AI Prompting Framework

Stop Chatting.
Start Directing.

Four phases to shift from open-ended AI conversations to high-leverage strategic output. Designed for design leaders who think in systems but need to ship in hours, not days.

⚡

The Core Shift

You're using AI as a conversation partner when you should be using it as an execution engine with strategic checkpoints. The conversation is where your time disappears. The contract is where your leverage lives.

Tell Claude what done looks like before you start working together.
✕ What you probably do now
"Help me think through the priority framework for the next phase of our design strategy..."
✓ The contract version
"I need a priority framework for [project]'s next phase. Here's the context: [2-3 sentences]. Produce a ranked list of 5-7 priorities, each with: the problem it addresses, why it's ranked here, and one key dependency. Format as a table I can paste into my strategy doc."
Rules
Name the artifact. "Produce a ranked list," "Write a brief," "Create an evaluation rubric." If you can't name what you want, you're not ready to prompt.
Specify the format. Table, numbered list, narrative paragraphs, slide outline. This alone cuts 2-3 rounds of back-and-forth.
Give the constraint. "5-7 priorities," "under 500 words," "3 options max." Constraints produce better output than open-ended asks.
State the destination. "I'll paste this into a Notion doc for my team," "This goes into a stakeholder presentation." Claude adapts tone and depth accordingly.
Everything Claude needs to do great work should arrive in the first message — not drip-fed across a conversation.
✕ The drip-feed pattern
Message 1: "Help me with our strategy doc" Message 5: "Oh, I should mention the timeline is Q3" Message 12: "Actually, the real constraint is legal approval" Message 18: "The audience is our VP, not the team"
✓ The front-loaded version
"Context: I'm leading [project] for [team]. We're in [phase]. The audience is [who]. The constraints are [timeline, dependencies, blockers]. Here's the current state: [paste existing doc/notes]. I need: [specific artifact with format]."
Rules
Use your PARA system. You've already built this infrastructure. Before prompting, pull the relevant project context from your second brain and paste it in. This is exactly what it's for.
Paste, don't summarize. If you have existing docs, drop them in. Claude can synthesize raw material faster than you can summarize it, and nothing gets lost in translation.
Name the audience and stakes. "This is for our VP review" produces fundamentally different output than "help me think through this." Two words of context save ten messages of correction.
State what you've already decided. "I've already committed to X and Y — don't relitigate those. Focus on Z." Prevents Claude from reopening settled questions.
After the first output, you're an editor — not a conversation partner. Redirect, refine, or approve. Don't explore.
✕ The exploration trap
Claude gives output → "Interesting, what about X?" → Claude expands → "Hmm, and what about Y?" → Claude adds more → 45 minutes later you have a rich discussion but no usable artifact
✓ The checkpoint version
Claude gives output → You evaluate against your contract → "Points 1-3 are strong. Point 4 misses [specific thing] — revise to account for [constraint]. Point 5: replace with [your idea]. Then give me the final version."
Rules
React to the output, not the topic. Your job after message 1 is editorial: what's right, what's wrong, what's missing. Stay concrete.
Use directive language. "Revise point 4," "Cut section 2," "Add a row for timeline." Not "what do you think about..." or "could you maybe..."
Batch your feedback. Give all corrections in one message, not one per round. Each round-trip costs you context and time.
Know when to stop. If it's 80% right, take it and polish manually. Chasing the last 20% through prompting often takes longer than fixing it yourself.
Not every task is a conversation. Choose the right interaction mode before you start.
✕ Using the same approach for everything
Long open conversation for a quick Slack message. Long open conversation for a strategy framework. Long open conversation for a competitive analysis. Long open conversation for everything.
✓ Mode-matched approaches
Quick tasks → single-shot prompt, no conversation Drafting → contract + 2-3 checkpoints Synthesis → paste sources + specific output request Exploration → time-boxed to 10 min, then switch to execution mode
Rules
Single-shot mode (Slack messages, emails, meeting agendas): One prompt, one output. Takes 1-2 minutes. Don't open a conversation for something that needs a single generation.
Draft mode (briefs, frameworks, strategy docs): Contract → first output → 2-3 checkpoints → done. Should take 15-30 minutes, not hours.
Synthesis mode (research summaries, competitive analysis): Paste all inputs → request specific structure → refine once → done. Let Claude do the heavy lifting of organization.
Exploration mode (genuinely new territory where you need to think): This is valid — but time-box it. Set a timer for 10-15 minutes. When it rings, switch to one of the modes above and produce something concrete.
Your Strategy Doc Test
That priority framework that took an entire day? Here's what it looks like with this framework applied:
5 minPull project context from PARA. Write the contract: "Produce a ranked priority framework with 5-7 items, each with problem statement, rationale, and dependencies. Format as table."
2 minClaude generates first draft with your full project context already loaded.
10 minYou review. Apply your strategic judgment: "Move item 3 to #1 because [reason]. Item 6 is wrong — replace with [your insight]. Add a column for effort estimate."
2 minClaude revises. You do a final pass.
10 minPolish manually. Paste into your strategy doc. Add any nuance that only you can add.
Total: ~30 minutes instead of a full day