How to Build a Content Ops Bot: Brief → Outline → Draft → Internal Links → FAQs

Quick answer: A content ops bot turns a structured brief into an outline, draft, internal-link suggestions, and FAQs — then stops for human edit before anything is published. Build it as a staged pipeline (intake → research pack → outline → draft → links/FAQs → edit gate → CMS draft), not as one “write a blog” button that ships unsupervised pages.

This is Part 4 of Build Real AI Automations. Parts 1–3 automated lead qualification, support drafts, and Shopify order status. Here we automate the agency content factory — without handing the publish button to a model.

At Let Start Design we ship a lot of SEO and AI-topic content. The win is not “AI wrote it.” The win is faster first drafts with brand constraints, internal links, and FAQ schema ready — while a human still owns accuracy, voice, and publish.

What the content ops bot produces

• Normalized brief — audience, search intent, angle, must-cover points, banned claims.
• Research pack — competing angles, questions to answer, source links to verify.
• Outline — H2/H3 map with answer-first notes.
• Draft — full article in your template (quick answer, sections, takeaways, CTA).
• Internal links — suggested anchors to existing site URLs.
• FAQs — 4–6 Q&As suitable for on-page accordion + FAQ schema.
• SEO packet — title options, meta description, slug, keyword list.

It does not: publish live, invent case studies/stats, ignore your style guide, or skip human review.

Architecture

1. Brief intake — form, Notion/Airtable row, or Slack slash command.
2. Site corpus index — your published posts/pages for internal linking (RAG over your sitemap content).
3. Stage runners — separate LLM calls with JSON schemas per stage.
4. Validator — checks required sections, banned phrases, link allowlist.
5. Edit gate — human UI (Notion doc, Google Doc, or CMS draft).
6. CMS writer — creates draft only (WordPress REST, Contentful, or your JSON/CMS).

Step 0 — Define the brief schema

Garbage briefs create garbage drafts. Require fields:

Field	Example	Why
Working title	How to build a content ops bot	Direction
Primary query	content ops automation AI	Intent
Audience	agency owners / marketing leads	Tone
Promise	pipeline they can ship in 2 weeks	Angle
Must include	edit gate, internal links, FAQs	Coverage
Must avoid	fake stats, “guaranteed rankings”	Risk
Series / cluster	Build Real AI Automations Part 4	Internal links
CTA	/contact	Business
Sources to prefer	official API docs URLs	Grounding

Store briefs in Airtable/Notion. Zapier/Make can trigger your runner, or call your API directly.

Step 1 — Build a research pack (before drafting)

Do not jump to prose. Generate a research JSON object:

• people_also_ask — questions the article must answer
• competitor_angles — what others cover (from titles/snippets you provide)
• evidence_gaps — claims that need a human-sourced example
• external_sources — URLs the writer may cite (allowlist only)
• internal_candidates — URLs from your site index with relevance scores

Feed the model your notes or exported SERP titles — or leave research semi-manual. Unbounded web browsing without an allowlist is how hallucinations and thin “AI SEO” pages happen. Google’s guidance on generative AI features still rewards people-first, evidence-dense pages (see Optimizing for generative AI features).

Step 2 — Outline stage (answer-first)

Force an outline schema:

• quick_answer (2–4 sentences)
• sections[] with h2, goal, bullets, optional h3s
• tables_needed
• faq_seeds
• cta

Human checkpoint #1 (optional but recommended): approve outline before drafting. This prevents a full rewrite when the angle is wrong.

Step 3 — Draft stage with hard constraints

System rules to encode:

• Follow the approved outline order.
• No statistics unless present in the brief/research pack.
• No invented client names or results.
• Use brand voice tokens (short paragraphs, concrete verbs, agency POV).
• Include a Quick answer block and Key takeaways.
• Output HTML or Markdown matching your CMS template.

For HTML CMSs, prefer the same Gutenberg-style blocks you already use — or Markdown that a converter turns into blocks. Validate that required H2s exist before accepting the draft.

Step 4 — Internal links from your real site corpus

Index your published URLs (title, slug, excerpt, category, embeddings). After the draft exists:

1. Retrieve top related posts/pages for the topic.
2. Ask the model to propose 3–6 anchors with exact URLs from the candidate list only.
3. Reject any link not in the allowlist.
4. Prefer series neighbors (Parts 1–3) and money pages (/services, /contact).

This is the same retrieval discipline as Part 2’s RAG support bot — your site is the knowledge base. For agent-oriented site readiness, see AI-agent ready websites.

Step 5 — FAQ generation for humans + schema

Generate 4–6 FAQs that match real buyer questions, not keyword spam. Each item needs:

• question
• answer (plain text, 2–5 sentences)

Validate: answers must not contradict the draft; no new claims. These map cleanly to FAQPage JSON-LD when you publish (as your blog system already does).

Step 6 — SEO packet

Separate call that returns:

• 3 title options (≤ ~60 chars awareness)
• Meta description (~150–160 chars)
• Slug suggestion
• 5–8 keyword phrases
• OG blurb

Human picks the title. Models are fine at options; weak at final brand judgment.

Step 7 — Human edit gate (the product)

Ship drafts to a place editors already work:

• Notion/Google Doc with checklist
• CMS draft status = draft / review
• Git PR if your content is code-managed (as on this site’s JSON posts)

Editor checklist:

1. Facts & sources verified
2. No banned claims
3. Internal links resolve
4. Examples feel first-hand where promised
5. Voice matches house style
6. FAQs match on-page content
7. Images/diagrams requested

Only after checklist → publish. Same philosophy as support: automate the draft, not the commitment.

CMS integrations

• WordPress REST API posts — create with status=draft
• Contentful / Sanity / other headless CMSs — draft entries
• Git-based content — open a PR; never push straight to production without review

Webhook security: signed requests, auth tokens in secrets, idempotent “brief_id” so retries do not spawn duplicate drafts.

Orchestration options

• Custom worker (Node/Python) — best control for agencies productizing the pipeline.
• Make / Zapier / n8n — fast MVP between form → LLM → Notion → Slack approve.
• LangGraph / workflow engines — when you need branching (outline reject → revise).

Keep stages separate. One mega-prompt is impossible to evaluate.

2-week MVP plan

Days 1–3 — Spec

• Brief schema + style guide excerpt + banned list.
• Article HTML/Markdown template.
• Success definition: “editor time to publish drops by X%.”

Days 4–7 — Stages

• Outline + draft JSON schemas + validators.
• Site corpus index for internal links.
• FAQ + SEO packet stages.

Days 8–11 — Edit gate

• Notion/Doc or CMS draft writer.
• Slack notification with brief + draft link.
• Run 5 real briefs end-to-end.

Days 12–14 — Measure

• Track edit distance / time-to-publish / factual fixes.
• Tune prompts only after measuring failures.
• Still no auto-publish.

Quality metrics

Metric	What it tells you
Outline acceptance rate	Brief clarity + angle quality
Editor time to publish	Real ROI
Factual corrections per draft	Grounding gaps
Internal links accepted vs rejected	Corpus/retrieval quality
% drafts needing structural rewrite	Outline stage failure
Published performance (later)	Do not optimize week 1

Guardrails

• Allowlisted external domains for citations.
• Ban auto-publish feature flag (default off).
• PII scrubbing if briefs include client secrets.
• Version prompts; log model + prompt hash per draft.
• Separate “legal/medical/finance” topics into mandatory SME review.
• Copyright: do not paste competitor full text into prompts — use notes/titles.

Example pipeline output (shape)

Brief: “Part 4 content ops bot for agencies.”

Bot returns: quick answer → architecture → steps 0–7 → MVP → metrics → FAQs → SEO packet → suggested links to Parts 1–3 and /contact. Editor adds first-hand screenshots, tightens voice, verifies API URLs, then publishes.

Common failures

• One-shot “write 2000 words”: split stages.
• No site index: internal links become generic or wrong.
• Fake stats: forbid numbers unless provided.
• Publishing from Zapier: always draft status.
• Skipping outline approval: expensive rewrites.
• Ignoring editor metrics: you will tune prompts blindly.

What Part 5 will cover

Next: Lead Follow-Up Sequences Without Looking Spammy — email/SMS automation triggered by form events, delays, stop rules, and CRM logging that protects deliverability.

Key takeaways

• Content ops bots are pipelines, not magic publish buttons.
• Brief schema + research pack beat bigger models.
• Internal links must come from an allowlisted site corpus.
• FAQs and SEO packets are first-class outputs.
• Human edit gate is mandatory; measure editor time saved.

Want this pipeline wired for your agency or content team — brief form, staged LLM workers, internal-link index, and CMS drafts? Talk to Let Start Design. We build the sites and the content systems that keep quality under human control.

Related: Part 3 — Shopify order status bot · Part 2 — Support triage bot · Part 1 — Lead-qualification bot · Agents vs chatbots vs copilots

Sources: Google — Optimizing for generative AI features; WordPress REST API — Posts; Contentful developer docs; Structured model outputs.