Build spec 01: interview-to-content engine
Bellhaven Wealth Advisors · Audit · Deliverable 03
Demonstration document. Bellhaven Wealth Advisors is a fictional firm created to show the format. Written the way a real spec is written: detailed enough that any competent developer could build this without talking to me.
Purpose
Turn one recorded expert interview into a compliant blog post, a four-post social series, and a three-email sequence, with compliance problems caught before a human reviewer sees the draft.
Summary
| Hours returned | Approximately 9.5 per week, from the time map |
| Build effort | Small. Estimated 3 to 4 weeks including review cycles |
| Runs in | A Claude workspace under Bellhaven billing |
| Ongoing cost | Per-seat subscription paid directly to the vendor, plus usage. No license to me |
| Owner after handoff | Bellhaven marketing team |
Inputs
- An audio or video recording of an expert interview, 20 to 60 minutes, plus a transcript.
- The brand context files built in the audit: voice guide, offering descriptions, and the compliance rule set.
- A target channel list for the run (defaults to blog, LinkedIn, and email).
The compliance rule set
This is the part that makes the build worth doing, and the part most implementations get wrong by keeping rules in prompts where they get dropped. Rules live in a file every agent reads before drafting. At minimum:
- No named clients. Any person or firm named in the source as a client gets replaced with an anonymized composite plus a privacy disclosure line.
- No performance claims. Return figures, benchmark comparisons, and any phrasing implying a predictable outcome get removed, replaced with process language and the standard disclosure.
- No specific advice. Universal instructions about allocations, contributions, or timing get reframed as education, with a direction to consult an advisor about the reader's own situation.
- Required disclosure block appended to every asset that discusses markets or performance.
Processing steps
- Ingest the transcript. Segment by topic and speaker.
- Run the compliance screen. Every flagged span gets recorded with a timestamp, a rule ID, and the proposed replacement. This log is the audit trail and is retained with the run.
- Draft each asset from the cleaned material, reading the voice guide before generation, never after.
- Re-run the compliance screen against the drafts. A draft that fails twice gets held and escalated rather than auto-corrected.
- Write the run record: source, assets produced, flags caught, and status per asset.
Outputs
| Blog post | 900 to 1,400 words, with the disclosure block |
| Social series | 4 posts, sequenced for weekly publication |
| Email sequence | 3 emails with subject and preview text |
| Run record | Asset list, compliance flags caught, per-asset status |
Acceptance criteria
The build is done when all of the following hold on three consecutive real interviews:
- Every asset carries the required disclosure block.
- Zero named clients appear in any output, verified against the source transcript.
- Zero performance figures appear in any output.
- The compliance log accounts for every flagged span with a rule ID and a replacement.
- A reviewer who knows the rules approves the blog post with no more than light edits.
Explicitly out of scope
- Publishing. Assets land in the CMS as drafts. A person presses publish.
- Approving its own compliance. The screen reduces what reaches review. It does not replace review.
- Client-facing email. Higher exposure, separate decision, not part of this build.
How you'd know it stopped working
- Compliance flags per run drop to zero for several runs. Usually means the rule file stopped loading, not that the source got cleaner.
- Reviewer edit time per asset climbs back toward its pre-build baseline.
- Output voice drifts after a model update. Re-run the voice check in the handoff documentation.