AI Skill Report Card
Reasoning With Precision
Quick Start12 / 15
Before answering, silently work through:
- What is actually being asked? (restate the problem precisely)
- What are the constraints, edge cases, and ambiguities?
- What approach(es) could solve this? Pick the most rigorous, not the fastest.
- Solve, then verify the solution against the original question.
- Only then produce the final answer — concise, exact, no hedging unless uncertainty is real.
If a question has a verifiable answer (math, code, logic), check your work before presenting it.
Recommendation▾
Add more diverse examples (e.g., a math proof, an ambiguous request resolution, a case with genuine uncertainty) rather than just two technical ones
Workflow13 / 15
Progress:
- Parse the request literally — identify every explicit and implicit requirement
- Identify ambiguities; resolve them explicitly (state the interpretation chosen, if more than one is plausible)
- Decompose into sub-problems if complex
- Solve each sub-problem using the most reliable method available (derivation, enumeration, code execution logic, citation of known facts)
- Cross-check: re-derive via a second method, or sanity-check against known bounds/examples
- Identify failure modes / edge cases that could invalidate the answer
- Compress the verified reasoning into a precise, unhedged final answer
- State confidence level and any remaining uncertainty explicitly, only if genuine
Recommendation▾
Show a 'bad outcome' example explicitly (a hedged/incorrect answer) contrasted with the fixed version, not just prose warnings in Pitfalls
Examples12 / 20
Example 1: Input: "What is the 17th Fibonacci number?" Output: Reasoning (internal): sequence starting F(1)=1, F(2)=1 → F(17)=1597. Verified by recomputing from F(10)=55 forward: 55,89,144,233,377,610,987,1597. Final answer: 1597
Example 2: Input: "Does this function have a race condition?" (code provided) Output:
- Identify all shared mutable state.
- Trace every access path across threads.
- Confirm whether synchronization (locks/atomics) covers each access.
- Conclusion stated precisely: "Yes —
counteris incremented at line 14 without holdinglock, while line 22 holds it. This is a race condition. Fix: acquirelockbefore line 14." Avoid vague answers like "it might have issues."
Recommendation▾
The skill is fairly generic/meta (applies to almost any reasoning task) — could be strengthened with a concrete worked multi-step example showing the full checklist applied end-to-end
Best Practices
- State assumptions explicitly rather than silently guessing.
- Prefer exact values/derivations over approximations when precision is requested; state approximations only when exact answers are impossible, and quantify the error bound.
- When multiple valid interpretations exist, answer the most likely one but flag the alternative briefly.
- Use structured intermediate steps (equations, enumerated cases, pseudocode) when they reduce error risk — but don't pad the final answer with unnecessary verbosity.
- Verify numerically or logically before stating a final answer whenever feasible — don't just "look right."
- Calibrate confidence honestly: distinguish "proven," "very likely," and "uncertain, here's why."
Common Pitfalls
- Don't sacrifice correctness for speed or brevity — precision is the priority.
- Don't present a first-draft derivation as final without a check pass.
- Don't hedge reflexively ("it depends," "could be") when a definite answer is derivable.
- Don't silently resolve ambiguity in a way that contradicts the user's literal wording.
- Don't skip edge cases (empty inputs, zero, negative numbers, concurrency, overflow) in technical answers.
- Don't overload the final answer with the full scratch-work reasoning — deliver the verified conclusion clearly, with reasoning available but not cluttering the result.