Help articles fail when they re-state the UI. Education articles win when they teach the underlying model, so the reader can solve the next, adjacent problem without opening a ticket. The prompt has to force that distinction, and it has to make the model put the answer in the first two sentences. A good help article and a page that ranks in 2026 are now the same thing: clear answer first, context second.
TL;DR
- Use the 12 templates below. Each one bakes in audience, a single outcome, a structure, and an answer-first opening.
- Draft with the right model: Claude Opus 4.7 for help docs that need accuracy and the least “AI voice,” GPT-5.5 when you want the most natural-reading prose and faster output. Both are fine; the difference is small.
- Coverage drives ticket deflection. A knowledge base with 50-200 articles deflects roughly 20-35% of tier-1 tickets; 200+ continuously updated articles reach 40-60% (best-in-class B2B SaaS, as of June 2026).
- AI writes drafts 1 and 2. A human ships draft 3. Never publish the first pass unedited.
Who this is for
Support leads, customer education writers, founders writing their first help center, and devrel teams writing concept docs.
When not to use these prompts
Don’t use them for marketing copy. Don’t use them when the underlying product is broken; fix the product first, then document. And don’t write an article for a one-off ticket. Document the clusters, not the outliers.
Why coverage matters more than polish
The single biggest lever on tickets is not article quality, it’s article count. Deflection scales with how much of your real ticket volume the knowledge base actually covers.
| Knowledge base maturity | Articles | Tier-1 ticket deflection |
|---|---|---|
| Early-stage | 0-50 | 10-20% |
| Growing, updated quarterly | 50-200 | 20-35% |
| Best-in-class, updated continuously | 200+ | 40-60% |
Figures are realistic B2B SaaS ranges as of June 2026; rates above 80% usually mean the bot is answering confidently and wrongly, so always pair deflection with a CSAT or resolution-accuracy check. This is exactly why AI drafting matters: it lets a small team get from 30 articles to 200 without the quality collapsing.
Pick a model for help docs
For drafting customer-facing help content as of June 2026:
- Claude Opus 4.7 (Pro $20/mo, or API at $5/$25 per 1M tokens in/out) reads least like AI and follows a detailed brief most reliably. Best when a wrong claim has a real cost, which is most help docs.
- GPT-5.5 (ChatGPT Plus $20/mo, or API at $5/$30 per 1M tokens) produces the most natural prose and adapts tone fast. Strong for friendly onboarding copy.
- Gemini 3.1 Pro (Google AI Pro $19.99/mo) has a 1M-token context, handy when you paste 20 existing articles for a tone audit (template 10).
Whichever you use, the prompt does the heavy lifting. Model choice is a second-order decision.
Prompt anatomy
Every education prompt should carry six elements:
- Audience: one specific reader, not “users.”
- Goal: one outcome the reader should reach.
- Voice: 2-3 anchor adjectives.
- Constraints: word count, banned phrases, must-include facts.
- Format: paragraph, bulleted, headed, or table.
- Examples: 1-2 tone samples. This is the strongest lever for matching voice.
In every template below, swap the backtick [placeholders] for your real values before running.
12 copy-ready prompt templates
1. Task article skeleton
Task: [task]. Audience: [userPersona]. Write a help article: (1) one-sentence answer-first opening (what the reader will achieve), (2) When you'd do this, (3) Prerequisites, (4) Numbered steps, each naming where the relevant button or screenshot lives, (5) Success check, (6) Troubleshooting (3 common issues), (7) Related links.Swap: task, userPersona
2. Concept article
Concept: [concept]. Write a 600-word article: (1) plain definition in the first 2 sentences, (2) Why it exists (the problem it solves), (3) How it works at a high level, (4) Where you'll see it in the product, (5) Two common misunderstandings. No "in today's world" openers.Swap: concept
3. Troubleshooting article
Problem: [problem]. Write a troubleshooting article: (1) Symptoms (exact error text if any), (2) Most likely cause and how to check it, (3) Less common causes, (4) When to contact support, (5) What info to gather before contacting (logs, IDs, screenshots).Swap: problem
4. Onboarding-week article
New users get stuck on [feature] in the first week. Write a "your first [feature]" article: (1) Outcome promised in the opening line, (2) 5-step walkthrough with a success check per step, (3) What changes once they do this, (4) Next thing to learn.Swap: feature
5. Glossary entry
Term: [term]. Write a 100-word glossary entry: (1) Definition in 1 sentence, (2) Why it matters here in 1 sentence, (3) Where the user encounters it, (4) Link to a deeper article. Be plain; skip "in plain English" wrapper sentences.Swap: term
6. Comparison article (help-center version)
Help-center article: "Should I use A or B?" Output: (1) Quick verdict in one line, (2) When to use A (3 specific signals), (3) When to use B (3 signals), (4) How to switch later if you chose wrong. Don't hedge.7. Best-practices article
Best practices for [task]. Output 5 practices. Each: (a) the practice in 6-10 words, (b) why it matters, (c) the anti-pattern to avoid. Don't restate the UI; talk about the decision behind it.Swap: task
8. Change-of-behaviour article
We just changed [behaviourChange]. Write a user-facing article: (1) What changed, (2) Why in 1 sentence, (3) What users need to do, if anything, (4) When the change is effective, (5) Where to opt out or get help.Swap: behaviourChange
9. Walkthrough video script
Turn this help article into a 90-second walkthrough script: (1) Cold open (5s): the outcome, (2) Steps with on-screen action notes, (3) Outro (5s) with the "next" link. Keep each section under 20s of screen time. [paste article]10. Bulk tone normalisation
Audit these help articles for tone: [articleList]. Flag articles whose tone clashes (some friendly, some terse). Pick one target tone, then suggest 5 concrete fixes per inconsistent article.Swap: articleList (paste the articles; Gemini 3.1 Pro’s 1M context handles 20+ at once)
11. Reduce-ticket article angle
Top support ticket clusters this month: [tickets]. For each cluster, design a help-article angle that lets users self-solve. Output: 3 article titles plus the core promise of each. Skip one-off tickets.Swap: tickets
12. Help-article hygiene audit
Audit this help article: (1) Does the title match search intent? (2) Does the opening answer the question in one sentence? (3) Do steps have success checks? (4) Outdated screenshots? (5) Does it link to neighbour articles? Output a 5-item fix list. [paste article]Common mistakes
- Vague audience, so the output reads generic.
- No tone anchor, so every variant comes back the same flavour.
- No constraints on word count, banned phrases, or length cap.
- Skipping examples. Examples are the strongest signal for voice.
- Trusting the first draft. AI lands on the safe middle.
- Burying the answer. In 2026, answer-first openings are what get cited in AI search results, so lead with the answer in the first 100-150 words.
- No fact-check pass. AI is confidently wrong often enough to matter.
How to push results further
- Give 1-2 real tone examples. “Be friendly” is noise.
- Constrain ruthlessly: word count, banned phrases, must-include facts.
- Read the draft aloud before publishing.
- Cut adverbs and adjectives that don’t carry weight.
- Let AI write drafts 1 and 2; a human edits draft 3, and 3 is what ships.
- Anchor every prompt in a real person from your audience.
- Test the headline standalone against the search query it should rank for.
FAQ
- Which model should I draft help articles with?: As of June 2026, Claude Opus 4.7 reads least like AI and is safest when a wrong claim costs you; GPT-5.5 writes the most natural prose. Both work on a $20/mo plan.
- How many articles do I need before tickets drop?: Deflection meaningfully starts around 50 articles (20-35%) and reaches 40-60% past 200 well-maintained articles. Coverage of real ticket clusters matters more than count alone.
- Can AI write the whole draft?: AI for the first two passes, a human for the third. Never publish the first pass.
- How long should a help article be?: As long as it takes to fully answer the question and no longer. Put the answer in the first two sentences regardless of length.
- How often should I refresh articles?: When the product, audience, or a screenshot changes, and a quarterly sweep for evergreen pages. Run template 12 as the checklist.
- Can I reuse these prompts for other content?: Yes. Swap audience, goal, and voice. The structure carries over.