Forest & Fiber
Free · Claude + ChatGPT

Make AI write like you.

Not like generic AI. Voiceprint turns a handful of things you've actually written into a reusable voice profile, and strips the AI tells that make every post sound the same. Works in Claude, ChatGPT, or whatever you already pay for.

1Get the file below. The .skill for Claude, the plain SKILL.md for everything else.
2Claude: drop the .skill into a Project, or save it as a skill.ChatGPT: open SKILL.md, copy the text into a Project's instructions or a Custom GPT.
3Add 5 things you really wrote. It writes as you from then on.
↓ Download voiceprint_linkedin_19_slim.pdf Prefer plain text? Read / copy SKILL.md · reference.md · the empty profile · humanizer extension · how to test it

What it does: it measures how you actually write, sentence spread, punctuation, connectors, the words you never touch, then asks you the things samples cannot show. What you leave out. What you never soften. When you go blunt. That becomes a profile in five layers, with a separate delta per platform and per register, because a rant and a tutorial are further apart than two platforms. It also keeps the analogies you reach for again and again, and it learns from every correction you make to a draft. It scores its own draft on six dimensions before showing it to you, so you see whether the problem is the voice or an unchecked fact, and it tells you when your style has drifted. There is also a test protocol, because a method like this is worth exactly as much as its evidence. No magic prompt. A file, plus a template you keep filling.

Under the hood

No black box. This is every step the file runs, what it stores, where that lives, the two loops it keeps running after the first draft, and how you test whether any of it works.

WHAT THE SKILL ACTUALLY DOESEvery box below runs inside your own assistant. Nothing is uploaded, there is no account, and noserver of mine ever sees your writing.AYOU BRINGINPUT5 to 10 things you really wrote, collected per surface, not in one pile2 to 3 anti-samples: writing in your field that makes you cringe, and whyNear misses: text that is 95 percent you and 5 percent not, which is sharperNothing AI-touched. One polished sample and it learns the robot, not youBGRILL ME, THE FIRST OF THREE TIMESTHE INTERVIEWLayer 1, identity: what you are for, who you are talking toLayer 2, beliefs: the take you repeat that people argue withLayer 3, decisions: what you leave out, what you never soften, what makes you bluntThe stock: what you compare things to, the story you keep tellingOne question at a time, and it always leads with its own guess so you only correctCCOUNTED, NOT GUESSEDMEASURE17 counted numbers: sentence spread,commas, colons, connectors, firstperson, word length, list share,contractionsA fingerprint to compare against later.Never a target to write toward, becausewriting to hit a number is just a newmachine tellDSEVEN LAYERS, PERSON WINSSTRIP THE TELLSSeven layers of tells: words, phrases,structure, rhythm, punctuation, stance,densityIf your own samples genuinely contain atell, you win over the list. You neverget sanded down to a neutral averageETHE ARTEFACTYOUR PROFILEFive layers, from identity down to your signature wordsThe stock: the analogies and examples you reach for again and againThe revision reflex: what you always change, learned from every correctionA delta per surface and per register, because a rant is not a tutorialA NEVER list, a tagged sample index, and a dated changelogWHERE THE DATA GOESIT STAYS WITH YOUThe profile lives wherever you put it: project instructions, a saved skill, or a file on yourmachine. It is text you can read and edit.Curated out before anything is stored: names, client identifiers, company internals, andanything touching health, money, politics or other people.LOOP 1, EVERY SINGLE DRAFTFRUNS EVERY TIMETHE DRAFTING LOOPLoads the core profile, the matching surface delta, and 2 or 3 indexed samplesDemands your raw material first: notes, real numbers, your actual opinionDrafts in sections, never in one shot, because one shot has one flat toneAnti-smoothness in two moves: mark the generic spots and change nothing, then rewrite only whatwas markedTHE GATESHIP-CHECK AND CONFIDENCEIt scores itself on six dimensions before you see the draft.EVIDENCEAUTHENTICITYSTRUCTUREVOICEFACTSDECISIONSThe gate acts on the LOWEST one, never the average, because an average of 80 is how a Factsscore of 40 disappears behind five good numbers.LOWEST 85+shipsLOWEST 60 TO 84ships, that dimension namedLOWEST UNDER 60stops and asks for what is missingA FAILING DIMENSION GOES BACK FOR MORE INPUTLOOP 2, EVERY 10 SAMPLES OR EVERY 6 MONTHSGTHE LONG LOOPDRIFT CHECKRe-measures the new samples with exactly the same recipeAnything that moved more than about 25 percent is worth a conversationDeliberate change goes into the core, a one-off becomes a noted exception, and a shift on oneplatform only goes into that surface deltaGrill me runs again here: was this on purpose, or was it just a strange monthOPTIONAL BRANCH, ONLY IF YOU ASK FOR ITTHE HUMANIZER LOOPIt lives in its own file, because your voice is stable for years and detectors change everyfew months. Length first, then a rewrite in a different model family, thendetector-in-the-loop on the flagged sentences only.The honest ceiling, measured: this took ZeroGPT from 100 to 0 and QuillBot from 55 to 0.GPTZero did not move once in 12 tries.HOW YOU CHECK THAT IT WORKSTEST IT, DO NOT TRUST ITFive texts you really wrote against five written with the profile, judged blind by someone whoknows your writing. Around 50 percent recognition is the target.There is also an ablation test, which removes one component at a time to find out which partsare actually earning their place for you. Protocol in evaluation.md.

Swipe the map sideways, or open it full size.

About this file

Voiceprint is free. You get it as it is, with no promise that it fits your case and no obligation on my side to keep it updated.

It is a set of written instructions for an AI assistant. It does not run code, it does not access, change or delete anything on your device, and it sends nothing anywhere.

Whatever the AI writes with it stays your text and your responsibility. Check facts before you publish, and do not put confidential information or other people's personal data into your writing samples.

What it will not do is get you past GPTZero. I measured 12 variants against it and every single one came back 100% AI. Other detectors are a different story: the same text dropped from 100% to 0% on ZeroGPT and from 55% to 0% on QuillBot once it was written through a voice profile. GPTZero was the only one that never moved, and I would rather tell you that than sell you a guaranteed pass.

As this is a free offer, I am liable only for intent and gross negligence (§ 521 BGB). Liability for injury to life, body or health remains unaffected.