Email drafter
Summary: A custom assistant loaded with your writing style, tone preferences, and relationship context. When you need to draft an email, you provide the recipient, the goal, and any context — the assistant produces a draft in your voice. You review and send.
What makes this context-level: The assistant knows your voice and adapts to relationship context, but you drive every interaction. It drafts; you review, edit, and send.
Map
The recurring workflow is email composition. It decomposes into: determine recipient and relationship context, decide appropriate tone, draft the email body, review for voice and tone accuracy, send.
The drafting step was classified as a good AI candidate — it's repetitive (multiple emails daily), structured (follows patterns by email type), and easy to verify (you read every draft and know immediately whether it sounds like you). The decision of what to say and when to send stayed with the human, as did final review: every email goes out under your name.
Test
Initial test was a simple reply to a colleague. The first prompt included basic tone guidance and a few example phrases. Output was close but too formal — generic closings, overly polite openers.
Adjusting with specific "do say / don't say" patterns fixed the voice issues immediately. Two examples from the source: "open with substance, not acknowledgment" and "no generic calls to action." The learning is that voice guidance needs to be specific about what not to do, not just what to aim for — generic instructions produce generic output.
Integrate
Built as a custom assistant with a style guide loaded as reference material. The style guide includes: opening patterns (what to say and what to avoid), closing patterns (value-add insights instead of generic sign-offs), emoji guidelines, language rules (when to use which language for bilingual users), and structure templates for different email types — reply, outreach, scheduling, follow-up, delay notice.
The assistant reads the style guide before each draft and adapts based on email type and recipient relationship. The structure templates ensure consistency across email types without having to re-specify the format each time.
Compound
The style guide itself is the compound mechanism. When the user corrects a draft — "too formal here," "I'd never phrase it that way," "this closing feels generic" — the correction gets folded back into the style guide as a new rule or example. Each correction makes future drafts more accurate.
Over time, the assistant needs fewer corrections because the style guide has captured more of the user's implicit preferences. The compound effect is measurable: early drafts might need three or four edits per email; after a few weeks of corrections feeding back, most drafts need zero or one edit.
The key mechanism is that corrections become rules rather than just one-off fixes. A correction applied once is a better draft. A correction written back into the guidelines is a better assistant.
Related pages
- compound-engineering — The style guide refinement is a concrete example of a continuous improvement loop
- ai-output-verification — Email drafts are easy to verify because the user reads every one before sending