LLM Reasoning Playbook

LLM Reasoning Playbook

A working reference for choosing, combining, and deploying advanced LLM reasoning frameworks, from your first stack to your tenth.

New to advanced LLM frameworks?

Start with Build Your Recipe: answer seven quick questions and get an ordered, paste-ready plan. Reach for the Reasoning Framework Playbook as the full reference once you know what you are reaching for.

The Reasoning Framework Playbook

12 core frameworks across 7 axes: taxonomy, choice matrix, boilerplate templates, and a selection flowchart.

Core reference

Build Your Recipe

Answer seven quick questions about your problem and walk away with an ordered, copy-pasteable reasoning plan, interactive builder included.

Workflow

Worked Stack Examples

Four end-to-end examples of composing frameworks across axes, plus a stack-selection cheatsheet to match your problem to a proven stack.

Applied

Output Eval Rubrics

How to judge each framework's output: universal dimensions, per-framework signature checks, and a cost/latency budget model.

Evaluation

Decision Log & Template

A per-task record of which stack you chose and why, plus the workflow for keeping the playbook current.

Practice

Anti-Pattern Gallery

Concrete before/after failures mapped to rubric violations. A living gallery you extend from real lessons.

Living

How the pieces fit together

The playbook is a lifecycle, not a pile of pages. You move through it in order, and most people stop after step two, which is exactly where reasoning stacks fail.

  1. Frameworks: learn the parts, the thirteen frameworks across seven axes.
  2. Build Your Recipe: compose a recipe for your specific problem.
  3. Run the prompt in your own LLM and get a reasoning trace back.
  4. Output Eval Rubrics: judge what came back. Is it right, faithful, and ready to ship?
  5. Anti-Pattern Gallery: concrete pictures of the failures the rubrics name.
  6. Decision Log: record what you chose and how it scored, so the next task is faster.

Quick reference: the 7 axes & 12 core frameworks

AxisFrameworks (core)
A · TopologyCoT · Thread of Thought · Tree of Thoughts · Graph of Thoughts
B · Sampling/AggregationSelf-Consistency
C · Grounding/ActionReAct · PAL / Program-of-Thoughts
D · Decomposition/PlanningLeast-to-Most
E · AbstractionStep-Back
F · Self-Evaluation/MemoryMetacognitive / Self-Refine / RaR · Reflexion
G · SteeringDirectional Stimulus Prompting

Tier-2 variants: Skeleton-of-Thought · Universal Self-Consistency · Chain-of-Verification · Plan-and-Solve · Self-Ask · Analogical.

Fast selection (4 gates)

  1. External info / tools / exact computation? → PAL (compute) · ReAct (tools) · CoVe (verify)
  2. Large / messy input context? → Thread of Thought
  3. Explore alternatives or merge sub-solutions? → ToT (search) · GoT (merge) · Least-to-Most (ordered)
  4. Dominant secondary need? → Self-Consistency (reliability) · Step-Back (principle) · Reflexion/Self-Refine (retry) · RaR (ambiguity) · DSP (steering) · CoT (default)

Axes are orthogonal, so layer one framework per axis for high-value stacks (e.g., ReAct × Self-Consistency × Step-Back).