The 70% Draft Problem: How to Make Claude Do the Thinking You Keep Postponing

Last Tuesday, a director of revenue operations at a 200-person SaaS company told me something that stuck. She had pasted a 14-page partner agreement into Claude, asked it to 'summarize the key points,' and gotten back a tidy five-paragraph overview. Clean grammar. Nice headers. Completely useless. 'It told me what was in the document,' she said. 'It didn't tell me what to worry about.' She spent another 90 minutes reading the contract herself, which was exactly what she'd been trying to avoid.

Her problem wasn't the tool. Claude can process that agreement and return something genuinely useful — specific obligations with deadlines, liability clauses that deviate from your standard terms, auto-renewal windows you'd miss if you skimmed. The gap was in the ask. And that gap is where most business teams lose hours every week without realizing it.

This piece isn't a feature walkthrough or a product comparison. It's about a specific failure mode I see in almost every team trying to use Claude for real work: they get output that looks like a draft but functions like a placeholder. And it introduces a diagnostic I call the 70% Draft Test — a framework for determining, before you hit Enter, whether your prompt will produce something worth editing or something you'll quietly rewrite from scratch.

Why Most Claude Output Lands in the Trash

I've reviewed over 200 AI tools for business use. The pattern is the same across all of them, but it's especially visible with Claude because the output reads so well. Claude's sentence-level quality is high enough that the result feels done. You read it, nod, and then realize ten minutes later that it didn't actually advance your thinking. It summarized without prioritizing. It drafted without deciding. It organized without opinionating.

This happens because most business users prompt Claude the way they'd query a search engine: 'Summarize this report.' 'Draft a memo about Q2 results.' 'Write a vendor onboarding checklist.' These are task descriptions, not briefs. They tell Claude what to produce but not what problem you're solving, who will read it, or what decision the output needs to support. The result is text that's structurally correct and strategically empty.

Claude is unusually good at following specific constraints — format, tone, length, structure. Among the major AI assistants, it's the one I trust most to respect an instruction like 'three sections, no bullets, 400 words, written for a CFO who is skeptical of this initiative.' But that strength only activates if you actually provide constraints. Without them, you get the default: competent, generic, disposable.

The 70% Draft Test

Here's the framework. Before you send any prompt to Claude for a business task, ask yourself three questions. If you can answer yes to all three, you'll get a draft that's 70% or more of the way to finished — meaning your job is editing and judgment, not rewriting. If you can't, you'll get filler.

• Does your prompt name a specific reader and their disposition? 'My VP of Sales, who is going to push back on the pipeline numbers' is a brief. 'Leadership' is not.
• Does your prompt state what decision or action the output should make easier? 'Help the reader decide whether to renew this vendor contract' gives Claude a target. 'Summarize the contract' gives it a genre.
• Does your prompt include at least one constraint that forces a tradeoff? 'One page max' or 'focus only on financial risk, ignore operational details' or 'argue against the proposal before arguing for it.' Constraints force Claude to prioritize, which is where the value lives.

That's it. Three yes answers before you press Enter. I've been running this test across my own workflow for six weeks, and the difference in output quality is not subtle. The drafts I get back now require 10–15 minutes of editing. Before, they required 45 minutes or a full restart.

What This Looks Like on Actual Business Tasks

The framework is abstract until you see it in action. Here are four workflows where I've applied the 70% Draft Test, with the specific prompt structures that passed.

Contract and document review: Claude's context window — up to roughly 200,000 tokens on Claude 3.5 Sonnet — means you can paste an 80-page agreement in one shot. But 'summarize this' fails the test. What passes: 'I'm evaluating whether to renew this vendor agreement. I need a one-page brief for my CFO that covers: (1) any liability clauses that differ from a standard SaaS vendor agreement, (2) auto-renewal deadlines and notice periods, (3) pricing escalation terms. Flag anything that gives them unusual leverage over us.' That prompt names a reader (CFO), states a decision (whether to renew), and includes a constraint (one page, three specific areas). The output I got back was usable after one editing pass.

Meeting follow-up: Paste a raw transcript from Otter.ai, Fireflies, or Teams. Failing prompt: 'Give me action items.' Passing prompt: 'This is a transcript from a cross-functional planning meeting. Produce a follow-up email I can send to the seven attendees. Include: (1) decisions that were finalized, with the owner's name, (2) open items where no decision was reached, sorted by urgency, (3) any commitments with a deadline mentioned. Keep it under 300 words. Tone: direct, not corporate.' The difference in output is the difference between something you forward immediately and something you sigh at.

Strategic writing: Investor updates, board memos, internal announcements. Claude handles tone better than any other tool I've tested — but only when you specify it. Passing prompt structure: 'I'm the CEO of a Series B SaaS company writing a quarterly investor update. Our ARR grew 18% but we missed our hiring plan by four heads in engineering. The audience is seven investors who are supportive but will notice if I bury bad news. Write a 500-word update that leads with momentum, addresses the hiring miss directly in the second paragraph, and closes with our plan to fix it in Q3. Tone: confident but honest.' I've used this pattern for three investor updates now. Each one required less than 15 minutes of revision.

Competitive analysis: Feed Claude a competitor's earnings call transcript or product changelog. Failing prompt: 'Analyze this.' Passing prompt: 'I'm a product marketing manager at a mid-market CRM company. This is our top competitor's Q2 earnings call transcript. I need a one-page internal brief that answers: (1) Where are they investing that we're not? (2) What did they de-emphasize that we should pay attention to? (3) What should our sales team know before next week's pipeline reviews?' Pair this with data from Crunchbase or SimilarWeb for numbers Claude can't access, and the brief starts to feel like something a paid analyst would deliver.

The Push-Back Prompt: Where Claude Earns Its Keep

The highest-value use of Claude I've found isn't drafting — it's stress-testing. Most business decisions die not because the idea was bad, but because no one in the room pressure-tested it before it shipped. Claude is very good at playing the skeptic if you ask it directly.

Try this: after Claude produces a draft strategy, memo, or proposal, follow up with 'Now tell me the three strongest reasons this plan will fail. Be specific and assume the reader is smart.' Or: 'A board member who used to run operations at a Fortune 500 company is going to read this. What will they find weak?' The responses are not always right, but they reliably surface angles I hadn't considered. That alone has saved me from sending at least two memos I would have regretted.

Choosing a Plan and Setting Up for Monday

Pricing is simple. Anthropic's free tier lets you test the framework on a few tasks. The Pro plan at $20/month is where individual operators should start — it gives access to Claude 3.5 Sonnet, higher usage limits, and the full context window. The Team plan at $30/user/month adds workspace sharing (useful if your team wants to build on each other's prompt patterns), priority access during peak hours, and a default opt-out of model training — which matters the moment you paste in a contract, financial data, or anything with a customer's name in it.

For teams already running Slack, Notion, or Zapier, Claude connects through those integrations. Companies on AWS can access Claude through Amazon Bedrock without standing up separate infrastructure.

Your Monday Playbook

Here's the specific sequence to run this week. Total investment: one task, 20 minutes.

• Pick one task you've been postponing. The contract review, the meeting recap, the quarterly update draft, the competitive brief. Something that normally takes 60–120 minutes of focused writing or analysis.
• Run it through the 70% Draft Test before you type anything into Claude. Name your reader and their disposition. State the decision the output should support. Add at least one constraint that forces prioritization.
• Paste your source material and your prompt into Claude. Time yourself.
• Edit the output. Don't rewrite it — edit it. If you find yourself rewriting more than 30% of it, your prompt failed the test. Go back, tighten the brief, and re-run.
• Log the time. First task, source-to-sent. Compare it to how long that task usually takes you. In my experience and in conversations with operators who've adopted this pattern, the number lands between 40–60% time saved on the first attempt.

Where Claude Doesn't Belong

Claude is not a replacement for every tool in your stack, and pretending otherwise will slow you down. For image generation, Midjourney and DALL·E are purpose-built. For spreadsheet-heavy financial modeling, Microsoft Copilot's Excel integration is tighter. For real-time web research with source citations, Perplexity is the better fit. Claude's strength is structured reasoning, long-form analysis, and writing where tone and precision matter. Treat it as your first-pass thinker for work that requires judgment, and route everything else to specialists.

The Actual Competitive Advantage

The 70% Draft Test isn't really about Claude. It works with any capable language model. But Claude rewards it more consistently than alternatives I've tested because it follows constraints tightly and doesn't hallucinate structure you didn't ask for. The advantage isn't having access to the tool — everyone has that. The advantage is knowing how to brief it like you'd brief a sharp new hire: here's who you're writing for, here's what they need to decide, and here's the box you're working inside. Teams that internalize this framework spend less time reformatting AI output and more time on the judgment calls that actually move their business forward. That's not a feature. That's a workflow change. And it compounds every week you practice it.