Sketch-of-Thought
Experts do not reason the same way for every problem. A mathematician scribbles symbols, a clinician writes shorthand, a planner links ideas in a quick chain. Sketch-of-Thought gives a model the same flexibility: a lightweight router picks one of three terse sketching styles to match the question, so the reasoning stays compact and the right kind of shorthand is used for each task.
*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.
One Shorthand Does Not Fit Every Problem
Efficient-reasoning methods that apply a single terse style do well on some tasks and poorly on others. A compact symbolic style is perfect for arithmetic but awkward for commonsense reasoning; a chain of linked concepts is great for commonsense but wasteful for a calculation. Sketch-of-Thought starts from the observation that experts switch shorthand depending on what they are reasoning about.
The method pairs a lightweight router with three reasoning paradigms. Before solving, a small routing model looks at the question and chooses the sketch that fits: Conceptual Chaining for linking ideas, Chunked Symbolism for math and quantities, or Expert Lexicons for dense domain shorthand. The chosen paradigm constrains the model to write only the essential reasoning, in the form best suited to the task.
The payoff is efficiency without a one-size-fits-all penalty. In the original study, this adaptive sketching cut reasoning tokens by up to roughly 84 percent while holding accuracy close to full verbose reasoning, because each problem got a shorthand that actually suited it.
Conceptual Chaining links the key ideas of a problem in a short ordered chain, ideal for commonsense and multi-step reasoning. Chunked Symbolism compresses quantitative reasoning into compact equations and symbols. Expert Lexicons use the dense shorthand and notation of a specialized domain. The router assigns one per question.
The Sketch-of-Thought Process
Route to a paradigm, sketch the reasoning, then answer
Route the Question to a Paradigm
A lightweight router model inspects the question and selects the sketch paradigm that best matches its character: conceptual, symbolic, or expert. This choice is made once, up front, and decides the shape of the reasoning that follows.
"If a 12 percent solution is diluted from 500 ml to 800 ml, what is the new concentration?" routes to Chunked Symbolism, because the reasoning is quantitative.
Sketch in the Chosen Style
The model reasons using only the compact notation of the selected paradigm. For Chunked Symbolism that means equations and symbols; for Conceptual Chaining a short chain of linked ideas; for Expert Lexicons the dense shorthand of the field. No verbose narration is written.
"mass = 0.12 x 500 = 60. new conc = 60 / 800 = 0.075." Pure symbolic sketch, no prose.
Stay Within the Paradigm
Because the sketch style is matched to the task, the reasoning stays both short and well-suited. A commonsense question would instead get a brief chain of concepts; a specialist question would get domain shorthand. The constraint is what keeps tokens low without hurting accuracy.
Commonsense route: "rain -> wet ground -> slippery -> caution." A few linked ideas, not a paragraph.
Produce the Final Answer
From the compact sketch, the model states the answer clearly. The user gets a short, paradigm-appropriate trace plus a clean result, having spent a fraction of the tokens a single verbose reasoning style would have required.
"New concentration is 7.5 percent."
See the Difference
A fixed style versus a routed, task-matched sketch
One Fixed Shorthand
A single terse style is applied to every question. It fits arithmetic well, but forces commonsense and specialist questions into a shape that does not suit them.
Efficient on tasks that match the style, but accuracy can slip on tasks that do not, because the shorthand fights the problem.
Sketch-of-Thought
The router sends arithmetic to Chunked Symbolism, commonsense to Conceptual Chaining, and specialist questions to Expert Lexicons. Each gets shorthand that fits.
Up to roughly 84 percent fewer reasoning tokens in the original study, with accuracy held close to full verbose reasoning across task types.
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.
Sketch-of-Thought in Action
One example per paradigm, showing how the route changes the sketch
"A car travels 240 km using 16 litres of fuel. What is the fuel economy in km per litre, and how far on a 50-litre tank?"
"econ = 240 / 16 = 15 km/L.
range = 15 x 50 = 750 km.
Answer: 15 km/L, 750 km."
"Why might a store place milk at the back of the shop, far from the entrance?"
"milk = common purchase -> forces walk through store -> passes other goods -> more impulse buys.
Answer: to route shoppers past more products and increase sales."
"A circuit has a 12 V source, two 4 ohm resistors in series. What current flows?"
"R_total = 4 + 4 = 8 ohm. I = V/R = 12/8 = 1.5 A.
Answer: 1.5 A."
When to Use Sketch-of-Thought
Best when a workload spans different kinds of reasoning
Perfect For
When questions range across math, commonsense, and specialist domains, routing each to the right sketch keeps every type both efficient and accurate.
Cutting reasoning tokens sharply across a high volume of varied queries lowers cost without sacrificing the accuracy a single fixed style might lose.
Fields that already have dense notation, engineering, chemistry, medicine, map naturally onto the Expert Lexicons paradigm for compact reasoning.
Shorter, task-matched reasoning returns faster, which helps interactive products keep step-by-step quality without making users wait.
Skip It When
If every question is the same kind, the routing step adds little; a single well-chosen terse style may be simpler to deploy.
When a full, readable rationale must be shown to a person, terse sketches in any paradigm are too compressed; use verbose reasoning instead.
The method depends on a small routing model to pick the paradigm. Where that extra component is not practical, a fixed terse style is the lighter option.
Use Cases
Where Sketch-of-Thought delivers the most value
General Assistants
Handle a broad mix of user questions efficiently by routing each to the sketch that suits it, keeping cost low across very different query types.
High-Volume Reasoning APIs
Serve varied reasoning requests at scale while controlling the token bill, since each request gets only the compact reasoning it needs.
Technical Support Tools
Lean on Expert Lexicons for domain-heavy troubleshooting, producing crisp specialist reasoning without long explanatory detours.
Tutoring Across Subjects
Give math problems symbolic sketches and reasoning problems concept chains, so each subject is worked in its natural compact form.
On-Device Reasoning
Fit step-by-step reasoning into tight compute budgets by always producing the shortest task-appropriate sketch rather than a long trace.
Agent Reasoning Steps
Keep each internal reasoning turn of an agent compact and well-matched, controlling context growth across a long sequence of mixed decisions.
Where Sketch-of-Thought Fits
The adaptive step in the efficient-reasoning family
Where Chain of Draft applies one shorthand to everything, Sketch-of-Thought routes among several. If your workload is uniform, the simpler Chain of Draft may be enough; if it spans many kinds of reasoning, the router earns its keep.
Related Techniques
Explore complementary efficient-reasoning techniques
Match the Sketch to the Task
Try guiding a model toward symbolic, conceptual, or specialist shorthand in the Prompt Builder, or explore the wider family of efficient-reasoning techniques.