exhibit a · the prompt that wastes 14 minutes
“hey can you reply respond to this angry customer, keep it nice but firm and don’t give too much please thanks”
no role · no audience · no rules→ ai-flavored mush
Write what you mean.Get a prompt that’s already proven to work.
Describe it like you’d tell a coworker. Meerkat drafts it across ChatGPT, Claude & Gemini, then runs it on real models to prove it holds — before you paste it.
What kind of prompt do you need?
in their words
“I can usually give a fairly rough explanation of what I'm trying to do and it turns that into something much more structured and useful. I gave Meerkat and Codex the same task and actually used both prompts, and the Meerkat one gave me the better result.”
§ 01the diagnosisfig. 01 — bad prompt vs good
vague. lazy. vibes.are out. Specifics ship.
Models don't need motivation, they need a job description. You tell Meerkat what you're trying to make; it asks the things a senior editor would ask before touching the keyboard, then writes the prompt for you in the format the model actually prefers.
exhibit b · the same ask, with its bones in place
- ROLE
- You are a senior support lead, calm under fire.
- TASK
- Draft one reply to the message pasted below.
- CONTEXT
- A paying customer, threatening to cancel.
Goal: keep them — without overpromising. - RULES
- Name their exact problem first.
Promise only what policy allows.
Escalate anything legal or safety to a human.
role · task · context · rules→ a reply that keeps them
§ 02the answer, live
Watch what you mean become a real prompt.
●build with meerkat
15.2s · livefix this for me:
"hey can you summarize this for me but make it good and professional and not too long please thanks"
reply to meerkat…claude
summarize-exec.txt
> waiting for your first answer
§ 03the panelfig. 03 — the panel, one prompt
Ask one model, get one read on it.
A single model only ever shows you its own take — you tweak against the one model that wrote it and hope it holds. Meerkat reads your brief across the frontier labs, keeps the strongest move from each, then runs the result on real models to see what actually holds. So you ship from proof, not a hunch.
Gemini
Reframes the ask around who actually reads it.
without it
the wrong audience, perfectly executed
Anthropic
Claude
Argues with the brief before it answers — catches the vague success metric.
without it
a confident answer to the wrong question
OpenAI
ChatGPT
Locks the output to a strict, copy-ready format.
without it
a good draft you still have to reshape by hand
one prompt
Meerkat’s merge
best of eachEvery move, kept. Nothing missed.
The right audience from Gemini, the sharpened brief from Opus, and the strict format from GPT — composed into a single prompt you copy once and ship. One model could only ever have given you a third of that.
paper vs. prooffig. 03b — the other kind of tool
most “prompt enhancers”
Graded on paper.
They read your prompt and hand it back tidier — a role, some bullets, a format, a tidy confidence score like 91/100.
It never finds out whether the prompt actually works. Neither do you.
meerkat
the difference that matters
Proven on models.
Meerkat runs the prompt on real frontier models, judges what comes back, and repairs what breaks — before you ever paste it.
You ship knowing it holds — not hoping it reads well.
We don’t just build the prompt. We prove it works.
our internal blind-judge test
Our own test, not a third-party benchmark: each prompt runs on a fresh input, then a separate model — told nothing about which side it’s judging — scores the result from 0 to 5.
Meerkat’s merge4.4/5
A typical prompt1.3/5
05
3.4×more complete
Completeness measures how fully a prompt holds its instructions across inputs — not just the easy one. Same models on both sides; only the prompt changed.
- the ask
- Pull clean, structured fields out of a messy job posting.
- the curveball
- The posting names a staffing agency, but the real employer is only described — never written out.
held
Meerkat’s merge: Resolved the employer by an explicit rule and returned valid, copy-ready JSON. Judged airtight.
slipped
A typical prompt: Left the hardest field empty rather than resolving it — complete-looking, but wrong on the input that mattered.
§ 04the frameworkfig. 04 — anatomy of a prompt
We get the bones right.You just talk.
Every strong prompt rests on four bones — role, task, context, rules. A template hands you four empty boxes; you still fill them. Meerkat covers whichever ones your task needs and shapes them to fit — tight schema for extraction, room to breathe for a brief — all pulled from a two-minute chat, not a framework you have to learn.
§ 05in the wild
Call your prompts from inside Claude.
Connect Meerkat once and it shows up in Claude like a teammate you can tag. You ask in plain English. Meerkat brings the brief. Claude runs with it.
- 01Save.Build prompts in Meerkat. Keep the ones you reach for.
- 02Connect.Paste your Meerkat URL into Claude’s connectors. Once.
- 03Call.Say “use Meerkat” in any chat. Claude pulls it in.
requires Claude Pro, Max, Team, or Enterprise
claude.ai · meerkat connector
have Meerkat write me a prompt that turns my raw meeting notes into a 3-bullet exec summary for my CFO. she hates fluff.
Got you. Your CFO is the audience, not you — so the prompt has to ban hedging, rank by dollar impact, and cap each bullet at one line.
draft promptcfo · 3 bullets
You are summarizing meeting notes for a CFO who has 90 seconds.
Rules:
- Exactly 3 bullets. Each one ≤ 18 words.
- Lead each bullet with the dollar impact, the deadline, or the decision owner — whichever is loudest.
- No hedging words: "potentially," "might," "could," "exploring."
- If a number isn't in the notes, say "TBD" — never invent one.
Notes:
{paste meeting notes here}Want me to harden it against jargon, or add a fallback bullet for when the notes are thin?
ban jargonthin-notes fallbackadd a one-line TL;DR
works with claude
§ 06the wall12 ways prompts fail
Meerkat doesn’t flatter. It roasts.
A gallery of the prompt mistakes everyone makes — one clean specimen of each, written by us and run through the real Meerkat. Every card names the failure it caught and opens the full roast and the rebuilt prompt.
Roast my promptfree · no signup · shareable
Verdict
“'Make it pop' is three words of pure wish — you've handed the model a mood board, not a brief.”
Vague verbread the roast
Verdict
“'Good and memorable' is a wish, not a brief — you've described every name ever made and ruled out exactly none.”
No format contractread the roast
Indictment
“'World-class expert with 20 years of experience' is a LinkedIn headline, not a brief — and 'help me with my go-to-market strategy' is the entire contents of a…”
Role as a crutchread the roast
Verdict
“'So people understand it' is the entire spec — which people, understanding what, to do what with it?”
Under-promptingread the roast
Verdict
“'Pour your heart into it' is instructions for a barista, not a language model — the model has no heart, no coffee, and no idea what makes yours different.”
Emotional directionread the roast
Indictment
“'Detailed but short, thorough but brief, professional but casual' — you've written four constraints that cancel each other out and called it a brief.”
Contradictionread the roast
Indictment
“Two hundred words of biography, zero words of an actual question — 'what to do next' is not a question, it's a shrug.”
Context dumpread the roast
Indictment
“'Gets replies' is the whole brief — you've described the desired outcome and left the model to invent the product, the sender, the recipient, the ask, and the…”
Under-promptingread the roast
Verdict
“'95% certain' sounds rigorous until you remember the model has no certainty meter — it'll ask one question or twelve, and neither feels wrong to it.”
Failure moderead the roast
Verdict
“Five 'don'ts' and zero 'dos' — you've handed the model a list of landmines with no map to the destination.”
Emotional directionread the roast
Verdict
“You buried a bakery brief under so many pleases and maybes that the only thing the model knows for certain is that you're polite.”
Polite fillerread the roast
Verdict
“You buried a perfectly clear question under a paragraph of grief — the model doesn't need your grandmother's Sunday routine to recommend a beginner loaf.”
Context dumpread the roast
§ 07the pricefig. 07 — pricing card
Free to try. Cheap to keep.
A free template doesn’t fix a weak prompt — it just moves the bill to the reroll spiral, where every “try again” burns tokens and minutes. The cheapest prompt is the one you only run once. The full engine is free here, 15 runs a month. $9 a month makes Meerkat yours: 150 runs a month, it learns the gaps you keep leaving in your briefs and covers them on its own, and plugs straight into Claude so it works where you do.
Free
spec · 01The full magic. Real models, every time.
$0forever
- 15 runs a month
- 10 saved specimens · forever
- One project shelf
- Manual filing
- Works in Claude and ChatGPT
Recommended
Pro
spec · 02Make Meerkat yours. It learns your habits.
$9/ month
billed monthly·cancel anytime
- 150 runs a month · add 100 for $5
- Learns how you write, covers the gaps automatically
- Unlimited shelf + projects
- Version history
- Export · JSON, Markdown
cancel anytime · no contract · no hagglesee full breakdown
§ 08your turn
Just talk.Ship a prompt.
Two minutes of talking, and you walk away with a prompt for ChatGPT, Claude, or Cursor — no signup, no homework. Every fix shows the why, so you get sharper each time.
fig. 08 — exit, stage left
