Diagram of Thought
Reasoning is rarely a straight line. Diagram of Thought lets a single model build its thinking as a growing diagram of propositions, critiques, and refinements, switching between proposing ideas, criticizing them, and summarizing what survives. A validator keeps the structure sound, so the model reasons like a careful thinker revising a draft, all within one model and one context.
*These are learning tools, not prompts. They teach you to write your own, think of them as training wheels that guide you while you learn, then fall away once you can ride on your own.
Reasoning as a Diagram, Not a Line
Earlier structured-reasoning methods often rely on an external controller to manage a tree or graph of separate model calls, orchestrating search from the outside. Diagram of Thought takes a different stance: the structure lives inside one model. The model itself builds a directed graph of reasoning, where each node is a proposition, a critique, or a refinement, and the edges record how ideas depend on and respond to one another.
It does this by switching roles. Acting as a proposer, the model adds new reasoning steps. Acting as a critic, it evaluates those steps and flags the flawed ones. Acting as a summarizer, it gathers the surviving, validated propositions into a coherent answer. Role-specific signals tell the model which mode it is in, so a single autoregressive process carries the whole back-and-forth.
A deterministic validator checks the reasoning trace against a defined structure, so refinements build on sound foundations rather than accumulating unchecked. The authors ground the framework formally in category theory, giving the diagram a rigorous mathematical interpretation rather than treating it as a loose metaphor.
The power of Diagram of Thought is that proposing, critiquing, and summarizing happen inside a single model rather than across separate orchestrated calls. The model criticizes its own propositions, discards the weak ones, and keeps refining, all while a validator ensures each step fits the reasoning structure. The result is self-correction built into the reasoning process itself.
The Diagram of Thought Process
Propose, critique, refine, and summarize within one evolving diagram
Propose Reasoning Steps
In the proposer role, the model adds candidate propositions to the diagram, partial steps toward an answer. These are not assumed correct; they are entries in a graph that will be examined before anything is trusted.
For a geometry proof: propose "triangles ABD and CBD share side BD" as a starting proposition.
Critique What Was Proposed
Switching to the critic role, the model evaluates the propositions, identifying ones that are unsupported, inconsistent, or wrong. Critiques become nodes too, recording why a step was rejected so the reasoning does not silently repeat the error.
Critique: "the claim that the triangles are congruent needs an angle or side condition that has not been established yet."
Refine and Validate
The model proposes refinements that address the critiques, and a deterministic validator checks that each new step fits the reasoning structure. Only steps that pass become part of the trusted diagram, so the graph grows on sound foundations.
Refinement: "establish the shared angle, then apply the side-angle-side condition," which the validator accepts as well-formed.
Summarize the Surviving Reasoning
In the summarizer role, the model gathers the validated propositions into a coherent final answer. Because the summary draws only on steps that survived critique and validation, it rests on reasoning the process has already checked.
Summary: the established shared side and angle give congruence by side-angle-side, completing the proof.
See the Difference
External orchestration versus structure inside one model
Externally Controlled Search
A separate controller spawns and scores many model generations, expanding a tree or graph of options from the outside and stitching the results together.
Powerful, but it needs orchestration machinery around the model and coordination across many calls, which adds engineering and cost.
Diagram of Thought
The model itself builds the reasoning diagram, switching among proposer, critic, and summarizer roles, with a validator checking each step as it goes.
Self-critique and structured refinement happen within a single model and context, with a formal grounding rather than ad hoc orchestration.
Practice Responsible AI
Always verify AI-generated content before use. AI systems can produce confident but incorrect responses. When using AI professionally, transparent disclosure is both best practice and increasingly a legal requirement.
Most US states are actively legislating AI transparency and accountability. Critical thinking remains your strongest tool against misinformation.
Diagram of Thought in Action
See how propose, critique, and summarize play out across problems
"Prove that the sum of the first n odd numbers is n squared."
Propose: "Try induction. Base case n = 1: first odd number is 1, equals 1 squared."
Critique: "The inductive step claims adding the next odd number gives (n+1) squared, but the next odd number was written as 2n, not 2n+1."
Refine and validate: "The nth odd number is 2n - 1, so the next is 2n + 1. Then n squared plus (2n + 1) equals (n+1) squared. Validator accepts."
Summarize: "By induction with the corrected odd-number formula, the sum is n squared."
"Is a lower interest rate always good for an economy?"
Propose: "Lower rates encourage borrowing and spending, which boosts growth."
Critique: "That ignores inflation risk and asset bubbles from cheap credit. The claim is too absolute."
Refine and validate: "Add the trade-off: lower rates stimulate demand but can overheat prices and inflate asset values if held too long."
Summarize: "Lower rates help in a slowdown but are not unconditionally good; the effect depends on inflation and timing."
"Plan the cheapest way to ship a fragile item across the country in two days."
Propose: "Ground shipping is cheapest."
Critique: "Ground rarely meets a two-day deadline cross-country, so it fails the constraint."
Refine and validate: "Among options that meet two days, compare expedited air with adequate fragile-item protection by cost."
Summarize: "Pick the lowest-cost two-day air service that includes proper fragile handling, since ground was pruned for missing the deadline."
When to Use Diagram of Thought
Best for problems that need self-critique and revision
Perfect For
Problems where an early flaw quietly invalidates the rest benefit from explicit critique and validation before each step is trusted.
Questions that need balanced reasoning are well served by a process that proposes a claim, criticizes it, and refines toward a nuanced answer.
When you want structured, self-critiquing reasoning without building external search machinery, the whole process runs inside one model.
Because critiques and refinements are recorded as nodes, the diagram leaves a trace showing what was rejected and why.
Skip It When
If a question needs no revision, the propose-critique-summarize cycle is overhead with nothing to catch.
The iterative critique and refinement spend extra tokens, so a tight budget may favor a leaner reasoning method.
For open-ended creative work without a notion of correct or incorrect steps, the validation machinery has little to act on.
Use Cases
Where Diagram of Thought delivers the most value
Theorem and Proof Help
Work through mathematical proofs where each step must be validated before the next, catching flawed lemmas through explicit critique.
Policy and Argument Analysis
Build balanced positions by proposing a claim, criticizing its weaknesses, and refining toward a defensible, well-qualified conclusion.
Design Reviews
Evaluate technical or system proposals by surfacing objections as critiques and pruning options that fail a constraint before committing.
Complex Diagnosis
Reason through layered diagnostic problems where competing hypotheses must be raised, challenged, and narrowed to a validated conclusion.
Self-Checking Answers
Produce answers that have already been critiqued internally, reducing the chance that an unexamined error reaches the final response.
Teaching Rigorous Reasoning
Show learners how a careful thinker proposes, doubts, and revises, making the critique-and-refine habit visible in the reasoning trace.
Where Diagram of Thought Fits
From external search to structure inside the model
Where Tree of Thought and Graph of Thought rely on an external controller to manage structure, Diagram of Thought folds proposing, critiquing, and summarizing into one model. Reach for it when you want structured self-correction without standing up separate orchestration.
Related Techniques
Explore complementary structured-reasoning techniques
Reason, Critique, Refine
Try prompting a model to propose, criticize, and then summarize its own reasoning in the Prompt Builder, or explore related structured-reasoning techniques.