[LifeSci Skill] Bound target-evidence retrieval

plugins/life-science-research/skills/clinicaltrials-skill/SKILL.md

@@ -1,8 +1,41 @@
 ---
 name: clinicaltrials-skill
-description: Submit compact ClinicalTrials.gov API v2 requests for study search, metadata, enums, search areas, and field statistics. Use when a user wants concise ClinicalTrials.gov summaries
+description: Submit compact ClinicalTrials.gov API v2 requests for study search, metadata, enums, search areas, and field statistics. Use for concise registry summaries and bounded target-program research paired with PubMed.
 ---
 
+## Target-program workflow
+When asked for target biology, prior programs, and safety, especially when this
+skill is paired with `ncbi-entrez-skill`:
+
+1. Run one `action=studies` wrapper call per target with `max_pages=1` and no
+   more than ten compact records. Set `save_raw=true` with a unique temporary
+   output path on this first call.
+2. Search target names and aliases in interventional studies. Keep completed
+   and terminated programs rather than filtering to currently recruiting
+   studies.
+3. Rank records by evidence value: posted results or mature outcome-bearing
+   programs first, then a negative or inconclusive program, then active
+   no-results programs.
+4. Retain at most five distinct programs per target. Do not treat termination
+   for business, accrual, or reprioritization as evidence of toxicity or target
+   failure.
+5. Extract intervention names, aliases, NCT IDs, phase, enrollment, status,
+   outcome signal, and safety signal. Feed intervention aliases into the
+   companion PubMed clinical-program search.
+6. Do not rerun the same registry query with a larger `max_depth` or a
+   different output projection. When compact output lacks a needed field,
+   reshape the saved response locally in one pass. Project only NCT ID, title,
+   status, stop reason, phase, enrollment, intervention names, `hasResults`,
+   primary outcome values, and compact serious-adverse-event terms or counts.
+7. Do not make direct-study follow-up requests in this bounded pass. Mark a
+   field unavailable when the saved response cannot support it.
+8. Stop when the mature positive, negative or inconclusive, and observed-safety
+   slots are covered or explicitly unavailable.
+
+Do not inspect wrapper source, probe Python environments, or use web search to
+re-verify registry results unless the documented invocation fails, records
+conflict, or the user explicitly requests current regulatory status.
+
 ## Operating rules
 - Use `scripts/clinicaltrials_client.py` for all ClinicalTrials.gov v2 calls.
 - Study searches are better with `max_items=10` and `max_pages=1`; only increase pages when the user explicitly wants more than the first page.

plugins/life-science-research/skills/ncbi-entrez-skill/SKILL.md

@@ -1,8 +1,50 @@
 ---
 name: ncbi-entrez-skill
-description: Submit compact NCBI Entrez E-Utilities requests for PubMed, Gene, Protein, Nucleotide, PMC metadata, and GEO metadata workflows. Use when a user wants concise Entrez search, fetch, summary, or link results; save raw JSON or XML only on request.
+description: Submit compact NCBI Entrez E-Utilities requests for PubMed, Gene, Protein, Nucleotide, PMC metadata, and GEO metadata workflows. Use for concise Entrez search, fetch, summary, or link results and for bounded target-evidence research paired with ClinicalTrials.gov.
 ---
 
+## Target-evidence workflow
+When asked for target biology, prior programs, and safety, especially when this
+skill is paired with `clinicaltrials-skill`:
+
+1. Plan the complete cross-source retrieval before invoking either wrapper.
+2. Create these evidence slots for each target:
+   - human biology or pharmacodynamic target validation;
+   - normal adult-tissue expression or translational safety;
+   - the most mature positive clinical program;
+   - one negative, terminated, or inconclusive program with outcome evidence
+     when available;
+   - observed human safety;
+   - preclinical mechanism, efficacy, or safety.
+3. Run the bounded ClinicalTrials.gov search first and reuse its intervention
+   names and aliases in the PubMed clinical-program query.
+4. Run at most two PubMed `esearch` calls per target: one for biology,
+   expression, normal tissue, and preclinical safety; one for named clinical
+   programs, efficacy, pharmacodynamics, and observed safety.
+5. Prefer primary studies with outcomes. Exclude editorials and comments from
+   the clinical query when possible.
+6. Select at most six papers per target. Across all targets, fetch selected
+   PMIDs in one batched `efetch`; do not add `esummary` when `efetch` supplies
+   the needed citation metadata.
+7. Keep NCBI calls sequential. Use compact `retmax`, `max_items`, and
+   `max_depth` values and avoid returning full objects or large abstract sets.
+8. Stop when every available evidence slot is filled. State unavailable slots
+   explicitly instead of expanding into open-ended searches.
+
+For one target, budget one ClinicalTrials.gov wrapper call, two PubMed
+`esearch` calls, and one batched `efetch`. For multiple targets, add one
+registry call and two searches per target but keep one shared `efetch`. Do not
+rerun an identical request to change formatting or selected fields. If fuller
+local inspection is needed, set `save_raw=true` on the first call and inspect
+the saved response without calling the network wrapper again. Wrapper failure
+is the only exception.
+
+For PubMed XML fetches, pair `retmode=xml` with `response_format=xml`, set
+`save_raw=true`, and parse the saved XML in one local pass. Do not inspect
+wrapper source, probe Python environments, or use web search to re-verify
+PubMed results unless the documented invocation fails, records conflict, or
+the user explicitly requests current regulatory status.
+
 ## Operating rules
 - Use `scripts/ncbi_entrez.py` for all Entrez calls in this package.
 - Use explicit `endpoint` values such as `esearch`, `esummary`, `efetch`, `elink`, or `einfo`.