The Tighter the Leash, the Better the Output
Why less freedom produces better AI writing.
The Architecture Series · 8 minute read · By Ed | Founder & CEO, DealiOS
Here's a counterintuitive finding from years of building production AI: the less freedom you give it, the better the output.
Not a little better. Dramatically better. More creative, more consistent, more usable, and less likely to produce something that makes you close your laptop and question your career choices.
The instinct is wrong. "Give the AI more freedom and it'll be more creative." No. Give the AI more freedom and it'll be more random. Random is not creative. Random is the kid in art class who threw paint at the wall and called it a statement.
Constraints aren't the enemy of creativity. They're the scaffolding that makes creativity useful.
Give the AI more freedom and it'll be more random. Random is not creative.
The Freedom Paradox
Ask an AI to "create a presentation about Q3 results." No template. No structure. No constraints on layout, slide count, content density, or visual treatment. Run it ten times.
You'll get ten different structures. Different slide counts. Different levels of detail. Three will be unusable. Two will be mediocre. Maybe one or two will be good. But you won't know which one you're getting until you open it. And neither will your sales rep. Or the next rep. Or the one after that.
Consistency across runs, across users, and across deals is the actual test of a production AI system. Not "can it produce one good output?" but "can it produce reliably good output, every time, for every user, regardless of what they asked for?" That's a fundamentally different engineering problem. And freedom is its enemy.
This isn't a new observation. Acar, Tarakci, and van Knippenberg published a cross-disciplinary review in the Journal of Management in 2019 examining how constraints affect creativity and innovation across six academic fields [1]. Their finding: the relationship between constraints and creative output follows an inverted U-shape. Too few constraints produce average output because the search space is too large for focused creative work. Too many constraints produce nothing useful because there's no room to maneuver. The productive zone is in the middle: enough freedom to be creative, enough structure to be excellent.
Damadzic and colleagues confirmed this empirically in a 2022 meta-analysis published in the Journal of Organizational Behavior, analyzing studies across multiple constraint types [2]. The pattern held across domains. Moderate constraints consistently outperformed both unconstrained and over-constrained conditions.
Unlimited creative freedom sounds great until you see the results. That's true for AI and for the guy at the open mic who says "I'm just gonna do whatever feels right." No sir. What feels right is apparently a 20-minute story about your roommate's cat that goes nowhere. The sonnet has 14 lines and a rhyme scheme. The haiku has 17 syllables. The three-minute comedy set has three minutes. These constraints don't limit art. They produce it.
The Slot Machine (The Good Kind)
The architectural fix is structural, not instructional. Instead of asking the AI to generate arbitrary output and hoping for the best, you define proven structures with typed content slots, then let the AI fill the slots.
Here's what that looks like in practice. A presentation template registry with 20+ proven layouts. Each layout has named content slots: a headline (max 80 characters), a body (max 300 words), a supporting metric, a visual directive. Each slot has a type (text, number, image reference), a character limit, a word count target, and validation rules. The AI selects which layout fits the content, then fills each slot within its constraints.
The AI never generates arbitrary HTML. It never decides to put 14 bullet points on a slide because that's technically possible. It never outputs a wall of text with no visual hierarchy because nobody told it not to. The structure prevents it.
Here's the rule that changed everything: "Fill every slot to at least 70% of maxLength."
That single constraint improved output quality more than any prompt engineering technique we tested. Before the rule, the AI would fill some slots fully and leave others sparse, producing slides that felt lopsided. After the rule, every slide had consistent information density. The AI had to find content for every slot, which forced it to think harder about what belonged where. The constraint didn't reduce the information. It distributed it.
This is the creativity research made practical. The AI's search space went from "generate anything" to "fill these specific slots within these specific limits." The search space shrank. The output quality increased. Exactly what Acar et al. predicted.
Why Constrained Generation Is Universal
This principle extends far beyond presentations.
Proposal sections have schemas. An executive summary has a defined structure: situation, complication, implied question, answer. Not because we're limiting the AI's expression. Because that structure (drawn from Minto's Pyramid Principle and adapted for sales contexts) produces executive summaries that executives actually read. The AI fills the structure. The structure ensures the argument is coherent.
Contract clauses have libraries. The AI selects from approved clause templates, filling in deal-specific variables. It doesn't generate novel legal language on the fly. Novel legal language generated by an AI is how you end up in a conference room explaining to your general counsel why the limitation of liability section includes phrasing that no lawyer would have written. The clause library constrains the AI to language that's been reviewed, tested, and survived contact with actual legal proceedings.
RFP responses have question-answer structures with word-limit extraction from the question context. The client asked for 500 words. The AI generates 500 words. Not 800 words that need to be cut. Not 300 words that need to be padded. The constraint is extracted from the client's template and enforced architecturally, not requested politely.
Even the "creative" escape valve is constrained. In our presentation engine, up to 60% of slides can use a freeform layout where the AI generates custom HTML for content that no template covers. But that HTML passes through CSS variable enforcement for brand consistency, content density checks (100 to 8,000 characters), and an 8-layer sanitization pipeline. The escape valve exists. It has walls on both sides.
The Blank Page Problem
Cognitive science has studied this for decades. The blank page is not a gift. It's a trap.
When you ask someone to "write a story about anything," they freeze. When you ask them to "write a 500-word story about a red-haired boy named Colin who is missing a toe," they start writing immediately. The constraints provided focus. They eliminated the paralyzing infinity of options and replaced it with a bounded creative challenge.
The same dynamic plays out in AI generation. A prompt that says "generate a proposal" gives the model an infinite space to explore. An infinite space means an infinite number of mediocre paths and a very small number of excellent ones. The model is statistically unlikely to find the excellent path because it's swimming in options.
A prompt that says "generate the executive summary section of a proposal, following the SCQA framework, for a SaaS company evaluating deployment platforms, with a maximum of 400 words, leading with the Cost of Inaction before presenting the investment" eliminates 99% of the mediocre paths. The model doesn't have to decide on the structure, the framework, the audience lens, or the section sequence. Those decisions are made. The model's creative energy goes entirely into the language, the specificity of the examples, the precision of the argument.
That's not less creative. It's more creative within a smaller space. And in that smaller space, the average output quality is dramatically higher than what freeform generation produces.
The doEverything() Problem
Most companies approach AI prompting as a single-layer problem. One system prompt. One instruction set. Everything in one block: identity, instructions, context, constraints, formatting rules, domain knowledge, examples, and prayers. That's doEverything() in English.
And it fails the same way doEverything() fails in code. Responsibilities bleed. Constraints from one domain interfere with another. The AI forgets it's writing a contract halfway through and starts sounding like a proposal. The formatting rules from the executive summary leak into the pricing section. The identity drifts because everything is fighting for attention in the same flat text block.
The fix is the same fix software engineering discovered decades ago: separation of concerns. Break the single prompt into layers with different responsibilities and different volatility (explored in depth in the Five Layers Deep paper in this library). Break the output into typed slots with different constraints. Break the generation into stages where each stage has a clear, bounded task.
The AI doesn't need more intelligence to produce better output. It needs more structure. The structure is the intelligence.
The AI's job isn't to decide what the structure should be. That's the architect's job. The AI's job is to be brilliant within the structure.
The Template Objection
The immediate pushback: "But templates make everything look the same."
No. Bad templates make everything look the same. Good templates make everything look consistent, which is different. A bad template is a fixed layout with no variation. A good template is a structural constraint with creative latitude within it.
A portfolio of 20+ proven layouts, each with typed slots that the AI fills contextually, produces more visual variety than freeform generation. Because freeform generation converges on the same default patterns. Left-aligned title, bullet points, maybe an image if you're lucky. The AI's "creative" choice, when unconstrained, is remarkably uniform. It gravitates toward the training data's center of mass, which is mediocre by definition.
Constrained generation forces the AI to populate different structures, which produces more actual variety than letting it choose its own structure. The paradox resolves: more constraint, more variety. Because the constraints prevent convergence on the default.
The Production Evidence
Here's what we observed after implementing constrained generation across document types.
Consistency across users went from roughly 40% (measured by structural similarity of outputs for similar inputs) to above 90%. Different reps asking for similar proposals got structurally similar, high-quality results. The quality floor rose dramatically. The ceiling stayed the same. But the floor is what matters in production, because your worst output is what your client sees when your best rep is on vacation and the new hire is building the proposal at 11 PM.
Information density improved because every slot has a minimum fill requirement. Slides that used to have a headline and three sparse bullets now have a headline, a supporting metric, a body paragraph, and a visual directive. Not because someone asked for more content. Because the structure required it.
Brand consistency became automatic. CSS variables enforced through template slots. Fonts, colors, spacing, and visual hierarchy controlled architecturally. The AI can't accidentally use the wrong font because the slot doesn't accept font specifications. The brand isn't a guideline. It's a constraint.
The Principle
The principle behind constrained generation applies everywhere someone is building AI that produces consequential output.
Code generation with type constraints produces fewer bugs than freeform code generation. Marketing copy within brand guidelines produces more on-brand content than asking the AI to "match our tone." Legal documents within clause libraries produce fewer liability risks than letting the AI improvise contract language. Medical documentation within approved terminology produces fewer errors than freeform clinical notes.
The pattern is always the same: define the structure, type the slots, set the bounds, let the AI fill them. The AI's job isn't to decide what the structure should be. That's the architect's job. The AI's job is to be brilliant within the structure. And brilliance within structure is what we've been calling "creativity" for as long as humans have been making things.
A sonnet isn't less creative than free verse. It's creative in a harder way. And the outputs are more consistently excellent because the structure prevents the lazy paths.
Your AI doesn't need more freedom. It needs better constraints.
The takeaway: Constraints don't limit AI creativity: they eliminate the lazy paths, raise the quality floor, and produce more consistently excellent output than open-ended generation ever will.
This paper is part of The Architecture Series by Ed, Founder & CEO of DealiOS. For the weekly series and more technical deep-dives, visit dealios.app/library.
References
[1] Acar, O. A., Tarakci, M., & van Knippenberg, D. (2019). "Creativity and Innovation Under Constraints: A Cross-Disciplinary Integrative Review." Journal of Management, 45(1), 96-121.
[2] Damadzic, A., et al. (2022). "[Re]thinking Outside the Box: A Meta-Analysis of Constraints and Creative Performance." Journal of Organizational Behavior, 43(8).
[3] Medeiros, K. E., Partlow, P. J., & Mumford, M. D. (2014). "Not Too Much, Not Too Little: The Influence of Constraints on Creative Problem Solving." Psychology of Aesthetics, Creativity, and the Arts, 8(2), 198-210.
This paper is part of The Architecture Series by Ed, Founder & CEO of DealiOS. For the weekly series and more technical deep-dives, visit dealios.app/library.
See how these principles come together. Explore the platform →