Prompt Engineering with Claude: How to Build Mega-Prompts for Content Pipelines

Anyone can ask Claude to write a blog post. That's not prompt engineering — that's prompting. Real prompt engineering is building a 4,000-token system prompt that produces consistent, brand-accurate output across 200 articles a month without a human editor catching every draft.

This guide walks through the exact mega-prompt structure we use to power content pipelines for IntellectVA clients running 50+ pieces per week. Tested. Refined. Battle-scarred.

Why short prompts fail at scale

Short prompts ("write a 1,500-word blog post about X in our voice") work for one-off drafts. They fail at scale because Claude has no persistent memory of:

• Your brand voice rules (formality, sentence length, taboo words).
• Your factual constraints (claims you can't make, statistics you must cite).
• Your structural patterns (intro length, H2 cadence, callout usage).
• Your conversion rules (where CTAs go, how product mentions appear).

Every short prompt forces Claude to guess. Guesses produce inconsistency. Inconsistency produces editor fatigue. Editor fatigue produces missed deadlines.

A mega-prompt eliminates guessing. You front-load every rule, every example, every constraint. Claude follows the system. You get consistent output.

The 7-section mega-prompt architecture

Every mega-prompt we ship follows the same skeleton. Section order matters — Claude weights earlier instructions more heavily.

# SECTION ORDER (~3,500-4,500 tokens total)

1. ROLE           // Who Claude is in this context
2. OBJECTIVE      // What output you want, in one paragraph
3. CONTEXT        // Brand background, audience, current state
4. CONSTRAINTS    // Rules: must-do, must-not-do, must-include
5. VOICE          // Tone, sentence length, taboo words, examples
6. STRUCTURE      // Section pattern, headings, CTA placement
7. OUTPUT FORMAT  // Exact markdown shape, frontmatter, metadata

Section 1: ROLE

Set the persona explicitly. Not "you are a writer." Be surgical: "You are a senior B2B SaaS content editor with 10 years of experience writing for technical founders. You write with the conviction of someone who has shipped 500+ articles and seen what converts."

This anchors Claude's vocabulary and reasoning depth before any task-specific instructions land.

Section 2: OBJECTIVE

One paragraph. What is the output? Who reads it? What action should it produce? Don't list 12 goals — pick one primary, one secondary. Claude optimizes for what you front-load.

Section 3: CONTEXT

This is where most mega-prompts compound value. Include:

• Brand description in 2-3 sentences.
• Audience profile (job title, pain points, awareness stage).
• 2-3 example URLs of competitor content you respect.
• Current content cluster context (what else has been published on this topic).

Section 4: CONSTRAINTS

Hard rules. Format them as imperatives:

• MUST include 1 primary keyword in H1, used once.
• MUST target 0.8-1.5% keyword density.
• MUST NOT use words: "delve", "elevate", "leverage", "seamlessly".
• MUST NOT claim quantitative results without a citation.
• MUST include at least one FAQ block, one comparison table, one numbered list.

Section 5: VOICE

Voice is the section that separates amateur prompts from production systems. Give Claude:

• Tone descriptors in pairs: "confident not arrogant, direct not blunt, technical not jargon-heavy."
• Sentence length rules: "Vary between 8-22 words. No sentences over 25 words."
• Voice samples: paste 2-3 paragraphs of approved past content. Claude pattern-matches voice from examples better than from descriptions.
• Anti-patterns: paste 1-2 paragraphs of writing that "sounds AI" and label them rejected.

Section 6: STRUCTURE

Lock the article skeleton:

• Intro: 2-3 short paragraphs, answer the query in first 2 sentences (AI Overview optimization).
• H2 cadence: every 200-300 words.
• Mid-article callout: 1 minimum.
• Comparison table or FAQ: 1 minimum.
• Conclusion: 2 paragraphs with conversion CTA.

Section 7: OUTPUT FORMAT

Specify the exact markdown shape. Frontmatter fields. Heading levels. Even the trailing newline if it matters to your build pipeline.

How token windows scale content pipelines

Claude 4.6 supports a 200K token context window. That's ~150,000 words. A 4,500-token mega-prompt uses 2% of available context. The remaining 98% is yours for:

1. Brand bible: paste your full style guide. Claude references it during generation.
2. Voice corpus: paste 10-15 approved past articles. Claude pattern-matches voice with high fidelity.
3. Topic cluster: paste related published content so Claude maintains internal linking opportunities.
4. Source material: research notes, interview transcripts, internal data — Claude weaves it in.

This is the unlock. You're not asking Claude to write blind. You're loading the system with everything a senior editor would carry in their head.

Prompt caching: the cost multiplier

Anthropic's prompt caching dropped the economics of mega-prompts by 10x. Cached input tokens cost 10% of standard input tokens. Cache TTL is 5 minutes by default.

Practical implication: your 4,500-token mega-prompt costs full price on the first request, then 10% on every subsequent request within 5 minutes. For a content pipeline pushing 20 articles in a batch, you pay full price once.

Architecture matters: load the cacheable content (mega-prompt + brand bible + voice corpus) at the top of the system message. Put the article-specific brief at the bottom. Cache the top, vary the bottom.

Common mega-prompt failures

• Conflicting constraints: "be conversational" + "be authoritative" + "be technical" produces mush. Pick two.
• Too many examples: more than 5 voice samples and Claude starts averaging them into a vanilla midpoint.
• Vague taboos: "don't use AI clichés" is useless. List the specific words.
• Buried output format: if structure rules appear in section 3, Claude weights them less than the structure section. Put format last.
• No self-check step: without "restate before writing," Claude drifts on 30% of generations.

What this unlocks for content pipelines

A working mega-prompt removes the editor-as-bottleneck problem. Articles ship 70-80% production-ready. Human editors review for fact-checking and brand-fit polish, not structural rewrites.

For IntellectVA clients running scaled content operations, this is the difference between publishing 10 articles a month and 80. Same team. Same budget. Same brand voice.

Next steps

If you're running content at volume and your team is drowning in editing cycles, the mega-prompt approach is probably your unlock. The setup work is real — expect 2-3 weeks to build, test, and refine a production prompt. The output gains compound for years.

Schedule a discovery audit. We'll review your current content output, audit your brand voice consistency, and prototype a mega-prompt against 5 sample briefs so you can see the lift before committing.