The 6-element prompt framework is incomplete. Here are 9.

Pattern: Framework evolution + Research anchoring

Angle: Position 9-field as evolution of generic 6-element frameworks

Keywords: prompt engineering framework, prompt anatomy

Priority: P0 (launch)

Length: 2,000-2,500 words

Thesis: Every prompt guide teaches 6 elements (Role, Goal, Context, Format, Examples, Constraints). They're missing Tone, Reasoning, and Tools — and testing shows those three change output quality significantly.

Outline:

The standard 6 elements everyone teaches (cite The AI Corner, Anthropic docs, OpenAI)
What's missing and why it matters
Tone: testing showed it significantly impacts output across all platforms
Reasoning: triggers different processing (Claude's <thinking>, CoT)
Tools: explicit tool declarations change how models plan
The 9-field anatomy system — cognitive flow ordering
Complexity levels: start with 3, scale to 9
CTA: try the 9-field editor

Draft

title: "The 6-element prompt framework is incomplete. Here are 9."

slug: nine-field-prompt-anatomy

description: Every prompt guide teaches 6 elements. They're missing Tone, Reasoning, and Tools — and those three change output quality significantly.

keywords: prompt engineering framework, prompt anatomy, 9 field prompt, prompt structure author: Mariano date: 2026-04-13

The 6-element prompt framework is incomplete. Here are 9.

Every prompt guide teaches 6 elements. They're incomplete.

You've seen the frameworks. Anthropic publishes guides. OpenAI has documentation. Popular frameworks use acronyms: RACE, CREATE, SCAMPER. They all converge on roughly the same 6 building blocks: role, context, task, examples, output format, constraints.

It's advice worth following. Six is better than zero. But it's not the full anatomy.

Three fields are missing. Tone, Reasoning, and Tools. And in practice they meaningfully affect output quality across platforms. Without them, you're optimizing incomplete prompts.

This is what the full 9-field anatomy looks like. Why each field matters. And why the order is cognitive, not arbitrary.

The standard 6 everyone teaches

Before we go deeper, let's acknowledge what's working. The 6-element frameworks are based on good thinking.

Role — tells the AI who it is. A helpful assistant. A technical writer. A customer support agent. Role primes behavior without instruction.

Context — tells the AI what it knows. Company history, technical details, customer data, domain background. Context is what the AI should consider before responding, but wouldn't know without you telling it.

Task — tells the AI what to do. Write an email. Classify this message. Generate a prompt. Task is directional.

Examples — shows the AI what good looks like. Few-shot learning. Demonstrations of the task in action. Examples are often the highest-leverage field on a prompt.

Output Format — tells the AI how to structure the response. JSON. Bullet points. Markdown. A specific schema. Output Format constrains shape.

Constraints — tells the AI what not to do. Don't use jargon. Don't exceed 100 words. Don't mention pricing. Constraints are safety rails.

Solid advice. Workable framework. But three critical fields are missing.

What's missing, and why it matters

Most prompt failures aren't because a field is bad. They're because a field is absent.

You skip Tone because you assume "helpful assistant" is fine. It's not fine. It's default. And default is why outputs feel generic.

You skip Reasoning because you think the model will figure it out. It won't. Different models have different reasoning modes. Some respond to "think step by step." Others respond to explicit reasoning instructions. Without guidance, the model processes at surface depth.

You skip Tools because your use case doesn't seem to need them. Then you get hallucinated function calls. Or the model plans wrong. Or it chooses the wrong tool for the job.

These aren't edge cases. These are structural gaps that affect every prompt at scale.

Three missing fields

Tone: controls register, personality, and formality

Tone is the voice of the response. It controls whether the AI sounds like a lawyer, a teacher, a peer, or a guide. It's not just word choice. It's how the model approaches the task.

Examples:

"Respond in the tone of a software engineer explaining to a junior developer." (Technical, patient, practical.)
"Respond in the tone of a venture capitalist evaluating a business." (Critical, metrics-focused, skeptical.)
"Respond like a Shakespeare scholar, formal and deeply learned." (Elevated, erudite, verbose.)
"Respond like a friend giving honest feedback." (Warm, direct, casual.)

Same task. Different tone. Completely different output.

Why doesn't this get taught? Because people assume tone is soft. It's not. Testing across Claude, GPT-4, and Gemini shows that explicit tone instructions significantly move output quality. Register matters. Personality matters. Without a Tone field, models default to "helpful AI assistant" voice — and that's rarely what you want.

Reasoning: triggers different processing modes

Reasoning tells the model how to think, not just what to output.

This is where Anthropic's research shows up. Chain of Thought. Extended Thinking. ReAct. These aren't tricks. They're processing modes. And explicitly declaring your reasoning expectations changes how the model solves the problem.

Examples:

"Reason through this step-by-step. Show your thinking." (Chain of Thought — the model verbalizes intermediate steps.)
"Use structured reasoning: First, identify the constraints. Second, list possible approaches. Third, evaluate tradeoffs." (Guided reasoning — the model follows a framework.)
"Use your extended thinking mode. This is complex. Take time to reason through it." (Deep processing — only Claude; activates extended thinking.)
"Use ReAct logic: think about the problem, then take an action, then observe the result, then reason again." (Agent reasoning — for agentic tasks.)

Without an explicit Reasoning field, the model does surface-level pattern matching. With it, the model engages deeper processing modes. Quality goes up. Consistency goes up. Hallucinations go down.

Tools: explicit tool declarations change planning

Tools tell the model what capabilities are available.

If you're building an agent or running a prompt in a system with function calls, declaring tools explicitly changes how the model plans. It stops hallucinating function names. It uses the right tool for the job. It chains tools in the right order.

Examples:

"You have access to: (1) Google Search API, (2) URL Fetcher, (3) Calculator, (4) Code Executor. Use these tools to answer the question."
"Available tools: AccountLookup (returns customer ID, transaction history, plan status), PolicyReference (returns billing terms), EscalationForm (routes to human agent)."
"You have access to a knowledge base search tool, a real-time pricing API, and an order creation tool. Use them in this workflow: search → verify pricing → create order."

Without explicit tool declarations, agents choose wrong tools or hallucinate tools that don't exist. With them, the model understands the problem space and plans effectively.

The 9-field anatomy: cognitive flow

Now the full picture. Not random. Ordered by cognitive flow — how a human learns to approach the task.

Role — Who are you? (20% impact)
Tone — How do you speak? (10% impact)
Context — What do you know? (15% impact)
Task — What are you doing? (25% impact)
Reasoning — How are you thinking? (5% impact)
Examples — What does good look like? (5% impact)
Output Format — How are you formatting the response? (10% impact)
Constraints — What are you avoiding? (5% impact)
Tools — What capabilities do you have? (5% impact)

The percentages are rough, directional weights — not benchmarks. Task is the heaviest lever. Role is second. Context is third. The rest are meaningful but lower-impact.

But "lower-impact" doesn't mean you can skip them. Tone, Reasoning, and Tools are force multipliers on the other fields. Nail Task without nailing Tone, and you get solid but generic output. Nail both, and you get exceptional output.

Complexity levels: progressive disclosure

You don't memorize all 9. You start where you are and add fields as needed.

Complexity 1: Simple (3 fields)

Role
Task
Examples

Enough for straightforward work. Classify emails. Generate summaries. Build quick prototypes. If it's working at 3, don't add 9.

Complexity 2: Standard (5 fields)

Role
Context
Task
Output Format
Constraints

This is production work. You're specifying what the model should know, what it should do, what shape it should return, and what boundaries matter.

Complexity 3: Advanced (9 fields)

All of them

For complex reasoning tasks, agent systems, and high-stakes work where tone register matters or reasoning process is critical.

Start at Complexity 1. Graduate to 2 when you hit consistency issues. Graduate to 3 when output quality depends on how the model reasons or what tone it uses.

Tone: the data

Testing across multiple models showed Tone's impact clearly:

Without Tone: Claude outputs formal-helpful voice. GPT-4 outputs slightly more casual. Gemini varies.
With explicit Tone: All three converge toward the intended register. Output consistency increases. Quality is higher across the board.

Example: customer support agent.

Without Tone field, you get different personalities from each model. Claude stays professional. GPT-4 softens toward warmth. Gemini leans formal.

With "Respond like you're helping a frustrated customer. Be warm, direct, and assume they've already read the docs" — all three models produce similar warmth and directness.

Tone isn't a luxury. It's structural.

Reasoning: the research

This comes from Anthropic's work on prompting and multi-step reasoning. When you explicitly declare how a model should think:

Chain of Thought prompts reduce hallucinations by requiring intermediate reasoning.
Extended Thinking (Claude) engages deeper processing for complex problems.
Structured reasoning frameworks force the model to follow a method rather than guessing.

Without a Reasoning field, models operate at surface depth. With it, they operate at process depth.

Tools: the necessity

If you're building agents, tools aren't optional. They're essential. Explicitly declaring available tools:

Stops hallucinated function calls
Reduces the model's decision space
Improves tool selection accuracy
Enables tool chaining and planning

This is why systems like AutoGPT, LangChain, and others explicitly list tools. It's not decoration. It's architecture.

The order is cognitive, not arbitrary

Why this order?

Role first — you can't think about how to speak or what to do until you know who you are.

Tone second — once you know who you are, you know how you sound.

Context third — you know who you are and how you speak; now you load knowledge.

Task fourth — only now, with role and context locked, do you specify what to do. Task clarity depends on role clarity.

Reasoning fifth — you understand the what; now you specify the how-to-think.

Examples sixth — you've specified direction; now you show it in action.

Output Format seventh — you know what you're doing; now you specify structure.

Constraints eighth — you've defined the positive; now you define the negative.

Tools ninth — you have the full picture; now you declare capabilities.

It's not random. It's how cognition works. Who → How → What → Do → Think → Show → Format → Avoid → Capable.

Why no product owns this yet

Most prompt platforms are text editors with syntax highlighting. They don't give you structure. They give you a blank box.

Some publish guides that mention 6 elements. Good. But they don't enforce it. They don't compile it for different platforms. They don't help you reason about what each field should contain.

And nobody teaches the full 9. Because most people don't know the anatomy exists. It's built from Anthropic's research, field observation, and empirical testing. But it's not canonized anywhere.

That's changing. Because structured prompts aren't a feature. They're the foundation of reproducible AI work at scale.

Start here

prompt-x implements the 9-field anatomy with built-in guidance. Start with 3 fields. Scale to 9. The editor understands each field's purpose. It guides your thinking. It compiles for any platform.

The result: better outputs, consistent quality, and prompts that travel.

Try it free

Related reading