Focused Cot Trace

Focused Chain-of-Thought Reasoning Trace

Goal

Use Focused Chain-of-Thought to answer the user's question through organize essential information first, then reason exclusively over the structured context, not through a generic checklist or interchangeable trace. The trace should extract the key facts, constraints, and principles from the query into a structured context block, then reason step-by-step exclusively from that context, reducing token waste on irrelevant detail, so the visible reasoning follows the same path the strategy is known for. The goal is to create a public scratchpad that a reviewer can audit without exposing hidden private chain-of-thought. Center the scratchpad on subquestions, assumptions, evidence, contrasts, tests, intermediate conclusions, and implications, because those artifacts make Focused Chain-of-Thought useful rather than decorative. Preserve the user's domain, constraints, definitions, and evidence standards so the trace stays tied to the actual task. Keep uncertainty visible by naming weak assumptions, missing evidence, rival interpretations, and confidence changes as they arise. Write the result to root memory/{question_name}.md so the reasoning trace becomes a durable project artifact.

Intent

Focused Chain-of-Thought trace is invoked when the user's query contains substantial extraneous information — background context, narrative framing, or redundant detail — that would cause standard chain-of-thought to waste tokens restating and re-navigating the verbatim query at every reasoning step. Standard CoT tends to re-read and re-express the original prompt throughout the chain, creating token bloat that dilutes the reasoning signal. Focused CoT eliminates this by inserting a structured organization phase before reasoning begins: the model extracts the essential facts, constraints, and principles into a compact context block, then reasons exclusively from that block without repeatedly referencing the original verbose query.

A user would select this trace over a generic trace when the input is long or information-dense — legal documents, technical specifications, research questions with extensive background — and they want the reasoning to be efficiently traceable to specific extracted facts rather than swimming in the original text. The strategy's value proposition is two-fold: it reduces token consumption by 2-3x compared to standard CoT while preserving reasoning quality, and it produces a cleaner audit trail because each reasoning step references a numbered context element rather than a sprawling paragraph. This makes the trace especially suitable for professional environments where conciseness and verifiability are both valued — the reasoning must be correct, and the path to correctness must be easy to inspect.

Instructions

Derive {question_name} from the user's actual question by lowercasing it, replacing non-alphanumeric runs with hyphens, trimming extra hyphens, and using reasoning-trace if no safe name remains. Create the root memory directory when needed, then write or replace memory/{question_name}.md with this trace. Start the file with the question, selected strategy, scale note, source constraints, and a brief statement of what the trace will inspect. Build the scratchpad by repeatedly applying the Focused Chain-of-Thought move: extract and structure the essential information from the query into a concise context block, then reason exclusively over that organized context. Use a medium-depth trace by default, usually around 100 numbered lines, unless the user asks for a different depth. Prefer concise public reasoning artifacts over hidden deliberation, and make every numbered item contribute a question, observation, test, comparison, inference, or synthesis. End the file with a synthesis and final answer that follow from the trace, including any important uncertainty that remains.

Background Information About The Reasoning Strategy

Focused Chain-of-Thought first organizes essential information from a query into a concise, structured context and then guides the model to reason exclusively over this context. This training-free, input-centric approach reduces the number of generated tokens by 2-3x compared to standard chain-of-thought prompting while preserving strong reasoning performance. Research from 2024 shows that a large fraction of reasoning tokens in standard CoT are redundant restatements ΓÇö Focused CoT eliminates this redundancy by separating organization from reasoning. The key insight is that the model benefits from a clean, structured context more than from re-reading the original verbose query. It should give the output document a visible organized context block followed by reasoning that references specific context elements, so the reader can see how the answer was reached without exposing hidden private chain-of-thought. Use it when the problem benefits from the core move "extract and structure the essential information from the query into a concise context block, then reason exclusively over that organized context" and when the final answer needs efficient reasoning that avoids redundant restatements and stays grounded in extracted essentials. A strong Focused Chain-of-Thought trace shows the organized context clearly, makes every reasoning step reference a specific context element, and delivers concise but complete reasoning.

Algorithm for the output document:

1. Restate the user's question, constraints, and evidence standard, then extract and organize the essential facts, constraints, and principles into a concise structured context block.
2. Verify that the organized context captures all information relevant to answering the question ΓÇö flag anything omitted.
3. Reason step-by-step exclusively from the organized context, referencing specific context elements by number for each reasoning step.
4. If the organized context proves insufficient for a step, expand it with the missing information and continue, noting the gap.
5. Synthesize the completed trace into the final answer, showing how the conclusion follows from reasoning over the organized context rather than the original verbose query.

Implementation Details

Focused Chain-of-Thought was introduced by Lee et al. (2024) in "Focused Chain-of-Thought: A Training-Free Approach for Efficient Reasoning," which demonstrated that separating information organization from reasoning reduces token consumption by 2-3x compared to standard CoT while preserving or slightly improving accuracy on reasoning benchmarks. The key empirical finding is that 40-60% of tokens in standard CoT responses are redundant restatements of information already present in the prompt — the model spends significant capacity re-expressing what it already has access to rather than generating new reasoning. Focused CoT eliminates this redundancy by front-loading extraction into a discrete organization step.

The technical mechanism is surprisingly simple: before any reasoning begins, the model produces a structured context block with numbered elements (facts, constraints, definitions, principles) extracted from the query. Each element is a single, concise statement representing exactly one piece of essential information. The model then reasons step-by-step, but instead of referencing the original query, each reasoning step cites specific context elements by number (e.g., "From context element 3, we know X, therefore Y"). This does two things: it prevents the model from re-expressing the query verbatim, and it creates a verifiable link between each reasoning move and its informational basis. The paper showed that this approach is training-free — it works with any instruction-tuned LLM without fine-tuning.

For Vidbyte reasoning traces, Focused CoT requires an explicit "Organized Context" block near the top of the trace, containing numbered elements that capture all essential information from the user's question. The reasoning section that follows must reference these elements by number for every inferential step. If the organized context proves insufficient — if the model discovers it needs information that was not extracted — the trace must show the context being expanded (with a new element number) and the reasoning continuing. The final synthesis should demonstrate that the conclusion follows from the organized context, not from re-processing the original verbose query.

The broader context includes work on retrieval-augmented generation (Lewis et al., 2020, RAG) and context distillation (Askell et al., 2021) which similarly separate information access from reasoning. Focused CoT differs by performing the extraction inline within a single generation rather than relying on external retrieval systems, making it simpler to deploy while achieving similar efficiency gains.

Output Information

Write the scratchpad as Markdown in root memory/{question_name}.md before responding to the user. Include this scale note near the top of the file: "Scale: default - aim for around 100 numbered lines, or roughly 2,000 to 3,500 tokens, while adapting to the real complexity of the question." Use numbered scratchpad items for scanability, but treat the number target as approximate and subordinate to usefulness. Keep the scratchpad public, inspectable, and concise enough per line that the structure remains easy to review. After writing the file, respond with the path, selected strategy, scale note, and final answer summary.

Specify Files And Length And Structure Of Output

Write the artifact to memory/{question_name}.md at the repository root, using the literal memory directory name. Structure the file with the sections Question, Strategy, Scale, Scratchpad, Synthesis, and Final Answer. The Scratchpad section should target around 100 numbered lines or roughly 2,000 to 3,500 tokens of public scratchpad detail, adjusted reasonably for very small or unusually broad questions. Use subsections inside Scratchpad when the trace becomes long enough that phases, branches, hypotheses, cases, or criteria would improve readability. If the user supplies a domain format, evidence source, or output constraint, preserve it inside this structure while keeping the default scale approximate.

Things Not to Do

• Do not skip the organized context block and jump directly to reasoning — the context block is the strategy's defining structural feature, and without it the trace is just standard chain-of-thought.
• Do not extract context elements that are trivial restatements with no informational content — each context element must capture a distinct fact, constraint, or principle that is actually used in reasoning.
• Do not let reasoning steps float without citing specific context element numbers — every inferential move must be traceable to a numbered element, because that traceability is the primary audit mechanism.
• Do not over-expand context elements when the original query is already concise — if the user's question is a simple one-liner, forcing a verbose extraction defeats the efficiency purpose.
• Do not let the organized context grow to include information not present in the original query — extracted elements must be faithful to the source material, not the model's own inferences or assumptions.
• Do not produce a trace where the context block has more tokens than the original query — if the focused version is longer than the original, the strategy is failing at its core purpose.
• Do not allow reasoning steps that claim context support without quoting the specific element — a bare citation number without the referenced content makes the trace unverifiable without cross-referencing.
• Do not write the trace to a location other than memory/{question_name}.md at the repository root.