关键要点
- Use topic brief → outline → draft → brand check → human edit flow; don't ask AI for final copy
- Define brand voice in system prompt with examples (tone, vocabulary, structure); refresh quarterly
- Enforce output format: Outline must be JSON; draft must be Markdown; no exceptions
- Add structural validation: Blog post must have title, 3+ sections, links, and CTA; reject if missing
- Human review always required; content teams use AI for speed, not to eliminate editors
Standard Content Workflow
Break content into 5 stages; each stage has input spec, prompt, and output validation.
- Stage 1 — Outline: Topic → AI generates outline (JSON); human approves structure
- Stage 2 — Draft: Outline → AI writes sections (Markdown); human edits for brand voice
- Stage 3 — Fact-check: Draft → Human verifies claims; adds links and citations
- Stage 4 — Format: Markdown → Publish format (HTML, Medium, Substack); last-minute adjustments
- Stage 5 — Publish & track: Post live; monitor engagement; feedback to writers
Stage 1: Outline Generation
AI excels at structure; have it generate 3–5 outline options, human picks best.
- Prompt: "Create 3 outlines for blog post about {topic}. Each outline: 1 intro section, 3–4 body sections, 1 conclusion. Output as JSON array."
- Output format: `"...", "sections": [{"heading": "...", "key_points": [...}] }]`
- Human step: Pick best outline; edit section headings for clarity
- Quality gate: No outline is used without human approval; AI can't judge what resonates with audience
Stage 2: Draft Writing
AI writes each section in parallel; human stitches together, edits for voice.
- Prompt per section: "Write {section heading} for blog about {topic}. Brand: {voice profile}. 400–500 words. Markdown."
- Brand voice profile: "Conversational but authoritative. No jargon. Analogies to real-world examples. Tone: friendly, slightly irreverent."
- Human edit: Read each section; rewrite openings/closings; ensure voice consistency; fix jargon
- Time estimate: AI 1 min per section, human 10 min total for brand pass
Enforce Brand Voice
Brand voice lives in the system prompt, not as post-hoc notes; refresh every quarter.
- Voice profile in system prompt: Include 2–3 examples of "good copy" and "bad copy" in your brand
- Example good: "Use analogies ("like a GPS for your prompts"), avoid marketing (no "seamless" or "powerful")"
- Example bad: "Avoid corporate-speak: don't say "leverage", "synergy", "drill down""
- Quarterly refresh: As company voice evolves, update examples; test on 5 recent posts
Stage 3: Fact-Check and Links
AI writes claims; humans verify claims and add sources.
- Before publishing: Human must verify 3+ claims (dates, statistics, product features)
- Link requirement: Blog post must include 4+ internal links and 2+ external sources
- Citation format: Standard BibTeX or Markdown links with hover text
- Common issue: AI cites non-existent studies; always fact-check before publishing
Output Validation Rules
Define what "complete" means; reject anything that doesn't meet spec.
- Blog post: Title (50–70 chars), Meta description (150–160 chars), Intro with date signal, 3+ H2 sections, 1+ image caption, CTA
- Email: Subject (40–50 chars), Preview text (100 chars), Greeting, 2–3 short sections, 1 CTA, signature
- Social post: <280 chars, 1 image, 1 hashtag, 1 link (or 2 hashtags if no link)
- Rejection rule: If any required element missing, return to writer with checklist
Common Mistakes
- Asking AI to write final copy without human edit—brand voice diluted or lost
- Using same prompt for all content types—blog prompts don't work for emails
- No fact-checking workflow—false claims published; damages credibility
- No output validation—missing CTAs, inconsistent formatting, tone varies wildly
- Not versioning prompts—can't reproduce good results; can't debug bad ones
Sources
- PromptQuorum content workflow case study, April 2026
- Brand voice guidelines from 5+ marketing teams, shared via industry research