Proposal · not built yet

Writing Harness Lite

A writing helper that learns your voice from your own writing and edits, and gets better the more you use it.

In one sentence: you say what you need written. It looks at how you've written similar things, what has worked before for that kind of message, and what you've corrected in the past. Then it drafts in your voice and remembers your edits for next time.

1. The problem

AI writing sounds like AI: too polished, too even, generic words. Today's fixes are either "humanizer" skills that strip obvious AI words, or voice profiles written once and never updated. Neither learns from you over time, and neither knows what works for a particular kind of message.

2. What it does differently

Learns from your edits

The gap between its draft and what you actually sent is the best lesson. It saves those changes and turns repeated ones into rules you approve.

Knows what works

It keeps patterns from messages that got results (a reply, a yes), so a cold email borrows from cold emails that worked.

Uses your real writing

It shows the AI real examples of your writing for that kind of task. That works better than describing your style in the abstract.

Everything is visible

Your voice is stored as readable notes and examples, not a black box. You can see each rule, edit it, or undo it.

3. The architecture, in plain terms

Think of it as four layers. Only the top two are new; the bottom two already exist.

You"Write a follow-up to the recruiter at Acme, friendly but direct."
The writing workflow New. The fixed steps every request goes through: understand the task → gather examples → draft → check → learn.
Your writing workspace New. A folder of your samples, approved voice rules, past corrections, patterns that worked, and outcomes.
The AI runtime you already use Existing. Pi or Hermes. It does the AI calls and the conversation.

Why not a whole new harness? A harness is the engine that runs an AI agent, and Pi and Hermes already are one. Building another means rebuilding a solved part. What's new here is the knowledge about your writing and a reliable process for learning it, so that's what we build. The engine stays swappable.

Why more than a skill? A skill is advice the AI may or may not follow. The workflow is real code, so saving, learning and undoing happen every time.

4. How it works when you use it

You ask. In any chat with your agent: "Draft a reply to Sam declining the meeting." No settings and no genre to pick.
It figures out the type (a decline email) and the audience (a colleague).
It gathers context: 3–5 of your real messages like this one, your approved voice rules, and patterns from similar messages that worked.
It drafts in your voice, fitting the structure to this task rather than one template for everything.
It checks that the draft says what you meant and nothing extra, and flags obvious AI habits.
You edit and send. It quietly saves its draft next to your final version.
It learns. When the same kind of edit keeps coming up ("Logani never opens with 'I hope this finds you well'"), it suggests a new rule. You approve it or not.

5. What's in your workspace

Folder	What it holds	Who adds to it
My writing	Things you actually wrote: emails, posts, docs	You (upload), or your sent final versions
Writing I admire	Other people's work you want to borrow from, kept separate from "your" voice	You
Voice rules	Short, readable rules, each backed by real examples	Suggested by the tool, approved by you
Corrections	Pairs of its draft and your final version	Saved automatically
What works	Structures from messages that got the result you wanted	From outcomes you mark (got a reply, got a yes)

It's all plain files with full history, so any change can be undone.

6. How it improves over time

When	What it can do
Day 1	Drafts from your uploaded samples. Already more "you" than plain AI.
Weeks	Has learned your common corrections, e.g. how you open, sign off, and how blunt you are with whom.
Months	Knows which structures get replies for your kinds of messages, and your voice in each setting.
Later (optional)	If there's enough data and it beats the simple approach, train a small custom model. Not before.

7. Safeguards

It must not learn its own writing. AI drafts stay labeled as AI drafts. Even approved drafts don't count as proof of your natural voice. Otherwise, over time it would copy itself instead of you.

"What works" mustn't flatten everything. Successful patterns are suggestions fitted to the task, not one template for every message.

Privacy, honestly: your workspace stays on your machine, but the examples it sends to a cloud AI to write the draft do leave it. A fully private option would need a local model.

8. How we'd prove it works before building more

A blind test on real writing tasks you haven't used for training. The same AI writes three ways:

Version	What it gets
A	A plain prompt
B	The prompt plus your real examples
C	Examples plus your learned rules and corrections

You score each one, without knowing which is which, on sounds like me, says what I meant, and ready to send.
Decision: if C clearly beats B, the learning part is worth it and we keep going. If not, we keep the simpler example-based version and stop. Either way, we know.

9. What we're building vs. not

Item	Status
Writing workflow + workspace on Pi / Hermes	v1
Learning from your edits, with approve/undo	v1
Tracking outcomes ("what works")	v1, simple
Callable from other apps (MCP)	when a second app needs it
Training a custom model	only if it beats v1
A brand-new agent harness	no
"Undetectable by AI detectors" promises	no

10. What we need from you

About 20–30 pieces of your own writing, mixed types are fine.
About 10 real upcoming writing tasks for the blind test.
Your call: is the cloud AI OK for drafting, or must it be fully local?
A "go" to build. Nothing gets built until you say so.