Voice Mining Kit / Productized / AI Workflows
Teach your agent to write like you.
It mines your own sent mail, builds a corpus out of what it finds, and turns that into a voice skill you can install.
// the problem
Ask an agent to write an email in your voice and you get something polished and anonymous. The usual fix is a paragraph of instructions saying warm but direct, which describes almost everybody. What actually carries a voice is mechanics: how long your sentences run, how you open a reply to someone you know, where you stop hedging, which words you would never type. All of that is sitting in your sent mail already, in thousands of messages you wrote without thinking about any of this. The work is pulling the evidence out and turning it into rules an agent can follow.
// the promise
Your agent stops guessing at your voice and starts working from evidence you wrote yourself.
// this is for you if
- You use Claude Code or Codex and keep rewriting what it drafts
- Years of sent mail are sitting there doing nothing
- Your voice should read the same in every harness you run
- Client work goes out under your name and you want a check before it ships
- Someone on your team needs this too, so the method has to be repeatable
// what’s inside
01 / The mining pass
One bounded pass over your sent mail. It confirms what it is allowed to read, runs a set of searches, keeps the messages you clearly typed yourself, and writes short excerpts into a corpus file with a register tag and a note on what each one teaches.
- $voice-mining-pass
- $corpus-template.md
02 / The skill builder
Reads the corpus, measures your sentence mechanics instead of guessing at them, and writes a voice skill into your skills directory. It follows the frontmatter contract both harnesses enforce, so the file actually loads.
- $voice-skill-builder
03 / The checker
Eight checks over any draft: em dashes, corporate filler, repeated sentence openings, negation runs, the not X it is Y construction, flat rhythm, short beat runs, and stacked triplets. Plain Python 3, so there is nothing to install.
- $voice-check
- $ban_check.py
- $ban-words.example.json
04 / Install and upkeep
The real directory each harness reads, why the folder name is the invocation, and what to do when a skill does not show up. Plus the monthly loop: mine again, rebuild, keep the corpus and throw the skill away.
- $README with the walkthrough
- $troubleshooting
// outcomes (from real implementations)
3
Skills in the bundle
8
Patterns the checker flags
2
Harnesses supported
0
Dependencies to install
// common questions
Does this work with my email provider?
The kit ships no mail integration and assumes none. It uses whatever read access your agent already has: a mail connector you have set up, an exported mailbox file, a folder of .eml files, or a batch of messages pasted into the chat. The first thing the mining skill does is ask you which of those it is working with.
Where does my mail end up?
Two answers here, since there are two halves. The corpus is a Markdown file in your project, and the checker is a Python script that opens one file and prints to your terminal, so both of those stay on your disk. Your coding agent is the other half: it reads the mail, which means those excerpts go wherever that agent already sends the files you show it, and your model provider's terms apply. Running a local model keeps the whole thing on your hardware. The mining skill reads sent mail only, confirms the reach with you first, and swaps names and account details for placeholders as it collects.
Claude Code or Codex?
Both. They read different directories, ~/.claude/skills/ and ~/.codex/skills/, but the folder layout is identical, so the same zip installs into either one. Codex reads its skills directory at launch, so restart it after installing. The README covers both paths.
What does the checker actually catch?
Mechanics. Em dashes, filler words, consecutive sentences opening on the same word, runs of negated clauses, the not X it is Y shape, sentence lengths clustered too tightly, strings of very short sentences, and three single words stacked as a beat. It prints a line number and a fix for each one, and exits non zero until the draft is clean. Whether the writing is any good is still your call.
Can I add my own banned words?
Yes, that is the point of the config file. You can add phrases, drop defaults that are normal in your field, ignore an opening word you repeat on purpose, switch off a whole check, and tune the rhythm thresholds. The builder skill also writes your ban list into the config as it goes, so the skill and the checker agree.
How long does the first run take?
About twenty minutes for a first mining pass over two years of mail, then a few minutes to build the skill. Come back monthly and run another pass. The corpus is the thing worth keeping, and the skill gets regenerated from it whenever you want.
Not sure where to start? The free scan answers that in about a minute.