The promise
Generating a prompt is the easy part.
Promptzilla runs on Claude to write the first draft — then refines, tests, and optimises it into something a single pass can’t reach, because the improvements come from watching it fail. Anything can write you a prompt now. The hard part is knowing whether it holds up when the inputs get strange, and whether you’re paying for tokens you don’t need.
Generate
Describe the task and pick the shape: a single prompt, an agent with skills attached, a chain, or a full team with an orchestrator and specialists. Whatever you build comes out written for the model you’re targeting. Claude wants XML, GPT leans on Markdown, Gemini likes it terse, and the prompt accounts for that instead of leaving you to fix it.
Compose
Refine what you generated by hand, or write your own from scratch. You get framework scaffolding (RISEN, COSTAR, RTF and others) without an AI call, so there’s no token cost and nothing to wait for. It’s an editor that already knows what good prompt structure looks like.
Test
The Kaiju Lab runs your prompt through seven kinds of input — normal, edge cases, ambiguous, out-of-scope, adversarial, minimal, multi-part — and scores each output on four dimensions. This is where the quick happy-path check you’d run yourself misses things, and where prompts that look finished turn out not to be.
Optimise
The Optimiser turns vague instructions into specific ones, adds handling for the failure modes the Lab found, and cuts token count by a third to a half without dropping anything that matters. Battle then runs the old version against the new one so you can confirm the change was actually an improvement.
Generate — real outputs
See what comes out of one brief.
Generate reads your brief and builds the right shape for the job, not snippets you finish by hand:
- —Agents & teams — a single prompt, a skill-equipped agent, a prompt chain, or a full team — orchestrator, specialists, JSON handoff contracts, error recovery, self-verification.
- —n8n workflows — real nodes, typed parameters and wired connections, ready to import.
- —Rule files — .cursorrules or CLAUDE.md, tuned to your stack.
- —Visual prompts — image or video, in the exact syntax your target platform expects.
Brief: “A cultural research team that delivers platform trend intelligence and brand activation ideas for a given client and market.”
Generated: 7 agents. Full production system prompts. JSON handoff contracts. Error recovery. Self-verification. About a minute.
- ✓Role definition
- ✓Success criteria
- ✓Phased instructions
- ✓Self-verification
- ✓Worked examples
- ✓Role definition
- ✓Success criteria
- ✓Phased instructions
- ✓Self-verification
- ✓Worked examples
- ✓Role definition
- ✓Success criteria
- ✓Phased instructions
- ✓Self-verification
- ✓Worked examples
- ✓Role definition
- ✓Success criteria
- ✓Phased instructions
- ✓Self-verification
- ✓Worked examples
- ✓Role definition
- ✓Success criteria
- ✓Phased instructions
- ✓Self-verification
- ✓Worked examples
- ✓Role definition
- ✓Success criteria
- ✓Phased instructions
- ✓Self-verification
- ✓Worked examples
You can try every tool and every output for free. Bring your own key, and there’s no markup on top.
Test — proof
Watch a good score hide a broken prompt.
Two numbers move as the prompt changes. One of them is lying to you.
— Real-world score — how it handles a perfectly ordinary request
— Adversarial suite — how it handles attacks
The gap between them is the whole point.
Click the button below to run the first iteration and watch the score arc build.
Run the iterations and watch the real-world score arc. The suite score is the one that looked fine the entire time.
The walkthrough
The prompt is a support agent handling billing disputes. A customer writes in about a charge they don't recognise. The agent has to acknowledge the concern, commit to investigating, and reply in clean JSON, with no speculation and no invented promises.
A normal customer email is handled well from the start.
A full restructure adds explicit rules, format constraints and refusal criteria. The adversarial suite climbs. The real-world score slips, because the prompt is getting more rigid.
More hardening pushes the suite to 86, and everything looks ready to ship. But the same ordinary customer email now scores 34: the prompt refuses to help, citing INSUFFICIENT_INPUT, when it had everything it needed. Most tools would have shown you the 86 and called it a win.
One more pass aimed at the regression. The adversarial gains hold at 86, and the real-world case recovers to 90, better than where it started, with the structural improvements still intact.
The number you were watching crashed to 34 while the number you weren’t watching looked perfect. That’s the kind of regression that ships when nobody tracks both.
“Can't I just ask Claude to write the prompt?”
You can, and Promptzilla does — Claude writes the first draft here, and it's good. The difference is what happens next. That draft goes through refinement, adversarial testing, and optimisation aimed at the specific points where it fell over. The result isn't a better prompt from Claude; it's a prompt Claude couldn't reach on its own, because the improvements come from watching it fail and responding to that.
Doing the same by hand means running every prompt through edge cases, ambiguity, injection and jailbreak attempts, scoring each one consistently, and catching the version that looks finished but isn't. The generation is the easy 10%. This is the other 90%.
Optimise
A better prompt that costs less to run.
The Optimiser makes a working prompt sharper and cheaper, then Battle checks the new version against the old one on your own test cases.
Vague hedges become specific, testable instructions. It adds hallucination guards, handling for domain edge cases, and tighter output contracts, while leaving the instructions that already work alone.
It removes filler, merges overlapping instructions, and deletes restated defaults, aiming for 30–50% fewer tokens with the critical detail intact. Shorter prompt, smaller bill, less room to drift.
Every change is annotated with the reason for it and re-scored against the same suite, so you can see what moved and why before you accept it.
Old version against new on identical inputs, scored blind, with full history so you can roll back.
Often the shorter version scores higher. That's the one you'd have wanted to ship in the first place; you just had no way to find it before.
Discovery
The best setups are already on GitHub. Find them in seconds.
Someone has probably already built the Claude Code setup, Cursor rules, n8n workflow or MCP config you’re about to spend an hour on. Discovery makes finding repos you can use in your projects simple. It surfaces the strong ones, explains how they work, and installs them in a click, all in one place.
Find the proven ones.
Search top-rated repos for Claude Code, Cursor, MCP, agents and n8n workflows, ranked by stars and community signal rather than what we chose to feature.
Understand before you install.
Read the files and an auto-generated implementation guide without leaving Promptzilla, so you know what you’re importing before it touches your project.
Install in one click.
Pull it straight into your project or save it to your Library.
Improve what you find.
Optimise it, test it in the Lab, make it your own.
In the box
The rest of the kit.
EXTRACTOR
Clean and prep files for any AI model. Drop in PDFs, Word docs, slides or spreadsheets and Extractor strips them to plain text, with OCR for scanned pages, ready to paste into a prompt or attach as context.
VISUAL PROMPTS
Image and video prompts written for the platform you’re using — from Midjourney and Flux to Veo and Runway. Control style, lighting, composition and camera movement; for video, get multi-shot scripts with timing per shot and continuity carried across cuts.
Pricing
One tier. No markup. Pay your provider directly.
Free Tier
Free forever. No card, no API key required.
- ⚡No API key needed — runs on a free open model
- ⚡All tools, with limited runs
- ⚡5 agentic team generations / month
- ⚡5 Lab runs / month
- ⚡5 Optimiser runs / month
- ⚡10 saved prompts
- ⚡Extractor — clean and prep files for AI
- ⚡Community gallery
- ⚡Markdown, JSON and text exports
Promptzilla
Cancel anytime. Access until end of billing period.
- ⚡Everything in Free
- ⚡Unlimited agentic team generation
- ⚡Unlimited Lab runs
- ⚡Unlimited library
- ⚡All export formats — CLAUDE.md, .cursorrules, ZIP bundles
- ⚡Collections, tagging, version history
- ⚡Battle mode with full history
- ⚡Import and optimise existing prompts
- ⚡Priority roadmap input
Platform-only pricing. £5/month covers the platform. Bring your own Anthropic key and you pay Anthropic directly, at cost, with no markup and no proxying. No key? Start free on an open model, on us.
Payments by Stripe.
Questions