Decision Check

Stop relying on shallow LLM advice for high-stakes technical and business choices. When you ask most agents for a recommendation, they give you a single, confident answer that often ignores codebase realities or critical assumptions. This skill changes that by producing a decision report that focuses specifically on what a single answer misses. It performs structured research, maps your code, and identifies the variables that would actually flip your decision.

What it does

• Codebase mapping analyzes ownership, state management, and blast radius for refactors or library migrations.
• Fact verification uses primary sources to check versioning, pricing, and known issues instead of relying on training data.
• Assumption triage identifies the 3-5 facts that would change the verdict and asks you for the ones only you know.
• Decision mapping breaks down what each option is betting on and suggests the cheapest way to test those bets.
• Pre-mortem analysis anticipates failure modes and provides warning signs to watch for after implementation.
• Independent review uses a sub-agent to re-check every claim against the code before the report reaches you.

How it works

1. Context scan triggered by phrases like decision check this or pressure-test this to identify the stakes.
2. Fact sorting to separate what is known, what needs research, and what requires user input.
3. Multi-agent research where a code reviewer or researcher validates technical claims and external facts.
4. Stress testing applying specific lenses like the "change test" to see if the design survives future requirements.

Frameworks & tools

This skill is designed for Claude Code, Cursor, and other agentic IDEs. It utilizes codebase search tools, web research agents, and filesystem access to decision-reports/.

Why this beats prompting it yourself

Standard prompts usually result in a "best guess" that sounds authoritative but lacks evidence. This skill forces a "claims ledger" approach where every technical assertion is independently verified against your actual files, preventing the hallucinated optimism typical of generic AI advice.

Use cases

• Architecture choices such as deciding between two libraries or microservice vs. monolith for a specific feature.
• Refactoring strategy to determine if a codebase migration is worth the blast radius and effort.
• Build vs Buy evaluating a specific vendor or tool against a custom internal implementation.
• High-stakes logic stress-testing a critical path or security-sensitive design before writing code.

Known limitations

Requires an environment with file-read and search capabilities. Best results require a sub-agent tool for independent review; without one, the agent performs a manual second-pass check.