AITech · EXP #08
PromptTrim
A single-purpose tool that takes a messy, conversational prompt and returns a short, direct instruction — preserving every requirement, constraint and exclusion while cutting the surrounding fluff.
All experimentsProblems solved
Context
Prompts written the way people actually talk — polite framing, backstory, restated instructions — cost tokens and dilute the actual instruction inside them. PromptTrim set out to strip a prompt down to what an AI model actually needs, without quietly dropping a requirement or a constraint in the process.
The challenge
The hard part wasn't shortening text — it was shortening it safely. The tool needed to reliably tell the difference between conversational filler that's safe to cut and a specific requirement, constraint or exclusion that must survive the edit untouched.
Research
Wrote test prompts with intentionally buried constraints — 'do NOT include X', specific tech requirements, tone notes — and used how often trimming preserved them to define what 'safe to cut' actually means for this tool.
Insights
The key findings that shaped the direction.
Politeness is safe to cut; constraints never are
Framing like 'could you please' and 'thanks so much' carries no instruction — but a buried 'do NOT include billing features' is load-bearing, and trimming has to tell the two apart every time.
Shorter isn't the goal — clearer is
A shorter prompt that loses a requirement is worse than the original — the tool is judged on whether the optimised version still gets the same result, not on character count alone.
People trust output they can immediately reuse
Returning plain text with no formatting or commentary meant the result could be copied straight into any model or tool, which mattered more than presenting it inside the product itself.
How decisions were made
- 01
Frame the product as trim, not rewrite. Positioned the tool explicitly around cutting fluff rather than rewriting intent, so the interaction model stays predictable — the output should still sound like an instruction the user meant to give.
- 02
Give people a way in without a real prompt to hand. Shipped three example prompts — dev brief, meeting notes, marketing brief — so the value is obvious on the first visit, before anyone pastes their own text.
- 03
Show the result as plain text, ready to reuse. Returned the optimised prompt as copyable plain text with no extra formatting, since the entire point is pasting it straight into another AI tool.
- 04
Say plainly where the prompt goes. Stated directly that a submitted prompt is sent to Groq to generate the optimised version and is never logged or stored, instead of leaving that as an assumption.
Solution
A single-purpose tool that takes a messy, conversational prompt and returns a short, direct instruction — preserving every requirement, constraint and exclusion while cutting the surrounding fluff.
- One-paste optimisation from messy prompt to short, direct instruction
- Three built-in examples — dev brief, meeting notes, marketing brief — to try before pasting a real prompt
- Explicit preservation of requirements, constraints and exclusions during trimming
- No account, no API key, and no logging of submitted prompts
Keep exploring
More experiments worth trying.
More time for what really matters.