MemoryLens

Inspiration

Every developer using GPT-4 or Claude quietly pays 30–50% more than they need to. Most prompts are padded with personas ("You are an expert…"), fillers ("Additionally, I want to mention…"), and hedging ("may or may not be relevant"). None of it changes the answer. All of it costs money.

I wanted to make this invisible cost visible — in a form any developer could use in seconds, without signup, without wiring up APIs, without reading documentation.

What it does

MemoryLens is a one-page prompt cost analyzer. Paste a verbose prompt, pick your model (GPT-4, Claude 3.5 Sonnet, etc.), hit Analyze. You immediately see:

How many tokens are wasted (huge red number) How much money you save per 100K requests (green, real $ figure) Which sentences to cut — each sentence color-coded by importance (HIGH = red, MEDIUM = yellow, LOW = grey + strikethrough) A compressed version that keeps only what matters, side-by-side with the original Compression is presented as a suggestion with full visual context, not as automatic deletion — every flagged sentence stays visible, and the developer decides what to cut.

No signup. No API keys. No server roundtrip. Just clarity.

How I built it with MeDo I spent roughly a week designing the specification — the UI structure, the scoring rules (weights for specific numbers, problem/action/tech keywords, filler penalties, position bias), the tier assignment algorithm, the token estimation formula, the cost rate table for six models, the edge cases, and the design system. Then I pasted the whole document into MeDo.

MeDo read the spec and generated a working, deployed, single-page web app. The sentence-level color-coding, the fade-in animations, the model dropdown with six pricing tiers, the side-by-side before/after view — all generated from natural language.

A week of thinking became minutes of building.

Challenges

The hardest part wasn't MeDo — it was writing a specification precise enough to survive translation into code without ambiguity. Vague prompts ("make it look nice") produce vague apps. Deterministic rules ("HIGH = top 30% by score, rounded up") produce testable apps.

The lesson: MeDo rewards engineering rigor as much as creative ideas.

The harder conceptual challenge was this: can a rule-based system really decide what an LLM "needs"? Ceremonial personas ("You are an expert...") are statistically verbose, but a short functional persona ("Respond as a pediatrician") can genuinely shift output. MemoryLens cannot tell these apart semantically — so I designed it not to. The app flags candidates and lets the user decide, instead of quietly deleting. Safety through visibility, not over-claiming.

Accomplishments I'm proud of

Replicated weeks of design work in minutes once the specification was ready, without losing any rule from the original scoring engine. Demonstrated that rule-based prompt compression works — deterministic, offline, zero LLM calls, yet consistently cuts 30–50% of tokens on real-world prompts. Shipped a public, signup-free tool that any developer can try in seconds and walk away with a measurable dollar number. Built for honest limits — MemoryLens uses statistical heuristics, not semantic understanding. Rather than hide this, the UI exposes every deletion candidate so users retain full control. Safety through visibility, not over-claiming. What I learned MeDo's own framing is "whether you're a developer or a creator" — it's built for anyone with an idea, not just engineers. And the experience confirms that: the barrier to start is simply typing in plain language.

What surprised me is how much of the outcome depends on how clearly I can describe what I want. When I wrote vague instructions, I got a vague app. When I wrote a structured spec — with explicit rules, edge cases, and numbers — MeDo produced something precise.

That's not a technical skill. A designer, a teacher, a product manager, a writer — anyone who can think carefully about what they want will get that same precision back. MeDo doesn't reward coding experience; it rewards clarity of thought. And clarity of thought is something every creator already practices, regardless of background.

In a sense, MeDo shifts the bottleneck away from how to build and toward what to ask for — which is a much more democratic place for the bottleneck to sit.

What's next

Semantic accuracy upgrade: Upgrade to semantic-level importance 
scoring with a continuous learning loop that improves compression 
accuracy over time.

LLM-powered verification: Add optional LLM comparison to validate 
that compressed prompts produce equivalent outputs, moving from 
estimated to verified compression quality.

Built With

• medo