AI Prompt Builder with BRAID

The AI Prompt Builder is Raison's visual prompt design tool powered by the BRAID framework. Instead of writing free-form prompts, you design structured reasoning flows through an interactive AI chat and diagram editor. Available on Team, Team Plus, and Enterprise plans.

What the AI Prompt Builder Is

Standard prompts are instructions written in natural language. They work well for straightforward tasks but can lead to inconsistent results on complex, multi-step reasoning problems — the model interprets the instructions differently each run, "drifting" as it generates tokens.

The AI Prompt Builder solves this by structuring prompts as directed graphs: discrete reasoning steps (nodes) connected by explicit transitions (edges). The resulting structure is expressed in Mermaid diagram syntax, which the model reads as a precise execution plan rather than loose natural language guidance.

Conceptual Foundation

The AI Prompt Builder is grounded in the BRAID framework — Bounded Reasoning for Autonomous Inference and Decisions — a structured prompting methodology developed by Armağan Amcalar and Eyup Cinar (2025) and validated across multiple reasoning benchmarks. Read the paper →

The core insight: traditional Chain-of-Thought prompting lets models "think aloud" in unstructured natural language, which increases token usage without proportional accuracy gains and introduces reasoning drift. BRAID replaces this with bounded, symbolic structures — Mermaid-based instruction graphs — that constrain the model's reasoning path to a deterministic logical flow.

Benchmark results from the paper:

• On the SCALE MultiChallenge dataset (complex multi-step reasoning), BRAID raised GPT-4o accuracy from 19.9% to 53.7% and gpt-5-nano-minimal from 23.9% to 45.2% — smaller models matching or beating larger ones using unstructured prompting
• On AdvancedIF, gpt-5.1-medium reached 71% accuracy with BRAID vs. 60% with classic prompting; gpt-5-nano-minimal more than doubled from 18% to 40%
• On GSM-Hard, BRAID consistently pushed near-perfect accuracy, e.g., gpt-5-medium from 95% to 99%
• The best configurations achieved 30–74× improvement in cost efficiency (Performance-per-Dollar) vs. monolithic large-model deployments — by using a high-capability model to generate the reasoning graph once, then caching it for repeated inference with a smaller, cheaper solver model

The paper calls this the BRAID Parity Effect: a smaller model with bounded reasoning often matches or exceeds a model one or two tiers larger using unstructured prompting.

Core Concepts

Nodes

Each node represents a discrete, atomic reasoning step. Effective nodes are focused — they describe one observation, decision, or action, not a paragraph of instructions. The paper found nodes under 15 tokens yield the highest adherence in smaller models.

Examples of good nodes:

• A[Extract user intent]
• B[Check account status]
• C[Draft response: empathetic tone]

Examples of overly broad nodes (avoid):

• A[Understand the user's question and figure out what they need and write a response]

Edges

Edges connect nodes and define the reasoning flow. Labeled edges act as explicit condition checks:

A -- "If account active" --> B
A -- "If account suspended" --> C

Conditions should be mutually exclusive to avoid ambiguity. This transforms the model's inference from probabilistic token prediction into a directed traversal of a decision tree.

DAG Structure

The full graph is a Directed Acyclic Graph (DAG) — no circular dependencies in the main flow. Terminal verification loops (critic nodes that route back for revision) are an exception used to emulate deliberate self-correction before the final output.

Token Allocation

Each node implicitly allocates a slice of the model's attention to a bounded subtask. By decomposing complex instructions into many small nodes, BRAID distributes cognitive load across the graph rather than front-loading it into a monolithic system prompt.

The Interface

The AI Prompt Builder has three panels:

• Flow Diagram — A live-rendered Mermaid diagram of the current reasoning graph. Nodes and edges update as you chat.
• AI Chat — A conversation interface where you describe what you want your prompt to do. The AI assistant designs or refines the graph based on your instructions.
• Conversation History — All previous exchanges with the AI assistant for the current session.

Creating a Prompt

1. Open an agent in the dashboard.
2. Click New Prompt > AI Prompt Builder.
3. Describe your use case in the chat: "I need a prompt that handles customer support inquiries, checks whether the issue is billing or technical, and routes to the appropriate resolution path."
4. The AI assistant generates an initial Mermaid graph.
5. Refine through conversation: "Add a verification step at the end to check that the tone is empathetic."
6. The diagram updates in real-time as you iterate.

Reading the Mermaid Diagram

A BRAID graph looks like this:

graph TD
    A[Identify inquiry type] --> B{Billing or Technical?}
    B -- "Billing" --> C[Check account status]
    B -- "Technical" --> D[Identify product area]
    C --> E[Draft billing response]
    D --> F[Draft technical response]
    E --> G{Check: Tone is empathetic?}
    F --> G
    G -- "Yes" --> H[End: Send response]
    G -- "No" --> I[Revise tone]
    I --> G

Each node label is a step the model will execute. Diamond nodes ({}) are decision points. Labeled edges are conditions. Terminal loops (like G → I → G) implement self-correction.

Editing: Chat vs. Direct Manipulation

• Via chat — The most natural workflow. Describe what you want to add, change, or remove. The AI assistant updates the graph.
• Direct editing — You can edit the raw Mermaid syntax directly in the diagram panel for precise control.

Both approaches update the same underlying graph.

Converting to a Deployable Prompt

When you are satisfied with the graph, click Convert to Prompt. Raison packages the Mermaid diagram as the prompt content — this structured diagram becomes the system message delivered to your model at inference time.

Deploy the prompt through the normal lifecycle: publish → deploy to Development → promote to Staging → promote to Production. The SDK delivers the structured graph as a string, which you pass as the system message to your LLM.

Usage Limits

AI Prompt Builder message limits are per user, per billing period (not shared across the organization):

Each exchange with the AI assistant — one user message plus one assistant response — counts as one message. The billing period resets when your subscription renews. You can check your current usage under Settings > Billing.

Best Practices

Drawing from the BRAID paper's design principles:

Decompose tasks into atomic steps. Each node should represent one thing. If you find yourself writing a multi-sentence node, split it.

Use deterministic branching. Edge conditions should be mutually exclusive. Avoid vague transitions — label every branch explicitly.

Add verification loops. Before the final output node, include a "Critic" phase: nodes that check tone, length, constraint compliance, or correctness. If a check fails, route back to a revision node.

Scaffold, don't generate. For open-ended tasks, nodes should describe constraints and semantic requirements of the output, not the output itself. This forces the model to apply its own language generation within the bounded structure.

Define node contracts. Each node should have a clear input and output. What does the model know entering this node? What should it produce before moving to the next?

Limitations

The AI Prompt Builder works best for well-defined problem spaces with clear reasoning steps and explicit success criteria. For highly creative, open-ended tasks where variability is desirable, traditional free-form prompts may be more appropriate.

Complex graphs with many nodes require more tokens in the system message. For high-volume, cost-sensitive deployments, the amortized cost model applies: generate the graph once with a capable model, cache it, and run inference with a smaller model — this is the configuration that achieves the highest PPD gains in the BRAID paper.