Chain-of-Thought Pattern
The most universal instruction pattern in the corpus. Instructs the LLM to think step-by-step before taking any action. Present in ~60% of tools (~31/52), consistently improving code quality and reducing errors.
Variants (confidence: high)
- Implicit (Anthropic Claude Sonnet): Relies on model’s native chain-of-thought capability. No explicit instruction.
- Explicit structured (Windsurf, RooCode): “Produce a plan, get approval, then implement.” Formalized as a multi-step workflow.
- Hidden thinking trace (Antigravity, Gemini CLI): CoT content goes into a private reasoning trace, not visible to the user. Public response is the distilled output.
Co-occurrence (confidence: high)
Chain-of-thought frequently co-occurs with:
- task-state-machine — CoT provides the reasoning for each task state transition
- plan-before-implement-gate — CoT produces the plan that the gate evaluates
- paired-thinking-response — Hidden CoT trace paired with public output
Implications: Chain-of-thought is the closest thing to a universal best practice in prompt engineering. Its effectiveness is consistent across tool types, prompt formats, and model families. The variant choice (implicit vs. explicit vs. hidden) reflects a tool’s stance on transparency — hidden traces keep the UI clean, explicit traces build user trust, implicit traces optimize for speed.
See also: task-state-machine, plan-before-implement-gate, paired-thinking-response
Relationships
- “is-sibling-of” chain-of-thought — same reasoning-before-answer technique; this page covers observed tool-implementation variants (implicit/explicit/hidden), the sibling covers the general prompting-guide technique
Sources
- [raw/prompts/articles/taxonomy-synthesis-2026-07-16.md]