Five steps across the weeks
Click any step to expand its detail: objective, activities, inputs → outputs, who owns it (and which agents help), the timebox, and the one thing to watch. The path runs left-to-right, but steps overlap. They don't wait in line.
Prompts & templates that do the work
The sprint above runs on a small set of reusable prompts. Steal these verbatim, swap the {placeholders}, and paste into your assistant of choice. Named tools below are illustrative, not endorsements.
# Role: You are a content strategist auditing an existing site. # Input: a list of {existing_pages_or_sitemap} and our {top_5_buyer_questions}. Cluster every page into no more than 8 top-level topics. For each topic return: - topic name + one-line definition - the buyer question it answers - existing URLs that belong to it - gaps (questions with no page) flagged as [GAP] Then propose a flat tag vocabulary (max 25 tags) covering format, funnel-stage, and audience. No nesting deeper than 2 levels. Output as a markdown table I can paste into a sheet.
# Goal: design a minimal content model for {content_type}. Define the type as structured fields, not a blob of HTML. Return JSON: - fields[]: name, type (text/richtext/ref/media/date/enum), required (bool), and a one-line "why this exists" - relationships[]: which other types it references (author, topic, CTA) - reuse: which fields are syndicate-able to email / social Rules: max 12 fields. Every field must earn its place, if a field won't be queried, filtered, or reused, drop it. Favour references over free text so we can change once and propagate everywhere.
# Task: tag this draft against our controlled vocabulary ONLY. Vocabulary: {paste_tag_list} Draft: {paste_draft} Return: topic (1), format (1), funnel-stage (1), audience (1 to 2), plus a 155-char meta description and 3 internal-link suggestions to existing URLs in {sitemap}. Hard rule: never invent a tag outside the vocabulary. If nothing fits, return [NEEDS-NEW-TAG: suggestion] for a human to approve.
## BRAND SYSTEM PROMPT, prepend to every generation VOICE: {3 adjectives, e.g. direct, warm, expert-but-plain}. We write at a {grade-8} reading level. Short sentences. Active voice. ALWAYS: lead with the reader's problem; one idea per paragraph; back claims with a number or a named source. NEVER: hype words ({"revolutionary","game-changing","seamless"}); em-dash-stuffed run-ons; claims we can't cite; first-person "we" before the reader's "you". IF UNSURE about a fact or a claim, write [VERIFY] inline rather than guessing. Match {brand_glossary} spellings exactly.
# PRE-PUBLISH EVAL, agent scores draft 0/1 on each, blocks on any 0 [ ] On-model: every required field present & valid [ ] On-voice: passes brand system prompt (no banned words) [ ] Tagged: topic / format / stage / audience all set, in-vocabulary [ ] Claims: every stat has a source or a [VERIFY] flag resolved [ ] Links: ≥2 internal links, all resolve (no 404) [ ] SEO: title ≤60 chars, meta ≤155 chars, one H1 [ ] No PII / no confidential client names unless approved [ ] Human eyes: owner skim-read & clicked "approve" Output: PASS / BLOCK + the failing line(s). BLOCK = do not publish.
These five artifacts are your engine's source code. Version them in a repo or a shared doc, date them, and treat a prompt change like a deploy. It is one.
Three ways this sprint goes sideways
Scope creep dressed as ambition
"While we're at it, let's also redo the design system." No. The sprint ships 4 to 6 content types. Everything else is a future sprint. Park it on a list, move on.
Committees instead of an owner
One named owner makes the calls. Guardrails (the prompts above) replace sign-off meetings. The moment three people must agree per piece, your weeks become quarters.
Agents with no eval
Generating fast is easy; generating trustworthy isn't. If the pre-publish eval doesn't exist by Week 3, you're shipping unreviewed AI output. Wire the gate before you wire the volume.
Ship on Friday. Really.
The deadline is the design. A live, imperfect piece every Friday beats a perfect plan that never ships. Tighten the loop after you've closed it once.
Before you move on
By the end of this chapter you've turned a plan into a running engine. The sprint is short by design, about four weeks, but the next chapter assumes the pipeline is live, so make sure these are in place before you turn the page.
What you should have now
- ✓ A time-boxed sprint plan, the five steps mapped across roughly four weeks, with one named owner and a Friday ship cadence.
- ✓ One content type modelled and shipped end-to-end, from structured fields to a live, published piece.
- ✓ The generate → eval → publish loop running, agents author and tag against the model, and nothing ships until the eval gate passes.
- ✓ A reusable prompt kit, taxonomy, model, tagging, brand and pre-publish eval, versioned as your engine's source code.
Why this sets up the next stage
Shipping your first content through the pipeline proves the loop works, but a loop running fast is also a loop that can run wrong, fast. Next, Guardrails hardens that pipeline with lightweight checks that keep every AI-generated piece safe, accurate and on-brand without slowing the cadence you just built.
So the order matters: ship first, then harden. You've closed the loop once, the next chapter makes it trustworthy enough to scale.