Using AI for writing without losing the draft's voice
The workflow that holds up: write a one-message brief (audience, length, tone, what to avoid), ask for a draft, ask the model to critique its own draft against the brief, then revise in the same thread. Branch when you want a different angle so the original draft survives.
Most bad AI writing comes from a bad first instruction. If the model does not know who is reading, how long the piece should be or what it must not say, it fills the gap with generic prose and you spend longer editing than writing.
ChatOTL is set up for the iterative version of this: one thread per piece, attachments for source material, branching for alternative angles and model switching mid-draft when one model keeps producing the same shape of sentence.
The four-step loop
- Brief: state the audience, the format, the length, the tone and the phrases to avoid.
- Draft: ask for the full piece, not an outline, so you can see where the structure fails.
- Critique: ask the model to list where the draft misses the brief before it rewrites anything.
- Revise: fix the listed problems one at a time in the same thread so context accumulates.
Give it your source material
Attach up to 5 files per message, 15 MB each: PDFs, images, plain text, Markdown, CSV, JSON and source code. Text and code files are extracted and passed inline to the model; PDFs and images are handed over through short-lived signed URLs. Drafting from your notes and transcripts beats drafting from the model's training data.
Branch instead of restarting
Fork any reply into a new thread to try a different angle, a shorter version or a harsher edit. The original draft stays intact in the parent thread, so you can compare rather than gamble.
Switch models when a draft plateaus
Switch between models from OpenAI, Anthropic and Google — or your own OpenAI-compatible endpoint — in the middle of a conversation. Different models fail differently, and a second opinion on the same brief is often faster than a third revision prompt.
Where it fits
Long-form pieces
Brief once, draft section by section, keep the whole piece in one thread.
Voice matching
Attach three things you have written and ask for a style sheet you can reuse.
Hard edits
Ask for the draft cut by 40% with nothing important lost, then check what went.
Example prompts
- Audience: technical buyers. 800 words, plain English, no marketing superlatives. Draft the full piece.
- Critique your last draft against my brief. List the misses; do not rewrite yet.
- Rewrite paragraph three only, keeping my sentence rhythm from the attached sample.
Limitations
- No web browsing, so it cannot research current facts, prices or sources for a piece.
- Model output can be confidently wrong; every factual claim needs your own verification.
- It imitates a voice from samples you supply — it does not know your voice on its own.
Frequently asked questions
How do I stop AI writing from sounding generic?
Constrain it before it drafts: name the audience, the length, the tone and the phrases you never want to see, and attach two or three samples of your own writing to match rhythm against.
Should I ask for an outline first?
Usually not. Outlines look fine and hide structural problems. Asking for the full draft surfaces where the argument thins out, and the fix costs one more prompt.
Can it write from my own research?
Yes — attach the notes, transcripts or PDFs and ask it to work only from the attachments. It cannot fetch anything from the web itself.
Is my draft private?
Row-level security scopes every conversation, message, setting and uploaded file to the account that created it, and you can bring your own provider API keys.
Try the loop on one real piece
Brief it properly once and see how much editing disappears.
Open ChatOTL