Claim Extraction

Claim Extraction

You cannot grade, compare, or pool what you have not pinned down. A paragraph of retrieved text hides its claims inside prose; a summary blurs them. This skill makes the agent extract each explicit claim as a structured item — what was found, how big the effect was, in how many people, in what kind of study, in which population — so every later step ([citation-faithfulness](../citation-faithfulness/SKILL.md), [conflict-detection](../conflict-detection/SKILL.md), [meta-analysis](../meta-analysis/SKILL.md)) has something precise to work on.

When to use

Right after retrieval, as the bridge between raw chunks and reasoning. Anything you plan to cite, contrast, or combine should first be extracted into claims.

The rule

Turn each source chunk into explicit, structured claims. Capture the finding and, whenever the text states them, the effect value, sample size, study design, and population — and keep each claim tied to its source.

For each chunk, extract one item per distinct factual claim:

1. The finding — the assertion itself, in the source's own terms ("metformin was associated with lower all-cause mortality"). Do not upgrade or soften it here; capture it as written.
2. Effect value — the number and its type when present: HR/OR/RR, mean difference, percentage, with the confidence interval bound to it.
3. Sample size — n, when stated.
4. Study design — RCT, cohort, case-control, case report, review, meta-analysis, in-vitro/animal, when identifiable.
5. Population — who the finding is about (condition, age group, setting).
6. Source link — the chunk/source id, so the claim stays traceable.

Doing it well

• Extract, don't interpret. This step records what the source says; grading whether a later claim is faithful to it is a different step. Keep them separate so the extraction stays neutral.
• One claim per item. A sentence with two findings becomes two claims. Splitting them is what lets conflict-detection and pooling work.
• Keep the effect size with its interval. An effect value that loses its confidence interval — or picks up the wrong one — becomes a downstream error. Bind them at extraction.
• Record "not stated" honestly. If the chunk gives no sample size or design, leave it empty rather than guessing. A missing field is information.
• Preserve direction and negation. "No significant difference" and "significant reduction" are opposite claims; capture the negation, don't drop it.

Output

chunk S3 →
  finding:   "metformin associated with lower all-cause mortality"
  effect:    HR 0.83 (95% CI 0.76–0.90)
  n:         12,482
  design:    cohort (observational)
  population: adults with type 2 diabetes
  source:    S3

chunk S7 →
  finding:   "no significant difference in mortality between arms"
  effect:    (p = 0.21)
  n:         640
  design:    RCT
  population: T2DM with heart failure
  source:    S7

Now the rest of the pipeline has something exact to grade, contrast, and — where appropriate — pool.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Hash-7777
• Source: Hash-7777/Hash-Medical-Reasearch-Agent-Skills
• License: MIT

Install and usage instructions live in the source repository linked above.