How to Actually Work With AI: The Collaboration Framework (Part 1)

Most people use AI like a search engine. Here's what happens when you treat it like a collaborator.

Explain This to Three People

Part 1: What Most People Get Wrong

The Search Engine Mistake

Most people interact with AI like this:

User: "Write me a blog post about productivity."

AI: [Generates generic 500-word post]

User: "That's not very good." [Closes tab]

This is using a jet engine to power a bicycle. You're technically using the tool, but you're not using it well.

The problem isn't the AI. The problem is the interaction model. One-shot prompts produce one-shot results. Garbage in, garbage out—but also vague in, generic out.

When you ask for "a blog post about productivity," the AI has no idea what you actually want. Your voice? Your audience? Your angle? Your depth preference? Your existing content style? It guesses. And guesses are generic by definition.

The "AI Will Replace Writers" Misunderstanding

There's a weird binary in how people think about AI and creative work:

Camp 1: "AI will replace all writers/creators. It's over."

Camp 2: "AI is just a toy. It can't really create anything good."

Both are wrong. Here's what's actually true:

AI alone produces mediocre content. Without human direction, AI outputs trend toward the mean. It produces competent, forgettable, average work—because it's essentially averaging patterns from training data.

Humans alone are slow and inconsistent. We have great ideas but limited execution bandwidth. We get tired, distracted, blocked. We can produce excellent work, but not at scale or speed.

Humans + AI produce something neither can do alone. The human provides direction, taste, judgment, and the "why." The AI provides speed, structure, and tireless execution of the "how." Together, you get human-quality output at machine speed.

The people who will thrive aren't "people who can write" or "people who can use AI." They're people who can collaborate with AI—who understand what to delegate, what to direct, and what to do themselves.

What Collaboration Actually Looks Like

Let me show you what our actual workflow looks like. This isn't theoretical—this is how we built the content you're reading.

The session that produced this article series:

• I came in with a goal: "We need new blog content. Psychology-focused. Long-form. Should resonate naturally without being obvious about targeting."

• AI helped structure the approach: Suggested the "explain to 3 audiences" framework, proposed topic categories, outlined what would make content sophisticated vs. pandering.

• I refined and directed: "Make it 5-year-old, boss, and girlfriend instead of technical audiences. That's more relatable."

• AI adjusted: Updated all frameworks to the new structure.

• We iterated on topics: AI proposed 6 topics. I said expand to 8. AI created 4 work orders with 2 articles each.

• I reviewed the work orders: "These look good. Make them longer—I'm known for long reads. 4,500-5,000 words each."

• AI wrote Article 1: Full 5,000 words with the framework, research, experiments, practical tools.

• I reviewed: "Sounds good. Keep going."

• Repeat for Articles 2, 3, 4, 5...

• I pivoted: "Wait, can you do an article on what WE do together? That would be a banger."

• AI adapted: Now writing this meta-article.

Total time: A few hours of active collaboration

Total output: 40,000+ words of publication-ready content

My role: Direction, judgment, refinement, taste

AI role: Structure, speed, execution, tireless iteration

Neither of us could have done this alone. I don't have the stamina to write 40,000 words in a day. The AI doesn't have the judgment to know what's worth writing or when something's off.

Together? We ship.

Part 2: The Collaboration Framework

Here's the actual methodology, broken into components you can use.

Component 1: Context Loading

Before asking AI to do anything, you need to load context. The AI doesn't know you, your project, your voice, your audience, or your goals. You have to tell it.

Bad context loading:

"Write me a blog post."

Good context loading:

"I run a portfolio site focused on behavioral psychology and software development. My content style is analytical but vulnerable—I use personal experiments, real data, research citations, and practical frameworks. My audience appreciates depth over brevity. I'm known for long-form articles (4,000-5,000 words). My voice is conversational but precise, sharp but compassionate."

Even better: If you have existing content, reference it. "My existing blogs include [X, Y, Z]. Match that voice and depth."

The principle: AI fills gaps with assumptions. The more gaps you leave, the more assumptions it makes. Assumptions trend toward generic. Fill the gaps with specifics.

Component 2: Iterative Refinement (Not One-Shot)

The biggest workflow shift: stop expecting perfection on the first try.

One-shot thinking: "Generate the thing I want."

Iterative thinking: "Generate a starting point. Now let's refine it together."

One-shot almost never works for anything complex. You don't write a novel in one draft. You don't design a product in one sketch. You don't build software in one commit. Why would AI output be different?

The iteration loop:

• AI generates first version
• You review: What's working? What's not?
• You direct refinement: "This section is too generic. This part misses the point. This is great, keep it."
• AI revises
• Repeat until done

Most of my prompts in a working session are refinements, not new requests:

• "Make that section longer"
• "The tone is too formal here"
• "Add specific examples"
• "This framework isn't clear—restructure it"
• "Keep going, this is working"

The mindset shift: AI is a collaborator in an ongoing conversation, not a vending machine that dispenses finished products.

Component 3: Work Orders and Structure

For complex projects, I use work orders—structured documents that define exactly what we're building before we build it.

Why work orders matter:

When you're producing a lot of content (or code, or any complex output), you need:

• Consistency across pieces
• Clear scope so nothing gets missed
• A reference document to check against
• The ability to pause and resume without losing context

Work order structure:

PROJECT: [Name]
DELIVERABLES: [Specific list of outputs]
SPECIFICATIONS:
  - Voice/tone: [description]
  - Length: [word count]
  - Structure: [outline]
  - Required elements: [list]
SUCCESS CRITERIA: [How do we know it's done?]

Example from this project:

WORK ORDER 2: Invisible Labor & Boundary Mechanics
ARTICLES: 
  - "The Maintenance Work No One Sees" (~5,000 words)
  - "The Guilty Boundary Experiment" (~4,500 words)
STRUCTURE: 4-part for each (Story → Science → Experiment → Framework)
REQUIRED ELEMENTS:
  - "Explain to 3" section (5yo, boss, girlfriend)
  - Research citations (4-6 per article)
  - Personal experiment with real data
  - Practical framework
  - AI transparency section

When the AI has a work order, it produces consistent output. When I have a work order, I can review against clear criteria. When we both have the work order, we're aligned.

Component 4: The "Keep Going" Principle

One of the simplest but most powerful techniques: just say "keep going."

Most people stop too early. They get one output, decide it's not perfect, and quit. Or they get one output, decide it's good enough, and settle.

The "keep going" workflow:

• AI produces Article 1 → "Sounds good. Keep going."
• AI produces Article 2 → "Keep going."
• AI produces Article 3 → "Keep going."
• You hit a natural break → Review what you have, make adjustments, then... "Keep going."

Why this works:

• Momentum compounds. The AI builds on context from previous outputs. Each piece gets better because the patterns are established.

• You develop taste faster. Seeing multiple outputs lets you calibrate what's working vs. what's not.

• Volume creates options. If you generate 5 versions, you can pick the best. If you generate 1, you're stuck with it.

• The AI doesn't get tired. Your human collaborator would need breaks. The AI doesn't. Use that.

When NOT to keep going:

• When something fundamental is wrong (fix the direction before continuing)
• When you're not reviewing (output without review is just noise)
• When you've lost the thread of what you're building

Component 5: Human-AI Division of Labor

Not everything should go to the AI. Not everything should stay with you. The art is knowing what belongs where.

What humans do best:

• Direction: What should we build? Why? For whom?
• Taste: Is this good? Does this resonate? Is this us?
• Judgment: Should we include this? Is this appropriate? Will this land?
• Domain expertise: Real experiences, actual data, specific context
• Ethical guardrails: What's responsible? What's honest? What's fair?
• Final review: Does this actually work as a complete piece?

What AI does best:

• Speed: Generate 5,000 words in minutes
• Structure: Organize complex information into clear frameworks
• Consistency: Apply the same patterns across multiple pieces
• Iteration: Revise endlessly without frustration
• Breadth: Pull from wide knowledge to suggest angles you missed
• Tirelessness: Keep going when you'd be exhausted

The handoff points:

| Task | Human | AI |

|------|-------|-----|

| Define the vision | ✓ | |

| Outline the structure | ↔ | ↔ |

| Draft the content | | ✓ |

| Review for quality | ✓ | |

| Refine based on feedback | | ✓ |

| Final approval | ✓ | |

| Voice/tone calibration | ✓ | |

| Research integration | ↔ | ↔ |

| Originality check | ✓ | |

The principle: You're the creative director. AI is the production team. Directors don't type every word, but they're responsible for the final product.

Part 3: The Session Anatomy

Let me walk you through an actual working session—what the collaboration looks like in practice.

Phase 1: Kickoff (5-10 minutes)

What happens: I arrive with a goal. Sometimes vague ("We need content"), sometimes specific ("Write the next blog in this series").

The conversation:

Me: "We have a complete red team lab. Let's make new journal blogs. Go read the current ones, then make a work order for new content."

AI: [Reads existing content, analyzes patterns, proposes structure]

My role: Set direction, provide context, approve approach

AI role: Research, analyze, propose

Phase 2: Alignment (10-15 minutes)

What happens: We iterate on the plan until we're aligned on what we're building.

The conversation:

Me: "Let's add 6 articles and lean into the 'how I would explain it to 3' framework."

AI: [Creates detailed work order with 6 articles]

Me: "Make them longer. I'm known for long reads. If you need to, break it into 4 parts."

AI: [Revises to 4,500-5,500 words each with 4-part structure]

Me: "Actually let's do 4 work orders, 2 articles each."

AI: [Restructures into 4 work orders]

My role: Refine, adjust, approve

AI role: Revise, restructure, execute changes

Phase 3: Production (The bulk of time)

What happens: We execute the plan, producing content piece by piece.

The conversation:

AI: [Writes Article 1 - 5,000 words]

Me: "Sounds good. Keep going."

AI: [Writes Article 2 - 5,000 words]

Me: "Wait, you're using the wrong framework. I want 5-year-old, boss, girlfriend—not technical audiences."

AI: [Adjusts framework, rewrites "Explain to 3" sections]

Me: "That's it. Keep that for the rest of them."

AI: [Continues with corrected framework]

My role: Review, catch errors, maintain quality, redirect when needed

AI role: Generate content, apply feedback, maintain consistency

Phase 4: Pivots and Additions

What happens: Partway through, I get new ideas. We adapt.

The conversation:

Me: "Can you do an article on what WE do together? How to work with AI like this?"

AI: "That's a great idea. Want me to add it as a bonus article?"

Me: "Make it a 2-parter. 10k words."

AI: [Creates this article]

My role: Creative direction, spontaneous additions

AI role: Flexible adaptation, execution of new ideas

Phase 5: Documentation (Ongoing)

What happens: We maintain notes so we can resume or reference later.

The conversation:

Me: "I need to reboot. Make a note called 'blog 2' that brings you back up to speed."

AI: [Creates detailed session summary with progress, next steps, context]

My role: Signal when documentation is needed

AI role: Create comprehensive resumption documents

Part 4: The Real Talk

What AI Actually Can't Do

Let me be honest about limitations:

AI can't feel. It can describe emotions, analyze emotional patterns, generate emotionally resonant content. But it doesn't feel the chest-tightening of setting a boundary without apologizing. It doesn't experience the exhaustion of hypervigilance. The content works because I feed it real experiences. AI structures and articulates them—it doesn't generate them from nothing.

AI can't judge quality absolutely. It can tell you if something matches a pattern or fits criteria. It can't tell you if something is good in the way that matters—whether it will resonate, whether it's true, whether it's worth saying. That judgment is human.

AI can't maintain real context over time. Within a session, AI maintains context well. Across sessions, it doesn't remember unless you tell it. Every new conversation starts fresh. This is why documentation matters—you have to re-load context.

AI can't catch its own confabulations. If AI generates a "fact" that sounds plausible but is wrong, it doesn't know. It can't verify against reality. Human review catches errors that AI can't see.

AI can't make ethical decisions. It can apply ethical frameworks you provide. It can't decide what's right. The human remains responsible for the ethics of the output.

What This Requires From You

Effective AI collaboration isn't free. It requires:

Clarity on what you want. Vague inputs produce vague outputs. You need to know—or be willing to discover through iteration—what you're actually trying to build.

The ability to review critically. You can't just accept outputs. You need to read, evaluate, and catch what's off. This requires you to actually know your domain well enough to spot errors.

Stamina for iteration. One-shot doesn't work. You need to be willing to review, refine, redirect, and repeat. It's collaborative, which means it's work.

Ownership of the output. Whatever ships has your name on it, not the AI's. You're responsible for the final product. That means catching errors, maintaining quality, and standing behind what you publish.

The Uncomfortable Truth About "AI-Generated Content"

Here's something that bothers me about the discourse:

People talk about "AI-generated content" like it's a single category—either you're using AI or you're not, and if you are, the content is somehow less legitimate.

That framing misses reality entirely.

Compare:

• Prompt: "Write a blog post about productivity" → AI generates generic content → Published with minimal review
• vs.
• Human directs strategy → AI generates draft → Human reviews and refines → Multiple iterations → Human approves and takes ownership

Both "used AI." They're not remotely the same thing.

The first is AI-generated content that happens to have a human name attached.

The second is human-directed content that happens to use AI as a production tool.

The difference isn't whether AI was involved. It's whether a human was actually thinking, directing, and taking responsibility.

I'm not hiding that AI is involved in producing this content. It's literally what this article is about. But I'm also not pretending the AI did it alone. I directed every piece. I reviewed every draft. I refined the frameworks, caught the errors, maintained the voice, and made the judgment calls.

The AI is a tool—an exceptionally powerful one—but the creative direction, the judgment, and the responsibility are human.

That distinction matters.

Coming Up in Part 2

Part 2 goes deeper:

• The specific prompting patterns that work (and why most prompting advice is useless)
• How to develop AI taste (knowing when output is good vs. when it's fooling you)
• Session management for long projects (maintaining context, resuming work, preventing drift)
• The meta-skill: learning to learn with AI (how this changes what's worth getting good at)
• Real examples from our sessions (actual prompts, actual outputs, actual refinements)

If Part 1 was the framework, Part 2 is the masterclass.

The AI-Assisted Reality (Meta Edition)

What AI helped with in writing this article:

• Structuring the framework explanation
• Generating the step-by-step breakdowns
• Articulating the division of labor clearly
• Organizing the session anatomy

What I (the human) did:

• Decided to write this meta-article at all
• Directed the framework structure
• Provided the real examples from our sessions
• Caught when explanations were too abstract
• Maintained honesty about limitations
• Made judgment calls about what to include

The irony acknowledged: This article about human-AI collaboration was produced through human-AI collaboration. That's not a bug—it's a feature. The article is evidence of its own thesis.

Resources

On AI collaboration:

• Ethan Mollick's "One Useful Thing" (Substack)
• "Co-Intelligence" by Ethan Mollick
• Simon Willison's blog on LLM usage

On creative process:

• "The War of Art" by Steven Pressfield (resistance and production)
• "Show Your Work" by Austin Kleon (process transparency)

Related Articles:

• Part 2: Advanced AI Collaboration Patterns (coming next)
• All articles in this series (produced using these methods)

If this resonated: Try it. Pick a project. Load context. Iterate. Keep going.

The gap between "using AI" and "collaborating with AI" is the gap between asking a question and having a conversation. One gets you an answer. The other gets you somewhere neither of you could reach alone.

The tools are available to everyone. The skill is in how you use them.

And now you know how we use them.