Content interview

Summary: A custom assistant that produces a week's worth of content (newsletter, social posts) by first interviewing the user topic-by-topic about what they've been working on, then drafting each piece in their voice using a loaded style guide and recent vault material as source. The user drives every interaction — answering interview questions, reviewing drafts, accepting or rejecting suggestions.

What makes this context-level: The assistant knows the user's voice, has the style guide loaded as reference, and can pull from recent work — but the user is in the conversation throughout. The interview structure means the human is generating the source material in real time; the AI is synthesising and drafting, not deciding what to write about. Compared to a fully automated content pipeline, this keeps the editorial judgement with the human.

Map

The recurring workflow is producing weekly content for a newsletter and social channels. It decomposes into: review the week's work and decide what's worth talking about, structure the topics, draft a newsletter piece around each, draft social posts derived from the newsletter, edit each draft against voice and style, schedule and publish.

Not every step is an equal candidate for delegation. The drafting steps were classified as good AI candidates — they're rewriting and synthesis tasks against well-defined inputs (interview answers, style guide, recent work). The decision of what to write about and the final voice editing stayed with the human, because both depend on judgement that's hard to delegate cleanly.

Test

Initial test was a single prompt to draft content directly from a folder of raw notes. The output was technically correct but generic — it missed the angle that makes a piece worth reading because the AI had no way to know which thread inside the notes the human actually wanted to pull on.

Adding an interview structure before the drafting fixed the genericness. The AI asks pointed questions topic-by-topic, the human answers in their own words, and the draft uses those answers as the spine. The interview surfaces the angle the human cares about without the human having to articulate it upfront — which is the part that's hardest to do before you start writing.

Integrate

Built as a custom assistant with three loaded artefacts: a voice and style guide describing how the user writes (sentence patterns, what to avoid, calibration examples), a recent-work folder the assistant can scan for material (notes, sources, relevant captures), and a content calendar template for the output structure.

The interview is the entry point — the assistant asks one topic at a time, captures the answers, and only starts drafting after the interview is complete. Drafts get written into a week folder where the user edits them in place and copies the final versions to publishing tools.

Compound

Two compound mechanisms run in this workflow.

The style guide grows. Every editing pass surfaces small corrections — a phrase the user would never say, a structural pattern that worked, a transition that didn't. These corrections get folded back into the style guide as new entries: do say / don't say patterns, opening templates that worked, closing patterns to avoid. Over time the assistant needs fewer corrections because more of the user's implicit preferences have been written down.

The interview prompts refine. Some questions reliably surface useful material; others tend to produce filler the user later cuts. The questions that work get reused across topics; the ones that don't get rewritten or dropped. The interview gradually becomes a sharper instrument for getting the human's voice onto the page.

Both mechanisms improve the workflow without rebuilding it. The assistant, the style guide, and the interview prompts are the same system; they just get more accurate through use.

example-email-drafter — The same compound pattern (style guide growing through corrections) shows up in a simpler form for email; this example is the more structured sibling
compound-engineering — Two parallel compound mechanisms (style guide refinement and interview-prompt refinement) running in the same workflow
productive-friction — The interview structure deliberately preserves human input where a fully automated draft would lose voice; this is friction kept on purpose