Nick QuirkMost agencies have already run the experiment. Someone types "write a blog post about local SEO for dentists," gets 900 words of confident mush, and concludes AI is not ready for client work. The model was not the problem. The prompt was a one-line brief, and a one-line brief would produce the same result from a human writer.
The teams getting real leverage treat AI prompts for marketing deliverables as production assets — versioned, reviewed, and built from the same inputs a good creative brief contains. This is how to structure them.
Why unstructured prompts produce unusable output
A deliverable has a buyer. A content brief is consumed by a writer, an audit summary by a client's marketing director, an ad variant set by a media buyer. Each has a format they expect, decisions they need made for them, and things they will reject on sight.
An unstructured prompt communicates none of that. It gives the model a topic and lets it infer audience, depth, format, and voice — four guesses, each of which will land somewhere near the statistical middle of the internet. That is exactly what "generic AI content" means: the average of everything ever written on the subject.
The fix is not clever phrasing. It is supplying the same six things a competent PM supplies when briefing a person.
The six-block prompt structure
1. Role (one or two lines)
Set the perspective, not a personality. "You are a technical SEO consultant writing for a non-technical ecommerce marketing manager" is useful. "You are a world-class 10x growth ninja" is noise that consumes tokens and adds nothing.
2. Context and inputs (the biggest block)
This is where quality is won or lost. Paste the actual material: the top five ranking URLs and what they cover, the client's positioning statement, three quotes from the discovery call, the Search Console queries with impressions but no clicks, the previous quarter's performance numbers.
A rough benchmark from our own testing: a prompt with 300+ words of real client context produces first drafts that survive editing roughly three times as often as the same prompt with a two-sentence context block. Everything you paste is something the model no longer has to invent.
3. Task (one sentence, one deliverable)
State the single artifact you want. "Produce a content brief for a 1,400-word comparison article." Not "produce a brief and also suggest some titles and maybe an outline for a follow-up." Multi-task prompts get you three mediocre things instead of one good one.
4. Constraints (five to ten bullets)
Constraints are where agency voice actually lives. Include both positive and negative rules:
- Word count range with a hard ceiling, e.g. 1,300–1,500 words
- Reading level or sentence-length ceiling
- Banned phrases — "in today's digital landscape," "unlock," "game-changer," "delve"
- Banned claims — no invented statistics, no percentage improvements unless present in the inputs
- Required entities — competitor names, product names, service tiers to mention
- Anything legal or compliance-sensitive for regulated clients
Negative constraints are underused and disproportionately effective. A shared list of 20 banned phrases across your agency will do more for consistency than any tone-of-voice adjective.
5. Output format (be mechanical)
Specify structure like a schema, not a suggestion. "Return: H2 heading, two-sentence summary, then a table with columns Issue / Severity (High/Medium/Low) / Affected URLs / Recommended fix / Estimated effort in hours." If it feeds a document template or a project system, describe the exact fields.
Format specificity has a side benefit: it makes output comparable. When every audit summary uses the same five severity fields, you can diff them across months and across clients.
6. Quality bar and examples
End with one or two examples of approved past work, or a short rubric the model should self-check against. Two real examples beat a paragraph of description. If you cannot paste a full example, paste a 150-word excerpt and say "match this density of specifics."
Three worked examples
Content brief
Inputs: primary keyword and monthly volume, the top five SERP results with word counts and their H2s, three gaps none of them cover, the client's product angle, two internal links to include, and the target reader's job title. Task: a brief a freelance writer can execute without asking questions. Format: working title, search intent in one sentence, angle, H2 outline with a bullet of guidance under each, required entities, internal links, and a "do not say" list.
Agencies who structure this way typically cut brief production from about 40 minutes to 10–15 minutes of editing, and — more importantly — writers stop sending clarification emails. Pairing this with real demand data rather than guesswork matters; if your topic selection is off, a perfect brief still produces a wasted article, which is why it helps to align your content calendar with search volume data before any prompt gets written.
Technical audit summary for a client
Inputs: the raw crawl export (or a filtered subset), site size, current organic sessions, and the client's stated business priority for the quarter. Constraint: no more than six issues, ranked by revenue impact rather than crawler severity score. Format: one paragraph of plain-English summary a CMO can forward, then the table.
The constraint that changes everything here is "no more than six." Crawl tools return 400 issues. The deliverable's value is the prioritisation, and prioritisation only happens when you force scarcity into the prompt.
Monthly client report narrative
Inputs: this month's metrics versus last month and versus the same month last year, hours logged by workstream, deliverables shipped, and anything that slipped with the reason. Task: a 400-word narrative that explains what happened and what happens next. Constraint: name at least one thing that did not go well.
That last constraint is deliberate. Reports that only contain wins train clients to distrust reports. This works best when the numbers come from your actual project record instead of being retyped — the mechanics of pulling them straight from delivery data are covered in this guide to AI client summaries from project data.
Chain prompts for anything long
Single prompts degrade past roughly 800 words of output. Structure long deliverables as a chain:
- Outline pass — produce structure only. You review and correct it. This takes two minutes and saves rewriting a whole draft.
- Draft pass — write section by section against the approved outline, passing back the previous section for continuity.
- Critique pass — a separate prompt that scores the draft against your rubric and lists specific fixes. Then apply them.
The critique pass is the one most teams skip and the one with the highest return. Ask for "the five weakest sentences and why," not "improve this."
Store prompts where the work happens
A prompt that lives in one strategist's notes app is not an agency asset. Prompt libraries need the same treatment as templates: versioned, named by deliverable type, attached to the project phase that uses them, and updated when a client's voice guidelines change.
In practice this means your prompt for "monthly SEO report narrative — enterprise retainer" sits next to the task that produces it, with the client's context block pre-filled. Keeping briefs, prompts and delivery data in the same workspace is a large part of why teams evaluating tools end up looking at PM platforms built for agencies rather than generic task boards — the prompt is only as good as the client context sitting beside it, and PeakKR was designed around that adjacency.
Know what not to prompt
Prompt structure improves execution, not judgement
