How Döppelscript learns your voice in under a minute

Every AI writing tool says it can write in your voice. Very few of them are willing to tell you how. The answer at most of them is: "we picked a friendly default and asked the model to be more like that." The answer at Döppelscript is different and this post walks through it step by step, because the whole product only works if you trust the voice profile it builds from your writing.

The four ways in

You can show Döppelscript your writing in four ways. Each has a different speed-versus-accuracy tradeoff, and the right one depends on what you have lying around.

Trait sliders take thirty seconds. You move eight sliders (warmth, formality, humor, verbosity, confidence, creativity, emotion, complexity) between one and ten. Döppelscript turns those numbers into a set of plain-English rules like "ALWAYS: open with a declarative statement" and "NEVER: use corporate jargon." This path is the fastest but also the least personal. It works if you just want to try the tool without uploading anything yet.

Paste samples takes about a minute. You paste three or four posts you have written, or paragraphs from a doc, or even LinkedIn comments you are proud of. Claude reads them, extracts the patterns (sentence rhythms, word choices, the topics you lean on, the stances you take), and writes the voice profile. This is the most common path for people testing the tool.

Upload a file takes a few minutes, because it runs against more content. You can upload a Substack export, a LinkedIn archive, a folder of markdown drafts, a PDF of a blog you wrote, or just a text file. The worker parses it, filters down to the posts that are actually long enough to learn from (anything shorter than about 80 characters gets dropped), and runs the analysis. This path is the most accurate because more samples means more signal.

Connect a platform is the highest-fidelity option. If you link LinkedIn, Medium, or Substack via the official OAuth flow, Döppelscript imports your historical posts directly. No copy-paste, no export-then-upload. The downside is that you have to share account access, which some people would rather not do. That is a fair tradeoff and nothing about connecting a platform is required.

What Döppelscript actually does with the samples

Here is the part most AI tools gloss over. When your samples arrive, Claude gets two passes at them.

The first pass is batched analysis. Döppelscript chunks your samples into batches of about fifty posts each and asks Claude one question per batch: "What are the notable patterns in how this person writes?" The model returns a structured JSON answer about tone, sentence length distribution, vocabulary preferences, stance, and opening and closing patterns. Each batch produces an independent analysis, so the whole thing does not depend on any single lucky pass through the content.

The second pass is synthesis. Once all the batch analyses are in, Claude reads them together and writes the final voice profile: a one-paragraph voice signature that describes how you sound in general, a list of eight to twelve plain-English rules (some ALWAYS rules, some NEVER rules), and a separate set of rules for professional register (when you want the tone dialed down a notch). There is also a statistical vocabulary extraction pass that picks out the words and phrases that appear in your writing significantly more often than in average English, so those can be preserved as "signature words."

At the end of all this, you get a voice profile you can read. Every rule is a sentence. You can edit any rule. You can add new rules. You can delete rules you disagree with. None of this happens behind a black box.

And then Döppelscript deletes the raw file.

Why the file gets deleted

This is the part that took me the longest to justify to myself as the person running a bootstrapped AI company. Keeping user corpus indefinitely would let me improve the voice analyzer over time by running new models on old uploads. It would let me debug bad generations by looking at the original source material. It would let me do all the things startup playbooks suggest when they talk about "compounding data moats."

I decided not to do any of that. Here is why.

The whole product is an anti-AI-theater bet. "Your voice, amplified" only works if you trust that your voice is yours, not an input to some training pipeline I run on the side. The moment I keep raw corpus around for "future improvements," I have to explain to a skeptical reader why that is different from the business model of every AI company whose training data controversy you have read about. The answer would have to be "trust me," and trust-me is the wrong ask in 2026.

So the corpus file is deleted the moment the voice profile is built. The voice profile itself, which is the abstract rules and patterns, is what persists. If you want to rebuild the profile tomorrow, you re-upload the file. It is slightly less convenient, and it is the honest default.

What "under a minute" actually means

I have been saying "under a minute" in the title and that is the realistic bound for the trait-sliders path and the paste-samples path with three or four samples. Uploads of larger files and platform imports take longer, usually a few minutes, because there is more content to process and Claude is doing real work.

What you do not spend any of that time on is waiting for approval, waiting for a human to review your profile, or waiting for the free trial clock to tick. There is no queue. There is no onboarding sequence. The voice profile goes from "building" to "ready" as soon as the work is done, and then you can hit generate.

The part where I tell you to go try it

You get a hundred free credits on signup, which is ten posts. No credit card. No trial window. No forced onboarding call with a sales development representative who has been instructed to be enthusiastic. If the voice profile it builds for you feels wrong, you can edit the rules directly and regenerate. If the whole thing does not fit your work, delete the account and everything is gone. The only way Döppelscript makes money is if you keep using it because it keeps being useful to you, so there is no incentive to trap you inside it.

That is the setup. Now you know how the sausage is made. If you want to watch it happen, go to doppelscript.com and upload four posts.