Free Your LLM to Pursue Goals, Not Follow Rules

This guide and its prompt tools are updated as the method evolves.

Schema-shots emerged from a practical problem: rules were brittle, detailed examples (shots) were powerful, but bulky. I needed a compact way to show models what good behavior looks like.

The usual prompt engineering instinct when something goes wrong is to add a rule. Schema-shots flip this. They turn the model from rule-following to goal-pursuing.

My field tests were clear: schema-shots outperformed rules and shots on prompts involving layout and analysis. Each approach has its strengths, but schema-shots were smaller than both and best at describing what good does rather than what it looks like.

When you give an LLM a stack of layout and formatting rules, it spends a lot of its time just working through them. Ask for heuristic analysis on top of that, and your analysis quality degrades.

Just as with people, focus helps LLMs deliver higher-quality work.

The schema-shot approach seeks to clear the road of rules, so LLMs can spend as much inference time as possible on desired behaviors. The model can then focus more fully on the goals you need it to fulfill.

By the end of this article, you'll have everything you need to build your own schema-shots. The prompt tools at the end will help too, but they work best when you can oversee the choreography.

What is a "schema-shot prompt?"

Schema-shot prompts work by outlining behavioral cues that choreograph each move. A shot demonstrates what good looks and feels like. A schema-shot orchestrates what good does.

A well-delivered schema-shot is more than just the shot itself. It also needs a disposition and a why statement. If the shots are choreography, then these two parts are the dancer.

The dancer

• A disposition — The role and attitudes the model brings to the task

• A why statement — Connects the disposition to the task. It's the bridge between who the model is and why the task matters.

The choreography

• A schema — A blueprint that defines structure, organization, and relationships.

• A shot — An example that demonstrates behavior, judgment, rhythm, or an output pattern.

How to choreograph a schema-shot prompt

Start with the disposition. "You are a..." This sets the archetype the LLM will draw inference from. It drives everything that comes after it.

Include why the model is doing the job. A model that knows why it's doing something will outperform one that only knows what to do.

Avoid imperative rules, or soften them. Rules constrain. A lighter touch frees the model to pursue the goal.

Think of the {{tags}} as the choreography moves. Put each move where you want it and describe the behavior you want inside the tag.

For this example, we'll focus mostly on the disposition at the top.

You are a journalist stationed on Mars. Deadpan, and world-weary, you cover your subjects with affectionate contempt in a world where catastrophe passes for inconvenience. 

Why: A great story brightens people's day. At least that's what your boss keeps saying.

# MARS MONTHLY — {{date}} 

---
{{topic tag}} 

**{{article headline}}**

{{short article description}} 

---
{# add 3 more articles here: #}

--- 
{{sign-off message}} 

--- 

{{footer layout: company name, ©, year, legal disclosure: long}}

The schema-shot tags above were left bare on purpose. This allows the disposition and why to shine through the structure.

Anthropic's Claude will take this to the extreme, allowing the personality of the disposition to influence the layout and text color choices. The flex is an example of goal-pursuing behavior.

While we could have added more detail to the tags, I found that I got better results by allowing unrestricted creativity of the disposition. The schema provides valleys for our dancer's inference to flow into. We'll let them cook.

Let's focus on the choreography

Let's back off on the disposition and focus on the dance moves. We'll use a new prompt for a product comparison analyst.

Unlike the last example, these {{moves}} are full-bodied. You can quite literally write whatever you want to happen, right where you want it.

You are a skeptical product comparison specialist.

Why: You work this way because people rely on this research to make real purchasing decisions.

# {{RESEARCH TOPIC}} — {{date the research was performed}}
{{table with relevant categories: Each cell includes a source link for follow-up}}

## What Matters Most
{{a few sections covering what's relevant for this product type. Each section names the area, summarizes how the compared products stack up, and notes any meaningful tradeoffs}}

## What to Research Next
{{unverified claims, what was not found, thin coverage, and what that means}}

## Decision Guidance
{{the best fit product for each relevant buyer type, explanations of why, notes of uncertainty that might affect their confidence in that recommendation}}

Why the tags?

The Jinja2 style {{brace tags}} send a loud templating signal to the model, automatically putting it in a template mindset.

• For behavioral moves, use {{double-braces}}
  Earlier iterations of this approach used single braces. Double braces evolved from testing on small local models. Single braces sometimes appeared printed in the output rather than being filled. Switching to double braces resolved this issue.

• For inline instructions, use {# comments #}
  These will be read as side conversations and not part of the template. You can use them to communicate instructions that behavioral moves can't perform.

What about stuff that's not in the tags?

Any information or structure outside the tags is highly likely to remain unchanged unless an edge case forces a change. The tags create a clear contrast between what is inside and outside them. In this case, the "What to Research Next" heading in Markdown is highly likely to stay static.

## What to Research Next
{{unverified claims, what was not found, thin coverage, and what that means}}

Shorthand symbols

Symbols are an optional yet useful way to express behaviors. They are not the method. The method is the behavioral choreography.

You can happily make schema-shots without ever using a symbol. I use them when I want to save space, and when they make it easier to say what I want.

You can have a lot of fun with these. The ones below are a small subset of the patterns you can use, but they are the ones I reach for most frequently.

• {{news: serious}} creates a relationship

• {{symptoms → diagnosis}} shows movement

• {{low|medium|high}} makes a choice palette

• {{news: funny/serious}} blends two things

• {{prioritize: cost > quality > speed}} describes an order

• {{news: short, recent, popular}} softly adds things up

• {{food: cheap + fast + tasty}} firmly adds things up

• {{news: -fluff}} excludes something

Symbol combinations

For models that understand them, the symbols can sometimes be more precise than natural language. Goal-oriented moves are a strong example of this.

• {{story angle: facts, affected people, open questions → fair frame}}
  A story angle shaped by the facts, the affected people, and the open questions, so the reporting lands in a fair frame

• {{restaurant: quiet|lively|romantic + budget → best fit}}
  A restaurant choice that picks a vibe, considers budget, and lands on the best fit

• {{style: precise/economical, goal → behavior, symbol_use: only when it clarifies, intent > rules, compressed + clean, -bloat}}
  Be precise and economical. Start from the goal and let it drive the behavior. Use symbols only when they clarify. When intent and rules conflict, honor the intent. Keep output compressed, clean, and free of bloat.

Let's apply these to our product analyst. We'll use shorthand symbols and toughen up the disposition. Now our prompt looks like this.

You are a skeptical product comparison specialist.

You pressure-test claims against independent sources and treat marketing copy as a lead, not evidence.

When sources conflict, you give more weight to clear, recent, low-bias, real-world evidence.

You only draw conclusions after weighing the available evidence.

Why: You work this way because people rely on this research to make real purchasing decisions.

# {{RESEARCH TOPIC}} — {{date}}
{{table: ~3 column, product-relevant categories, source link per cell}}

## What Matters Most
{{product type → key areas: stack up, tradeoffs}}

## What to Research Next
{{what was not found → unverified claims, thin coverage, open questions}}

## Decision Guidance
{{buyer type → best fit: reason, uncertainty}}

Take note of how the rules were softened into personal traits rather than imperative instructions. A model that owns a behavior as part of its character will pursue it. A model following a rule will comply with it until something more demanding comes along.

• ❌ "Only draw conclusions after weighing the available evidence."

• ✅ "You only draw conclusions after weighing the available evidence."

Now that we've made our updates, let's try our product comparison analyst out:

Prompt: "Please compare Amazon Alexa vs Google Home."

The result is a genuinely helpful, fully fact-checkable comparison that adheres firmly to the disposition. Even pitching in with a fair mention of Home Assistant at the end.

Flexibility

The Product Comparison Analyst was designed solely to compare products. What happens if we ask it something that's an edge case?

I experimented with researching single products, company-to-company comparisons, single-company research, and stock ticker comparisons.

The disposition and choreography hold fast. Improvisation occurs when the flex is needed for the edge case.

When running a rework of an experimental stock analyst prompt, the LLM consistently added a risks section that wasn't in the prompt when it found a substantial risk. If I had limited it with rules, the risks might not have surfaced. This is the goal-seeking behavior we want

Thinking traces

Reading the thinking trace of Gemma4-31b executing the Mars News Maker prompt showed that it spent the majority of its time on story-building, with a quick 3-line compliance check as an afterthought.

For the Product Comparison Analyst, the model spent its reasoning time on the analysis: pressure-testing claims, planning searches, weighing sources, and building buyer personas.

Layout barely entered the equation. It simply followed the schema.

Start making your own!

Add the schema-shot-maker prompt to your favorite frontier model to quickly transform any prompt into a schema-shot. It can also help you make new prompts from scratch from a simple description.

The maker uses shot-based and schema-shot techniques to craft prompts, so it's worth checking out just to see how it's crafted.

Schema-Shot Maker Prompts & Skills on GitHub:
https://github.com/AG3design/schema-shot-maker/

How to work with schema-shot-maker

• Give it a prompt and ask it to transform it

• Or, give it an idea, and it will craft you a prompt from scratch

• It will default to schema-shot best practices and will not ruin rule-based prompts. It will move what it can into the disposition and why statement and retain rules where needed. It will not over-rotate. I've experimented with transforming pure-rule-based, shot-based, and schema-shot-based prompts. The outcomes subjectively outperformed my originals in every instance.

• If you want it to be ruthless, tell it to be aggressive about making your prompt schema-shot-based. See what breaks. This can be a good technique for identifying which rules are load-bearing in your prompts.

• If you want it to preserve rules in a very heavily rule based prompt, ask it to turn the rules into traits.

• By default, it will only use symbols when they best clarify the meaning. If you want heavy compression or natural language only, tell it.

• When it finishes, ask it to do a step-by-step check of your original prompt to ensure everything you wanted survived. Ask it what it thinks. It will help you craft towards what you want.

Thanks for reading!