uygnoey Purpose
Use this skill to plan an AI workflow for regulated domains where outputs must be verifiable against underlying documents (e.g., financial filings).
Instructions
Separate responsibilities
- Assign Claude to interpretation, ambiguity detection, and planning.
- Use deterministic components for retrieval, computation, and assembling final outputs so results are provably correct.
Build for provenance and auditability
- Ensure each number/claim can be traced to a specific source document and location (page/line item).
- Maintain audit logs and enforce security/access control at every step.
Engineer the surrounding context
- Provide structured domain knowledge: definitions, formulas, and hard boundaries.
- Maintain an ontology that maps domain concepts to precise definitions (customizable where needed).
Use multi-stage pipelines and match models to stages
- Decompose complex workflows into stages (intent decomposition → retrieval → compute → assemble → review).
- Use higher-reasoning models for ambiguity resolution and plan generation; use faster models for constrained stages.
Invest in evaluation before (and beyond) prompt tuning
- Test prompt and model changes against a large set of known-correct cases.
- Build automated evaluation pipelines that help localize failures (reasoning vs. context vs. execution).
Examples
Example: design a verifiable “answer with citations” workflow
User: Build an assistant that answers questions about quarterly results and can be audited.
Assistant:
- Claude stage: interpret question, list required data points, flag ambiguity (period definitions, restatements).
- Deterministic stage: retrieve the exact filings and tables, compute metrics, and record provenance for each value.
- Assembly stage: produce a report where each number links back to the source document location.
- Evaluation: run the workflow against a regression suite of questions with expected outputs.