DEV Community

Hamza KONTE
Hamza KONTE

Posted on

A colleague told me my prompt was garbage. He was right.

Role, constraints, examples, output format. All in one paragraph. The model has to guess what's what.

I kept rewriting my prompts, re-running, tweaking one sentence, re-running again. Ten tries to get something decent. Not because the model was bad. Because my instructions were a mess.

The problem nobody talks about

When you write a prompt, you think in categories. You know which part is the role, which part is a constraint, which part is an example. But the model doesn't. It sees one block of text and has to infer the boundaries.

Sometimes it gets it right. Sometimes it treats your example as an instruction. Sometimes it ignores your output format because it blended into the context paragraph above it.

The fix isn't a better model. It's showing the model where each piece starts and ends.

So I built flompt

flompt pre-interprets your prompt before the model does.

You paste your prompt. AI splits it into typed blocks: role, audience, objective, constraints, examples, chain of thought, output format, response style, language. You see exactly how your instructions will be read.

If the interpretation is wrong, you fix the blocks. Then you compile and get a structured prompt that the model parses correctly on the first try.

Same model. Same task. Night and day.

What it looks like

Three ways to use it

Web app at flompt.dev. No account, no install. Paste, decompose, edit, compile.

Browser extension for Chrome and Firefox. Adds an "Enhance" button right inside ChatGPT, Claude, and Gemini. You structure your prompt in a sidebar without leaving the conversation.

MCP server for Claude Code. Your agents can call decompose_prompt() and compile_prompt() programmatically:

claude mcp add --transport http --scope user flompt https://flompt.dev/mcp/
Enter fullscreen mode Exit fullscreen mode

50+ templates

Don't want to start from scratch? There are 50+ templates built from the best prompts on GitHub. Code review, blog post, sales pitch, data analysis, lesson plan. Load one, tweak the blocks, compile.

Stack

React 18 + TypeScript + React Flow + Zustand (frontend), FastAPI + Python 3.12 (backend), Caddy (reverse proxy). 10 languages supported.

Free. Open-source. MIT.

Try it: flompt.dev
Repo: https://github.com/Nyrok/flompt

If you find it useful, a star on the repo would mean a lot. Solo project, every star helps with visibility.

Top comments (0)