VoiceCheck
An AI editor that reads your copy the way a sharp human would — margin notes included.
Overview
VoiceCheck is an AI writing editor built on an unfashionable belief: writers don't want a score, they want an editor. Paste your copy and it comes back as a marked-up manuscript — flagged lines underlined, margin notes explaining what's off, suggested rewrites a tap away.
It treats tone not as a metric to grade but as an identity to protect, closing with a Voice Fingerprint: a portable card describing how you actually write.
The Challenge
AI writing tools output sterile dashboards — 'Clarity: 74' — that nobody acts on, because a number carries no instruction. Meanwhile the feedback writers do act on has always looked the same: a trusted reader's marks in the margin. The challenge was building AI feedback that writers trust enough to accept, which is a voice problem before it is a model problem — the same diagnosis, phrased like a scold, gets dismissed.
Goals
- 01
Replace scores with line-level diagnosis anchored to the exact sentences that triggered it.
- 02
Make the feedback voice feel like a sharp, generous editor — direct enough to trust, warm enough to hear.
- 03
Keep the writer in a manuscript, not a dashboard — the page itself is the interface.
- 04
Degrade gracefully: when model output falls short, fall back to structured JSON rendering rather than hallucinated polish.
Research
The design research was a study of how feedback is actually received: editorial letters, workshop marginalia, and the specific phrasings human editors use to criticize without triggering defensiveness. The consistent pattern — name what works, locate what doesn't, propose the fix in the writer's own register — became the product's response template.
A parallel audit of AI writing tools confirmed the gap: every one of them surfaces aggregate scores first and buries the line-level 'why.' None of them write like editors; all of them write like report cards.
Strategy Thesis
“The tone of the feedback IS the product.”
Everything else follows from that sentence. The interface is a manuscript because editors work in manuscripts: your text renders as a document with flagged underlines, and each flag opens a margin note — observation, reason, rewrite — in an editorial voice that was prompt-engineered as carefully as the UI was designed.
Charts were banned deliberately. The one artifact that summarizes you — the Voice Fingerprint — is a designed identity card, not a radar plot: something a writer would actually keep and share. And because trust dies the first time the product fakes competence, the system prefers an honest structured fallback over a confident mess.
Design Process
- Step 01
Write the editor before building the app
The feedback voice was drafted as literal sample margin notes — dozens of them, tuned until they sounded like a specific person — and the model was then constrained to that register. The UI was designed around what those notes needed.
- Step 02
Choreograph the reveal
Analysis doesn't dump onto the screen. Flags surface progressively down the page — a reading rhythm, not a report load — so the writer processes each note in the context of their own line.
- Step 03
Engineer the fallback path
Model responses are demanded as structured output; when a response fails validation, the product renders the structured JSON honestly rather than reflowing broken prose into fake margin notes.
Final Solution & Gallery
The live product takes any pasted text — a loaded corporate sample is one tap away — and returns the marked-up manuscript: underlines to tap, margin rewrites to accept, and a Voice Fingerprint card to export at the end.
Results & Metrics
What the build commits to, verifiable in the linked product:
- no 0–100 number anywhere in the product
- 0 scores
- every note is anchored to the exact flagged sentence
- Line-level
- from underline to margin note to accepted rewrite
- 1 tap
- honest structured fallback when model output degrades
- JSON
Key Takeaways
- 1
For AI products, voice design is interaction design — the register of the output determines whether it gets acted on.
- 2
Anchoring feedback to the user's own lines converts criticism from a grade into a conversation.
- 3
An honest fallback state builds more trust than a polished failure.
Concept client, real build — every interaction shown is live in the linked product.