Query planning

Skill Hash-7777/Hash-Medical-Reasearch-Agent-Skills/skills/query-planning

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.

Install

npx -y skills add Hash-7777/Hash-Medical-Reasearch-Agent-Skills --skill query-planning

Assembled 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

Decompose a clinical or research question into targeted sub-queries — concepts, hypotheses, and synonym-expanded search terms — so retrieval is thorough instead of a single naive lookup. Use before searching literature for anything more than a trivial question.

SKILL.md

3.2 KB, as published. Nobody here has run it

Query Planning

A hard question asked as one flat search gets a shallow answer. "Does metformin reduce mortality in heart failure with diabetes?" is really three questions — a drug, an outcome, and a population — and each needs its own retrieval. This skill makes the agent plan the search before running it, so the evidence it gathers actually spans the question.

When to use

Before prisma-search, whenever the question contains more than one concept: a drug and an outcome, a gene and a disease, an intervention and a comparator, a population qualifier, or a time frame.

The rule

Turn the question into a plan before you search: the core concepts, the hypotheses worth testing, and a set of targeted sub-queries with their synonyms.

Produce four things:

Concepts — the distinct entities in the question (drug, outcome, population, comparator). Strip filler and generic academic words ("study", "effect", "analysis") so only real concepts remain.
Hypotheses — the specific claims the search should try to confirm or refute (e.g. "metformin lowers all-cause mortality", "the effect holds in the HF-with-diabetes subgroup").
Sub-queries — one focused search per concept pairing, not one blob. Retrieve for each, then combine.
Synonym expansion — for each concept, add the terms a paper might actually use: brand and ingredient names, abbreviations, MeSH-style variants, British/American spellings.

Doing it well

Cover every concept. If the question names a population or a comparator, a plan that ignores it will miss the studies that matter most.
Expand, don't drift. Synonyms broaden recall; unrelated terms poison it. "MI" for myocardial infarction — yes. A different disease that co-occurs — no.
One decomposition, then retrieve per sub-query. Running the sub-queries separately and pooling the results beats a single over-stuffed query that matches nothing.
Fall back gracefully. If you cannot plan richly, a plain concept split with synonyms still beats the raw question. Never let planning become a blocker.

Output

Question: Does metformin reduce mortality in HF patients with diabetes?

Concepts:     metformin | all-cause mortality | heart failure | type 2 diabetes
Hypotheses:   (1) metformin lowers all-cause mortality in T2DM
              (2) the effect holds in the HF subgroup
Sub-queries:  • metformin AND all-cause mortality AND type 2 diabetes
              • metformin AND heart failure AND mortality
              • metformin AND heart failure (safety / contraindication)
Synonyms:     metformin → biguanide, Glucophage
              heart failure → HF, cardiac failure, HFrEF, HFpEF
              type 2 diabetes → T2DM, NIDDM

Retrieve each sub-query, then hand the pooled results to the rest of the pipeline. A planned search finds the subgroup study a flat search would never surface.