Senior · Core path · 1 chapter
Chain-of-thought, scratchpads
When reasoning models make this redundant
Visible chain-of-thought is often an expensive and brittle interface; treat it as an experimentally justified fallback rather than a default prompting habit.
Read these first
What you will know
Get a measurable improvement on a reasoning task by asking for steps
Use a fixed held-out set and report the baseline, absolute accuracy change, and variance across runs rather than relying on a compelling example.
Measure the token and latency cost of visible reasoning
Long traces consume output budget and can push requests into slower or more expensive tiers, especially when concurrency is high.
Run the same task with and without explicit reasoning on a reasoning model, and compare quality, cost, and latency
Internal deliberation may already provide the benefit, while an added scratchpad instruction mainly exposes verbose text or changes the model’s response policy.
Keep the reasoning out of the parsed answer
A leaked or variably formatted trace can violate schemas, expose sensitive intermediate content, and make downstream extraction fail.
Show a task where forcing steps makes the answer worse
Extra intermediate claims create more opportunities for arithmetic, factual, or instruction-following errors on tasks that direct answering handles cleanly.
Chapters
- 1
Chain-of-thought and scratchpads
What chain-of-thought prompting is, what a scratchpad is, and how the two differ: one is an instruction that asks a model to produce intermediate reasoning before its answer, the other is the workspace those steps live in. Works one authorization decision through both, changes a single field to show the outcome move, and shows what happens when the reasoning and the answer share an output region a parser is reading. Ends with a decision procedure that can be run as written.
External resources
- Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelsarxiv.org
The original CoT paper and the evidence behind step-by-step prompting.
- Chain-of-thought prompting: reasoning steps as scratchpadzeroentropy.dev
This is a substantive explainer covering CoT mechanics, variants, use cases, tradeoffs, calibration risks, and production implications rather than merely promoting a tool.
- Prompting Science Report 2: The Decreasing Value of Chain of Thoughtarxiv.org
Measures how little manual CoT buys you on reasoning models.
- Chain of thought prompting (Claude docs)platform.claude.com
Copy-ready CoT patterns plus when extended thinking replaces them.
- Chain-of-thought prompting - Explained!CodeEmporium
Short visual intuition for why intermediate steps help.
- Chain of Thought Prompting Elicits Reasoning in Large Language Models [2022]AI Paper Podcast
Paper read-through covering the experiments and their limits.