Reasoning Architect DocsBack to app

Reasoning frameworks

A framework decides how the model reasons through your task. The research step recommends one of these 6, and you can pick any of them in the builder. Your choice is written into the prompt itself, so the prompt keeps working wherever you run it.

Chain-of-Thought

Default

One path, reasoned step by step.

ParsePlanSolveCheckAnswer

When the recommender picks it

Linear, single-path problems where step-by-step reasoning reaches one clear answer (factual, analytical, or procedural topics).

Best for

  • Structured problems with clear steps
  • Known solution paths
  • Sequential reasoning

Examples

  • Mathematical calculations
  • Code debugging
  • Financial analysis
Speed
Fast
Cost
Low
A single reasoning pass. The baseline the other ratings are measured against.

The protocol it adds to your prompt

These steps are inserted into the prompt’s <meta_strategy> block, word for word.

  1. 1First, I will analyze the problem and identify key components.
  2. 2Then, I will evaluate each component systematically.
  3. 3Finally, I will synthesize findings into a coherent solution.

Tree-of-Thoughts

Several branches explored, weak ones pruned, the best one kept.

ProblemIdea AIdea BIdea CEvaluateBest path

When the recommender picks it

Exploratory, multi-path problems with trade-offs or competing solutions that benefit from generating, scoring, and branching among alternatives (design, strategy, open-ended analysis).

Best for

  • Complex, open-ended challenges
  • Multiple competing solution paths
  • Decisions with real trade-offs

Examples

  • Strategic planning
  • Product and system design
  • Creative brainstorming
Speed
Slow
Cost
High
Each expert proposes three paths, then the best are branched and re-scored over two rounds. Expect several times the output of Chain-of-Thought.

The protocol it adds to your prompt

These steps are inserted into the prompt’s <meta_strategy> block, word for word.

  1. 1GENERATEEach expert proposes 3 divergent solution paths.
  2. 2EVALUATECritique each path against constraints. Score: Relevance (0-10), Novelty (0-10), Impact (0-10).
  3. 3BRANCHExpand top-scoring paths with 2-3 sub-branches.
  4. 4ITERATERepeat evaluation cycle for 2 iterations.
  5. 5SYNTHESIZEConverge on the optimal path with justification.

Expert panel: the builder asks you to define the experts whose ideas it branches and scores.

Self-Consistency (CoT-SC)

Three independent reasoning paths, then a majority vote.

QuestionPath 1Path 2Path 3DisagreesVoteAnswer

When the recommender picks it

Single-answer arithmetic, factual, or classification tasks where reliability matters: sampling multiple independent reasoning paths and taking the majority answer beats a single chain.

Best for

  • Questions with one correct answer
  • Tasks where a wrong answer is costly
  • Reducing run-to-run variance

Examples

  • Arithmetic and word problems
  • Classification and labelling
  • Fact checking
Speed
Moderate
Cost
Medium
Three full reasoning passes plus a vote, so roughly three times the output of Chain-of-Thought.

The protocol it adds to your prompt

These steps are inserted into the prompt’s <meta_strategy> block, word for word.

  1. 1SAMPLEIndependently reason through the problem step by step 3 times, varying your approach each time.
  2. 2EXTRACTState the final answer each path arrived at.
  3. 3VOTESelect the answer that the majority of paths agree on.
  4. 4RESOLVEIf the paths disagree, reason explicitly about why and pick the best-supported answer.
  5. 5REPORTGive the consensus answer and note the level of agreement.

Step-Back Prompting

Find the governing principle first, then apply it to the question.

QuestionPrincipleKnown factsApplyAnswer

When the recommender picks it

Conceptual or knowledge questions that benefit from first establishing the governing principle or general case before reasoning about the specific instance.

Best for

  • Conceptual and knowledge questions
  • Problems that hinge on a general principle
  • Avoiding surface-level answers

Examples

  • Physics and chemistry questions
  • Policy or legal reasoning
  • Explaining why something works
Speed
Fast
Cost
Low
One pass with a short abstraction step up front, so close to Chain-of-Thought.

The protocol it adds to your prompt

These steps are inserted into the prompt’s <meta_strategy> block, word for word.

  1. 1ABSTRACTStep back from the specific question to the broader principle, concept, or general case behind it.
  2. 2ESTABLISHState what is known and true at that higher level.
  3. 3REASONWork back down from the principle to the specifics of this question.
  4. 4SYNTHESIZEGive the answer, showing how it follows from the principle.

Grounded Synthesis

Answers built only from your files or repo, with every claim traced to a source.

FilesRepoEmailsExtractVerifyCited answerGap flagged

When the recommender picks it

Corpus-grounded tasks: summarizing, extracting from, or synthesizing across provided documents or repository source code, where every claim must be traceable to the supplied material rather than general knowledge.

Best for

  • Summarizing or extracting from your documents
  • Answering questions about a codebase
  • Claims that must trace back to a source

Examples

  • Contract and report summaries
  • Repository architecture overviews
  • Notes from uploaded papers
Speed
Moderate
Cost
Medium
Output is similar to Chain-of-Thought, but the model reads your whole knowledge base as input, so cost grows with the size of your files.

The protocol it adds to your prompt

These steps are inserted into the prompt’s <meta_strategy> block, word for word.

  1. 1INVENTORYSurvey the material in <knowledge_base>; identify each source and what it covers.
  2. 2EXTRACTPull out the facts, claims, and structures relevant to the task, citing the exact source (file name, section, or heading) for each.
  3. 3SYNTHESIZECombine the extracted material into a coherent answer, resolving conflicts by preferring the more specific source.
  4. 4GAPSState explicitly what the task asks for that the provided material does NOT cover — never fill gaps from general knowledge.
  5. 5VERIFYCheck that every claim in your draft traces back to step 2; remove or flag any that do not.

Needs: Uploaded files or a connected GitHub repository.

ReAct (Reason + Act)

A loop of thinking, acting with a tool, and observing the result.

ThoughtActionObservationReflectAnswer

When the recommender picks it

Research or fact-grounded tasks where the answer depends on external lookups or current information, AND the prompt will run in a tool-enabled environment (search, retrieval, or APIs available). Never choose this for a single-shot prompt without tool access: the model would have to fabricate its observations; prefer CoT, Step-Back, or Grounded-Synthesis there. Assume no tool access unless the topic or context says otherwise.

Best for

  • Answers that depend on current information
  • Multi-step lookups and verification
  • Agents with search or API access

Examples

  • Market and competitor research
  • Fact-finding across several sources
  • Tool-using support agents
Speed
Slow
Cost
Medium
Every search or tool call is another round trip, so time depends on how many loops the task needs.

The protocol it adds to your prompt

These steps are inserted into the prompt’s <meta_strategy> block, word for word.

  1. 1THOUGHTState what you currently know and what is still missing.
  2. 2ACTIONRetrieve or verify the missing information from an available source.
  3. 3OBSERVATIONRecord what the source returned.
  4. 4REFLECTDecide whether to loop again or answer. Do not assert facts you have not grounded.
  5. 5ANSWERSynthesize a grounded conclusion, citing what each claim rests on.

Needs: A tool-enabled environment (search, retrieval or APIs). Without tools the model would have to invent its observations.

Not sure which one fits? Read the guide to choosing a framework.