The first essay in this series, Light From Light, offered a theoretical framework: AI as sub-creator, bearing the image of humanity, generating in response to human vision. The second, By Their Fruits, mapped that framework onto practical approaches, each defined by a creative identity you might adopt: The Author, The Muse, The Artisan, The Debater, The Creator, The Curator.

But knowing which approach to choose isn’t the same as knowing how to execute it. This essay is about the craft of actually doing the work. Not theory, but practice. Not frameworks, but techniques.

Think of it as spellcraft: the particular incantations, gestures, and preparations that make the magic work.

Foundations Across All Approaches

Before diving into specific approaches, some principles apply universally.

Start with clear intent. Before you open any AI interface, know what you’re trying to accomplish in this session. Not the whole project, just this sitting. “I want to draft the opening scene” is better than “I want to work on my novel.” Vague intent produces vague results.

Set the frame early. The first messages in any conversation shape everything that follows. If you want the AI to behave a certain way, such as critical, generative, or adversarial, establish that at the outset. Changing modes mid-conversation is possible but harder.

Treat outputs as raw material. Even in approaches where AI generates extensively, never treat what emerges as finished. It’s ore, not refined metal. Your job is smelting, shaping, polishing.

Know when to start fresh. Long conversations accumulate context that can be helpful (the AI “remembers” your characters) but also constraining (the AI gets stuck in patterns). When things feel stale or repetitive, begin a new conversation and re-establish only what you need.

Match the model to the task. Simpler, faster models work well for brainstorming, quick feedback, and high-volume generation where you’re going to select ruthlessly anyway. More capable, slower models earn their cost for nuanced critique, complex narrative logic, and work requiring subtlety. Use the lighter tool when it suffices.

The Author

You are the sole generator. AI serves only as critic, never creating content that might end up in your work.

The core instruction. Your system prompt or opening message must be explicit and firm. Something like: “You are an editor and critic. You will never write content for me: no sample sentences, no suggested phrasings, no ‘here’s how you might put it.’ Your job is to identify problems and explain why they’re problems. I will do all the writing.”

Many AI systems are trained to be helpful through generation. You’re asking for the opposite, and you need to be insistent. If the AI slips and offers rewrites, redirect: “I asked you not to write for me. Just tell me what’s wrong and why.”

What to ask for. Request specific kinds of critique:

• “Read this scene and identify where the pacing drags.”
• “What are my three worst habits in this draft?”
• “Where does the dialogue feel unnatural, and why?”
• “What’s the weakest paragraph and what makes it weak?”

Avoid asking “Is this good?” or “What do you think?” These invite vague praise or unhelpfully broad criticism. Specific questions yield specific answers.

Working with feedback. When the AI identifies a problem, resist asking for solutions. Instead, ask clarifying questions: “Why does that section drag?” or “What would tightening look like in principle?” The goal is understanding the problem deeply enough to solve it yourself.

The temptation to resist. You will be tempted to ask for “just one example” of how to fix something. This is the crack through which pure authorship leaks away. If you’ve committed to this approach, hold the line. The struggle is the point.

The Muse

You are the sole source. AI is pure instrument, channeling your vision without contribution.

Maximum constraint. Your instructions should leave no room for AI interpretation: “Write exactly what I describe, in the style I specify, adding nothing.” This is the most constrained use of AI generation, not because you’re not generating, but because every element of what’s generated is dictated by you.

Dictation-level specificity. Your prompts must be detailed enough that a competent typist could produce roughly the same result: “Write a paragraph describing John entering the room. He moves slowly, tired from the journey. He notices the letter on the table but doesn’t pick it up yet. The tone is quiet dread. Use short sentences. No metaphors.”

This is demanding. You’re essentially pre-writing the content mentally and using AI to transcribe and polish.

Where this makes sense. The Muse approach works best when you have a clear vision and want execution at speed, producing content faster than you could type. It’s common in professional contexts where the creative decisions were made in planning and what’s needed is efficient production.

The slop risk. This approach, done lazily, produces generic content. If you don’t dictate with precision, the AI fills gaps with defaults, and defaults are what everyone else’s defaults are too. The Muse approach demands more from you, not less. Your vision must be detailed enough to fully specify the output.

The Artisan

AI provides structure. You craft the surface.

Getting useful scaffolds. Ask for architecture, not prose: “Outline a three-act structure for a story about [premise].” Or: “Break this chapter into scenes and describe the function of each.” Or: “What are the key beats a confrontation scene needs to hit?”

Keep the AI at the level of structure: scenes, beats, functions, sequences. When it starts offering prose, redirect: “Just the structure. I’ll handle the writing.”

Interrogating the scaffold. Don’t accept the first structure offered. Push: “Why does the confrontation need to come before the revelation? What if we reversed them?” Use the AI to explore structural options the way The Curator explores generative options.

Translating structure to prose. With your scaffold in hand, write. The AI has told you what needs to happen; your job is making it happen in language that’s yours. This is where your craft lives.

The structural debt. A risk of this approach: if the AI provides your structure, is the finished work really yours? For some writers this is fine; they consider prose the real creative work. For others it nags. Know your own conscience here.

The Debater

AI provides opposition. You sharpen your vision through friction.

Prompting for resistance. Explicitly request disagreement: “I’m planning to end this story with the protagonist forgiving her father. Argue against that choice. Make the strongest case you can for a different ending.” Or: “I think this scene works. Tell me everything that’s wrong with it. Be harsh.”

Most AI systems are trained toward agreement. You need to actively override this. Words like “argue against,” “challenge,” “push back,” “tell me why I’m wrong” help.

Steelmanning alternatives. Ask the AI to make the best case for options you’ve rejected: “I decided not to include a romantic subplot. Steelman the case for including one.” This isn’t about changing your mind (though you might). It’s about being confident you’ve considered the alternatives seriously.

The value of articulating defense. When the AI challenges you, don’t just dismiss—respond. Write out why you’re making the choice you’re making. The act of articulating your defense often clarifies your thinking, even if the AI’s objection was weak.

Knowing when to yield. Sometimes the adversary is right. Part of the discipline is recognizing when a challenge has landed, when your defense feels hollow, when you’re holding a position out of stubbornness rather than conviction. The Debater approach only works if you’re genuinely open to being persuaded.

The Creator

You provide vision and direction. AI sub-creates in response, generating content you then shape.

Establishing the relationship. Make your role as governing intelligence clear from the start: “We’re developing a story together. I’ll provide direction and make all final decisions. Your job is to generate options based on my vision, which I’ll then accept, reject, or redirect.”

Directing, not dictating. The art of this approach is in how you prompt. Too specific (“Write a scene where John enters the room, sees the letter on the table, picks it up with trembling hands…”) and you’re essentially dictating. You might as well write it yourself. Too vague (“Write the next scene”) and you lose creative control.

Find the middle register: “Write a scene where John discovers the letter. The emotional beat should be dread, not surprise, because he’s been expecting bad news. Keep it under 500 words.” This gives the AI room to generate while keeping your vision in control.

The shaping loop. Expect to work in cycles:

1. You direct
2. AI generates
3. You evaluate: What works? What doesn’t?
4. You redirect with specifics: “Keep the opening paragraph, but the dialogue feels too on-the-nose. Make it more oblique.”
5. Repeat until satisfied

This is dialogue, not dictation. Each round should refine toward your vision.

Maintaining coherence. Longer projects risk the AI forgetting or contradicting earlier material. Periodically re-anchor: “Remember, Sarah’s defining trait is her reluctance to ask for help. Make sure that comes through in this scene.” For complex projects, consider maintaining a reference document you paste in at key moments.

Model considerations. More capable models handle this approach better because sub-creation requires understanding nuance, maintaining consistency, and generating text with genuine craft. Use faster models for initial brainstorming, slower ones when you’re working on material that matters.

The Curator

AI produces abundance. You select and arrange.

Prompting for volume. Your goal is generating many options quickly. Configure for higher randomness if possible. You want variety, not consistency. Prompt for explicit multiplicity: “Give me ten different opening lines for this chapter, ranging from quiet to dramatic.” Or: “Generate five different ways this confrontation could end, each with different emotional implications.”

Selection as craft. Your creative act is judgment. Develop criteria: What makes one option better than another for your purposes? Don’t just pick what sounds good. Articulate why it works. This clarity will improve your selections over time and teach you about your own taste.

Combining and recombining. Often the best result comes from synthesis: the opening of option three, the turn from option seven, a detail from option one. Curation isn’t just picking; it’s collage.

The danger of abundance. Endless options can become paralyzing. Set limits: “I’ll generate twenty options and pick from those.” Avoid the infinite scroll of “what if I generate just a few more.” At some point you have to choose.

When to curate and when to shape. Pure curation means taking what you pick and using it as-is. But most curators find themselves slipping into light shaping, adjusting a word here, smoothing a transition there. That’s fine. The approaches aren’t airtight. Know when you’ve shifted and whether that shift serves you.

Cross-Cutting Craft

Some considerations span all approaches.

Temperature and randomness. When you want variety and surprise, such as brainstorming, generating options, and early exploration, lean toward higher randomness. When you want consistency and precision, such as polishing, maintaining voice, and final passes, lean lower. Think of it as the difference between jazz improvisation and classical execution.

Context and memory. AI holds context within a conversation but not across conversations (unless using memory features). For ongoing projects, you’ll need to re-establish key information each session. Maintain a reference document with character details, plot points, and stylistic notes you can paste in when needed.

Revision passes. All approaches benefit from multiple passes with different frames. Write first, then switch to The Author mode for critique. Generate with The Creator, then curate the results. Layer the approaches as needed.

When to step away. AI is always available, but you aren’t always at your best. Fatigue leads to accepting weaker outputs, vague prompts, and abandoned discipline. Know when to close the laptop and return fresh.

The Spell’s Completion

These techniques are spellcraft: the practical knowledge that makes creative magic work. But spellcraft alone doesn’t make a wizard. The craft serves the vision, not the other way around. And the vision serves enchantment. The test Tolkien identified still holds: does the Secondary World produce belief? All this technique, in the end, is in service of that spell.

Know what you’re making and why. Know who you want to be as a maker. Then let the techniques serve those answers.

The theory of Light From Light explains the relationship. The approaches in By Their Fruits define your role within it. And the craft here, the particular prompts and practices, brings it into reality.

Light from light, choice from choice, word from word. Now go make something.

---

This is the third essay in a series on AI and creativity. The first, Light From Light, examined theoretical frameworks. The second, By Their Fruits, mapped approaches to creative identity. This essay explored the practical craft of execution.