Mimic My Writing#

Force any draft to sound like a specific human by extracting a measurable fingerprint from their samples and writing to those constraints. Stops the model from defaulting to LLM voice.

Quick start#

bash

# 1. Drop the user's writing samples here (markdown or plain text)
samples/<author-slug>/

# 2. Extract the fingerprint
scripts/analyze_voice.py samples/<author-slug>/

The script prints a JSON report. Read it, then draft.

The fingerprint, in one breath#

The analyzer measures rhythm (sentence-length burstiness, fragment share), vocabulary (TTR, top content words, profanity rate, AI-filler hits), punctuation (em-dash, exclaim, semicolon rates), contractions, signature 2- and 3-grams, sentence openers, and quirks (all-caps emphasis, rhetorical Q+A, "fuck" as intensifier, etc).

Each metric maps to a concrete writing rule. See references/fingerprint.md for the translation table.

Workflow#

1. Get samples. Need 2-4 pieces, ~1k+ words total. Ask if not provided. Save under samples/<author-slug>/.
2. Run the analyzer. scripts/analyze_voice.py samples/<author-slug>/. Takes <1 second, stdlib only.
3. Translate the JSON using references/fingerprint.md. Write out the constraints (sentence rhythm targets, vocab anchors, must-use signature phrases, quirks to preserve).
4. Draft against those constraints.
5. Self-audit with references/anti-ai-tells.md -- rip out LLM defaults (delve, leverage, tricolon stacks, etc).
6. Deliver.

Detailed variants (cold mimic, warm mimic, hybrid voice, critique mode, sample organization) live in references/workflow.md.

Non-negotiables#

• Never paraphrase signature phrases. They're the author's verbal tics. Drop 2-3 of them into any mimic draft verbatim.
• Match burstiness, don't average it. If the author swings between 2-word fragments and 30-word runners, do the same. Don't write a uniform-length string of sentences.
• Match the profanity rate. Diluting it sanitizes the voice; inflating it caricatures. Within ±50% of the measured rate.
• Reuse anchor words; do not synonym-cycle. If the author says "call" 10x, you say "call" -- not "conversation, dialogue, exchange."
• Honor every quirk in the quirks array. They're flags for hard constraints, not suggestions.

When to load which reference#

• Translating JSON metrics into writing rules → references/fingerprint.md
• About to ship a draft, doing the AI-tell sweep → references/anti-ai-tells.md
• Edge case (no samples, hybrid topic, critique mode, sample layout) → references/workflow.md

Failure modes to avoid#

• Mimicking the topic, not the voice. If samples are about sales and the user asks for a poem, the rhythm/vocab quirks still apply. Topic ≠ voice.
• Surface mimicry only. Copying a few catchphrases without matching sentence rhythm reads like a bad SNL impression. Stats first, vocab second.
• Bleeding authors. If samples/ has multiple authors, only analyze the requested one's folder. Don't mix fingerprints unless explicitly asked for a fusion.
• Outdated samples. If the user provides a new sample, drop it in and re-run the analyzer. Don't trust an old fingerprint from prior session memory.
• Skipping the script. Eyeballing samples and "writing in their voice" without the fingerprint is how you end up with delve+leverage soup. Run it every time.

Example mini-fingerprint readout (Chad)#

From samples/chad/:

• burstiness 0.82, fragment share 30%, long-sentence share 18% → swing hard between one-liners and 30+ word runs
• TTR 0.48 → moderate vocabulary; repeat anchor nouns
• profanity 7.7/1k words → curse freely; "fuck" as intensifier confirmed
• em-dash 8.7/1k chars → dashes are structural, not ornamental
• ALL-CAPS emphasis confirmed → use it on intensifier words (ENTIRE, MOMENT, NOT)
• Signature phrases: "ask for the money", "you can't", "that's the", "i've seen"
• Rhetorical question + self-answer → confirmed move

A draft that hits those numbers reads like Chad. One that doesn't reads like ChatGPT cosplaying Chad.