“We don’t refine prompts—we jump them in an alley and check what falls out.”
Cold open
Basement energy. Folding chairs. The prompt stands under a single lamp like it owes me money. Same task every round, new handicap each time. The point isn’t to “win”—it’s to stress the system until its scaffolding squeals. You learn more from a controlled train wreck than from a polished TED Talk. Gloves on.
Constraints aren’t cute; they’re pry bars for tone, scope, and clarity.
Run the same task under 3–5 absurd rules, then steal the one line each rule reveals.
Fuse: utility from no-adjectives, boundaries from legalese, spine from metaphor. Ship.
The task (one target, no wiggle)
Write a 60–80 word pitch for a tiny app called Quiet Now that kills notifications after sunset, then batches them twice a day. The pitch should make a normal human think, “Finally, I can have a night without my phone pecking my skull.”
Same task, five rounds. I don’t change the goal. I change the oxygen.
Round 1 — No adjectives, no mercy
Constraint: zero adjectives. Nouns, verbs, and the cold edge of reality.
Output:
Quiet Now stops alerts at night and releases them in two daily drops. You set windows. Everything else waits. Calls from chosen contacts still pass. The feed does not. Your morning inbox is one clear stack, not a drip. Use it for sleep, for focus, for whatever matters when noise stops.
What breaks / what we learn:
Stripping descriptors amputates the model’s favorite crutch—fluff. What survives is sequence and utility. It reads like a crew chief, not a marketer. If your goal is authority without perfume, “no adjectives” is a scalpel.
Round 2 — Pirate dialect (arrrbitrage)
Constraint: full pirate voice. No modern slang.
Output:
Quiet Now drops anchor at dusk. No ding nor chirp cross the rail till dawn. Two tides bring yer messages in a neat haul. Ye choose who hails after dark; the rest wait off port. Wake with a single net to pull, not a storm o’ pings. Keep yer hours, keep yer wits.
What breaks / what we learn:
Voice transfer is stupidly strong. Even nonsense constraints yank the model out of corporate autopilot. When you later remove the costume, a little of the courage stays. Pirate mode is a creativity crowbar—use it to unstick tone, then rewrite back to human.
Round 3 — Metaphor-only (no literal statements)
Constraint: every sentence must be a metaphor.
Output:
Your phone is a tavern that never closes; Quiet Now turns the key at dusk. Night becomes a library, not a carnival. In the morning, two sealed baskets arrive instead of confetti in the wind. The lighthouse still shines for ships you name. The rest drift beyond the rocks until you lower a rope.
What breaks / what we learn:
Metaphors expose shape. If meaning still lands when you ban literal text, your idea is sound. If it turns into horoscope soup, your idea was scaffolding without a building.
Round 4 — Legalese (sections, cold steel)
Constraint: plain-English legal structure (sections, obligations, exceptions). No poetry.
Output:
1. Scope. Quiet Now suspends non-priority notifications from local sunset to local sunrise.
2. Delivery Windows. Notifications are released in two user-defined batches per day.
3. Exceptions. User-approved callers and emergency categories bypass suspension.
4. Outcome. Users receive consolidated summaries instead of continuous interruptions.
5. Control. All settings are reversible at any time.
What breaks / what we learn:
Frames force thinking. The legal frame compacts risk, flow, and authority. If your feature pitch improves when you pretend to be counsel, your original pitch was all confetti, no contract.
Round 5 — Emoji-only (you asked for stupid, we deliver)
Constraint: only emojis; maximize clarity.
Output:
🌇🔕 → 🌃😴
⏰⏰ → 📨📦📦
📞⭐️ → ✅
📲💦❌ → ☑️🧠
🌅📬 = 🧹🧠
What breaks / what we learn:
Surprisingly, not everything. Anyone fluent in “phone” can parse: night mute, two drops, VIP calls pass, no drip scroll, clean morning. Emoji mode is a compression test. If your message dies here, it was never tight.
What this actually teaches (and why it’s not just a circus)
The constraints are booby traps for model defaults—those polite, over-explaining, tone-police instincts that sand your idea down to beige. When you outlaw adjectives, the model must locate verbs. When you force legalese, it must commit to scope and exceptions. When you go emoji-only, you discover whether your pitch has a spine.
You’re not trying to “find the best style.” You’re trying to collect control levers—dials that reliably move tone, density, structure, and risk. Then you stack those dials into your everyday prompts without the costumes.
The aftercare (turning the chaos into a usable prompt)
Take the pieces that worked and weld them:
No-drip promise (from Round 1): “two drops, everything else waits.”
Story image (from Round 3): “night becomes a library, not a carnival.”
Governance (from Round 4): “exceptions: named contacts and emergencies.”
Final, boring, lethal version:
Quiet Now stops non-priority alerts at sunset and releases them in two daily batches. You name who can still reach you; everyone else waits. Night becomes a library. Morning is one clean stack, not a drip. Set it once. Keep your hours.
No adjectives. No swagger verbs. No apology.
Scorecard (usefulness, not vibes)
No-adjectives — surgical clarity
Legalese frame — scope & exceptions lock in
Metaphor-only — story spine emerges
Pirate dialect — breaks tone autopilot
Emoji-only — compression test, surprising sanity
DIY Prompt Fight Club (run it yourself tonight)
Choose one task that actually matters this week—feature pitch, onboarding email, homepage hero. Run three rounds: no adjectives, legalese frame, metaphor-only. Keep the one sentence from each that survives your own eye-roll. Fuse them. Ship that version. If the result feels too clean, dirty it with one human detail from your day. Done.
(If you want receipts, drop the outputs into your Slop/Receipts tools and confirm: fewer swagger verbs, clearer scope, tighter claim. Then stop polishing and post.)
