Meta analysis
Skill Hash-7777/Hash-Medical-Reasearch-Agent-Skills/skills/meta-analysis
Drop-in agent skills that make an AI appraise medical literature like a reviewer — grade every citation against its source, run reproducible PRISMA searches, pool studies safely, and defend RAG against prompt injection. Framework-agnostic, MIT.
npx -y skills add Hash-7777/Hash-Medical-Reasearch-Agent-Skills --skill meta-analysisAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 23 days oldThe repository was created 23 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Pool effect sizes across studies only when the studies are genuinely comparable; otherwise plot the individual estimates and state why they cannot be combined. Use whenever an agent wants to produce a single summary number from multiple studies.
SKILL.md
2.8 KB, as published. Nobody here has run it
Meta-Analysis
A pooled number is powerful and dangerous. Done on comparable studies it sharpens the truth; done on studies that measured different things it manufactures a precise-looking lie. The judgment this skill enforces is not "how do I average these" — it is "should these be averaged at all?"
When to use
Whenever the agent is tempted to report one effect size (a risk ratio, an odds ratio, a mean difference, an HR) drawn from more than one study.
The rule
Only pool comparable primary studies. When they are not comparable, show them individually and say why.
Before combining, every study in the pool must share:
- Population — same clinical group, not "adults" vs "elderly with comorbidity" flattened together.
- Intervention / exposure — same drug, comparable dose and route; same exposure definition.
- Comparator — placebo vs active control are different questions; do not mix them into one pool.
- Outcome and its measure — the same endpoint, defined and measured the same way, on the same scale.
- Design tier — do not pool RCTs with observational studies into a single estimate. Report them in separate strata.
If any of these differ materially, do not pool. Plot each study's own estimate with its confidence interval (a forest-style layout) and let the reader see the spread.
Doing the pool responsibly
- Weight by precision, not by vote. Larger, tighter studies carry more weight; do not average raw numbers as if each study counted once.
- Report heterogeneity. State how much the studies disagree (e.g. an I² and the between-study variance). High heterogeneity is a reason to stop pooling, not a footnote.
- Keep confidence intervals bound to their estimate. A CI that has drifted from the number it belongs to is a parsing bug that becomes a clinical error. Verify each estimate carries its own interval.
- Prefer primary studies. Do not pool a review's summary figure together with the primary trials it already summarizes — that double-counts.
When you cannot pool
Say it plainly and show the evidence anyway:
These 6 studies cannot be pooled: 4 measured 28-day mortality, 2 measured
90-day; comparators differ (placebo vs standard care). Individual estimates:
Study A (RCT, placebo, 28d) RR 0.81 [0.70–0.94]
Study B (RCT, placebo, 28d) RR 0.88 [0.79–0.98]
Study C (obs, standard, 90d) RR 0.72 [0.55–0.95]
...
A row of honest individual estimates is more useful than one dishonest summary. The skill is knowing the difference.