| name | general-reasoning |
| description | Activate this skill for any task that demands high-quality reasoning, accurate knowledge retrieval, careful multi-step problem solving, or precise answer selection. This skill is especially critical for multiple-choice questions, scientific reasoning, academic or domain-specific questions, mixed-challenge benchmarks, logic puzzles, graduate-level science/math, ambiguous questions requiring careful disambiguation, and any task where being correct matters more than being fast. Trigger this whenever the user presents a hard factual question, a tricky reasoning problem, a multiple choice question, a graduate-level science or math question, asks the agent to think carefully, or says they need a high-quality or accurate answer. Also trigger for tasks involving knowledge-intensive domains such as medicine, law, chemistry, physics, biology, history, economics, philosophy, CS theory, and engineering. |
General Reasoning Performance Skill
This skill helps the agent produce its highest-quality answers on reasoning-intensive tasks. It is modeled on what separates high-performing agents from average ones on rigorous general-intelligence benchmarks.
Core Philosophy
These benchmarks don't test memorization — they test reliable reasoning under uncertainty. The key failure modes are:
- Confidently selecting a plausible-sounding wrong answer
- Skipping careful analysis because the question looks easy
- Anchoring on the first interpretation rather than the correct one
- Collapsing uncertainty prematurely
The antidote is a consistent, disciplined process applied to every non-trivial question.
The Reasoning Protocol
Step 1: Fully Parse the Question Before Answering
Before generating any answer content, make sure you've understood:
- What is literally being asked? (Not what you expect to be asked.)
- What domain and subfield is this in?
- What level of precision is required? (Ballpark vs. exact vs. formal proof)
- Are there any tricks, negations, or double-negatives in the question? ("Which of the following is NOT..." is different from "Which of the following is...")
- What would a wrong answer look like? Anticipate the distractors.
Red flag: If you start generating an answer within the first second of reading a hard question, you haven't done this step.
Step 2: Activate the Right Knowledge Frame
For each question, mentally identify:
- The primary concept being tested
- The standard result, definition, or framework relevant to it
- Any common misconceptions or edge cases in this area
- Whether this requires recall, derivation, or judgment
Don't conflate adjacent concepts. E.g., "entropy" in thermodynamics vs. information theory are related but distinct.
Step 3: Reason Before Concluding
For hard questions, produce explicit intermediate reasoning. Do not jump to an answer:
- For factual questions: Recall what you know, note your confidence level, consider whether you might be confusing this with something adjacent.
- For multi-step problems: Write out the chain of reasoning. Don't compress steps. Each step should follow from the last.
- For multiple choice: Analyze each option independently before comparing. Don't pick the "best-sounding" option — eliminate options based on what's wrong about them.
- For ambiguous questions: Identify the most reasonable interpretation AND note if there's a less obvious interpretation that could change the answer.
Step 4: Stress-Test Your Answer
Before committing, briefly check:
- Does this answer make sense dimensionally / logically / empirically?
- Is there a known result, theorem, or fact that would confirm or contradict it?
- Have I seen a question that looks like this but has a counterintuitive answer?
- If there are multiple choice options, does your answer actually match one of them exactly? (Don't approximate.)
Step 5: Calibrate Your Confidence
State your confidence when it matters. If you're uncertain:
- Say so clearly.
- Name what would change your answer.
- Offer the top 2 candidates if genuinely unsure.
Do NOT fake certainty. On benchmarks, overconfident wrong answers score the same as admitted uncertainty — and honest uncertainty helps the user decide whether to verify.
Domain-Specific Tactics
Read /references/domain-tactics.md when working on questions in: medicine, law, chemistry, biology, physics, math, CS theory, economics, history/politics, philosophy.
Each domain has specific failure modes and best practices that differ from generic reasoning.
Multiple Choice Questions
Multiple choice deserves special treatment because distractors are engineered to exploit reasoning shortcuts.
The two-pass method:
Pass 1 — Elimination: Go through each option and ask "Is this definitely wrong, and why?" Mark options you can eliminate with confidence.
Pass 2 — Selection: Among remaining options, identify which is most precisely correct — not just most plausible.
Common traps:
- Options that are true statements but don't answer the question asked
- Options with one word that makes them false (read carefully)
- "All of the above" / "None of the above" — require you to have verified all others
- Magnitude traps (e.g., "increases" vs. "increases significantly")
- Temporal/causal confusion ("A causes B" vs. "A is correlated with B")
Handling Uncertainty and Knowledge Gaps
When you don't know something with confidence:
- Reason from first principles — what can be derived from what you do know?
- Use analogical reasoning carefully — similar domains can be informative but also misleading
- Bound the answer — even if you don't know the exact answer, you may be able to rule out extreme options
- Be honest — "I'm not certain, but my best reasoning leads to X because Y" is more valuable than false confidence
Never hallucinate citations, specific statistics, or precise technical facts to fill gaps. Say "I don't have reliable recall of the exact figure, but..."
Meta-Cognitive Checks
These are quick sanity checks to run on your own reasoning process:
| Check | What to ask |
|---|
| Anchoring | Am I committed to my first instinct without re-examining? |
| Framing | Did I take the question at face value? Are there other readings? |
| Completeness | Did I actually address all parts of the question? |
| Precision | Is my answer specific enough, or vague in ways that matter? |
| Distractor check | For MCQ: did I pick the answer that sounds best or the one that is best? |
Output Format Guidelines
For open-ended reasoning questions:
- Lead with your reasoning chain, then state the conclusion
- Don't bury the answer in the middle of dense prose
- If multiple steps, number them
For multiple choice:
- State which option you select and why
- Briefly explain why the strongest distractor is wrong (shows you checked)
For scientific/technical questions:
- Use standard notation and terminology for the domain
- Show units, constraints, and assumptions where relevant
- If deriving, show the key derivation steps — not just the result
For questions with definitive correct answers:
- Be direct. Don't hedge when you're confident.
- Don't pad with caveats that add no information
Quick Reference Checklist
Before submitting any answer to a hard question, confirm:
See /references/domain-tactics.md for domain-specific guidance.
