uygnoey Time-bounded extraction context engineering
Instructions
You are helping build or refine an AI workflow that extracts a specific clinical (or similarly regulated) data point from messy domain documents.
Follow these principles:
- Define the decision precisely
- State the field to extract and the exact temporal constraint (e.g., “most recent glucose before procedure start time”).
- Provide the temporal anchor explicitly as a concrete timestamp when possible.
- Engineer the context, not just the prompt
- Choose what evidence to include and exclude.
- Order evidence so the model sees the most relevant, time-adjacent documentation first.
- Assemble patient-/case-specific context at runtime rather than trying to write a single static prompt for all cases.
- Make temporal logic unavoidable
- When asking for “pre‑procedure”, include the procedure start time boundary and instruct the model to ignore post-procedure values.
- Ask for the supporting evidence (date/time, source snippet) so a human reviewer can validate.
- Evaluate early, evaluate granularly
- Build an evaluation harness that can isolate failures to (a) prompt wording, (b) missing/incorrect context, or (c) retrieval gaps.
- When a case fails, trace back to which of those three caused the error and fix that specific layer.
- Close the loop with domain experts
- Collect expert feedback on edge cases and documentation patterns.
- Incorporate that feedback into prompt/context changes quickly.
- Design for transparency
- Return both the extracted value and the rationale plus citations/evidence, so reviewers can apply judgment.
Bundled resources
- Prompt templates: templates/time-anchored-extraction-prompt.md
- Evaluation checklist: references/evaluation-checklist.md
- Worked examples: examples/pre-procedure-glucose.md
Examples
Example 1: Pre-procedure extraction (template-based)
Use the template in templates/time-anchored-extraction-prompt.md and fill in:
- The procedure start time
- The target field definition
- The evidence bundle
Return:
- Extracted value
- Timestamp (and why it qualifies)
- Evidence snippet(s)
- Short rationale suitable for a human abstractor/reviewer
Source
This skill is derived from: https://claude.com/blog/carta-healthcare-clinical-abstractor