xuzhougeng Indication dossier
Five research phases, each writing one waypoint JSON under
<workdir>/waypoints/, ending in a cited Markdown report. Waypoints make the
run resumable: a later invocation reads which files exist and continues from
the first missing one. The only pause for user input is after Phase 1.
The framing rule
Treat the indication as a patient population, not a disease entry. Every section answers a population question — who are these patients, how are they identified and managed, which trials would help them — rather than a textbook question about the condition. Nesting is population nesting: everyone in the child indication is in the parent.
Some inputs are not billable diagnoses at all: a biological state ("immunosenescence"), a non-accepted indication ("ageing"), an iatrogenic population ("GLP-1 induced sarcopenia"). Detect and label this early — it changes the epidemiology evidence base, the regulatory path, and what a "complete" dossier even looks like.
Inputs
| Input | Required | Meaning |
|---|---|---|
indication |
yes | e.g. "sarcopenia", "idiopathic pulmonary fibrosis" |
additional_context |
no | focus areas, parent indication, framing |
workdir |
no | waypoint/report location; default ./do_not_commit/indication-dossier-<slug>/ |
Tooling
Preferred: clinical-trials MCP for CT.gov, pubmed MCP for literature,
WebSearch/WebFetch for FDA guidance, specialty-society guidelines
(NCCN, AASLD, …), and CDC/WHO data; WebFetch for remote PDFs, Read for
local ones; Agent subagents for parallel evidence gathering. When a listed
MCP is not connected, say so and fall back to WebSearch against the public
site itself.
Run protocol
Read references/standards.md first — it defines what counts as a citable
finding, the anti-fabrication rules, and the report style. Phase-by-phase
instructions live in references/phases.md; waypoint formats in
references/waypoints.md.
- Identity. Resolve definition, ICD codes, aliases, parent, diagnostic
status; quick CT.gov landscape count. Write
meta.json. Then show the resolved identity and end the turn asking Proceed / Revise identity / Stop — the expensive phases wait for the answer (Wisp has no separate interactive-question tool, so this is a normal turn end). - Epidemiology. Case definition, prevalence/incidence, demographics,
natural history →
epidemiology.json. - Biology & standard of care. Mechanism, biomarkers, approved
therapies, guidelines, unmet need →
biology_soc.json. - Regulatory & trials. Accepted endpoints, precedents, design
parameters, landmark trials, failures →
regulatory_trials.json. - Synthesis. No new research threads (single targeted gap-fills only).
Write
indication_dossier_report.mdandresearch_output.json, then markprogress.jsoncomplete.
After each of phases 2–5, write the waypoint, emit a ≤200-word summary of findings and open uncertainties, and continue directly.
Resuming
When workdir already contains waypoints: list which phases are complete
(file exists and is non-empty), show the meta summary, and ask which phase to
run. Never overwrite an existing waypoint without confirmation.
Output layout
<workdir>/waypoints/
├── progress.json # loop control, flipped last
├── meta.json # phase 1
├── epidemiology.json # phase 2
├── biology_soc.json # phase 3
├── regulatory_trials.json # phase 4
├── sources_evaluated.json # appended by every phase
├── research_output.json # phase 5, structured
└── indication_dossier_report.md # phase 5, the deliverable