OpenPrompt

Context

Rough prompts often contained duplicated examples, unnecessary context, inconsistent formatting, and unclear output requirements, increasing cost and making results less predictable.

Problem

Simply shortening a prompt can remove details that materially affect its outcome. The tool needed to reduce waste while preserving intent, constraints, context, and expected output.

What I built

I built OpenPrompt with Next.js, the ChatGPT API, Supabase, Google authentication, and a private Vercel deployment. It was initially used by the developer team and is now maintained as a private tool for my own use.

How it works

1. 01Accept a rough prompt and identify its goal, required context, constraints, output contract, and missing information.
2. 02Apply semantic compression to remove repetition, filler, duplicated examples, and unnecessary formatting.
3. 03Generate a compact natural-language prompt or a structured JSON/XML version when machine readability is required.
4. 04Tokenize the original and enhanced prompts using the target model's tokenizer and calculate exact token and estimated cost differences.
5. 05Flag potentially important removed details so the user can balance efficiency with output quality.
6. 06Save reusable prompts in Supabase as stable artifacts for future engineering and product workflows.

Outcome

• Typically reduces prompt token usage by 20–40%.
• Makes code-generation, agent-behavior, evaluation, content, and setup prompts easier to reuse.
• Provides exact token and estimated cost comparisons rather than subjective claims of improvement.

What I learned

Prompt engineering is strongest when prompts are treated like product components: structured, measurable, reusable, and explicit about the tradeoff between compression and lost context.