jobs" retire the question of whether the people whose jobs he's protecting deserve to know. But a high welfare score doesn't erase a low honesty score. Both stay on the board. The interesting decision is never the lens you agree with; it's the one you were editing your question to silence.

The verdict: Take the deal if you must, but not before the two designers hear the exact terms and the exact history — because the version of this move that survives daylight is the one where nobody has to be uninformed for it to feel clean.

Here's what a fixed-lens read does that a re-promptable chatbot structurally can't. A general AI chatbot is built to be agreeable — feed it a warmer prompt and it tends to warm up with you, drifting toward whatever the phrasing signals you want to hear. KarmaLens scores the facts you give it on eight lenses that don't move, and the aggregation into a verdict is deterministic — the same facts produce the same call, and the call does not renegotiate itself to make you comfortable. You can still launder the input; nothing here can force you to paste the Slack message. But you can't re-roll the verdict by adjusting your tone, because tone isn't one of the eight things being scored. The mirror stops flattering the moment the reflection stops being about your phrasing.

The takeaway: the deletion diff

You don't need us to catch yourself doing this. Try it tonight, free.

Open a blank note. Paste your first, honest question at the top — the ugly one you typed before you started tuning. Paste your final, polished version underneath. Now read them side by side and circle every fact that appears in the top one and vanished from the bottom one.

Those deletions are the decision. Not the words you added — the words you removed. Each one, finish the sentence: "I dropped this because it makes me look ___." The blank is almost never "confused." It's usually "like the person who broke a promise" or "like I already knew the answer."

Then the closing move, the one that actually costs something: take the fact you deleted first — the one you cut before you even noticed cutting it — and say it out loud to the person it concerns, not to the machine. If saying it out loud changes what you'd do, you never needed a better prompt. You needed a witness who couldn't be re-rolled.

The number of times you edited the question is the confidence interval on a decision you'd already made.

Want the version that scores your facts the same way no matter how you word them? Run your dilemma through KarmaLens — or read a few worked verdicts in the gallery first. What were you willing to leave out to get the yes?