Setup: chronic-illness research-partner project

This skill runs ONCE to scaffold a long-running personal research project for someone with a refractory or complex chronic illness. It does not run again per session — the persistent discipline lives in the generated repo's CLAUDE.md.

Activation

Activate when the user invokes this skill by name, OR describes wanting to:

• "set up a long-running research project on my [condition]"
• "use AI as a research partner for my chronic illness"
• "scaffold a research project for [condition]"

Do NOT activate for:

• General medical questions, including one-off questions about a condition (e.g. "what is POTS?").
• Mid-project sessions in an already-generated repo. Look for an existing CLAUDE.md referencing this skill — if found, decline and tell the user to work in the generated repo directly.
• Symptom tracking or journaling (existing condition-specific tracker apps already do this).
• Acute medical advice or triage (route to clinician/ER as appropriate).
• Diagnosis-shopping or self-diagnosis. The skill assumes the user has, or is actively pursuing with a clinician, a working diagnosis.
• General medical-literature search (PubMed, Semantic Scholar, Elicit handle one-off questions).
• Re-running on an existing project (this skill is one-shot per project; subsequent work happens inside the scaffolded repo).
• Non-chronic / acute-only conditions with no multi-year arc.
• Caregivers seeking emotional support, care logistics, or insurance navigation rather than literature synthesis. Caregiver-as-researcher is in scope; caregiver-as-general-support is not.
• Clinician workflows (case-series databases, EHR workflow, trial-screening tools need different scaffolding).
• Researching a condition the user does not have or care for.

Pre-interview preamble

Before Q1, print this paragraph and wait for acknowledgement:

Three things before we start:

1. The repo this generates will accumulate personal health information. Do not push it to a public remote without a redaction review.
2. You'll be asked about working preferences including whether to drop routine medical disclaimers. That setting relaxes routine hedging only — for any acute red flag (suicidal ideation; signs of stroke, sepsis, anaphylaxis, new neurological deficit; severe medication reactions), I'll always escalate to "call emergency services or your clinician" regardless of disclaimer setting.
3. Any answer can be skipped with skip or unknown. Skipped answers leave the relevant template section as a TODO prompt rather than seeding it. You can fill skipped items later.

Ready to start?

Wait for the user to acknowledge before continuing to Q1.

Interview

Discipline rules — apply to every question:

• One question per turn. No batching.
• Wait for the user's answer before moving to the next question.
• Do not summarize prior answers conversationally between questions. Just ask the next one.
• For enum questions, list the valid options verbatim. If the user types something that doesn't match an enum, re-prompt once with the valid options. Second invalid response: treat as skip and apply the default.
• skip and unknown are always accepted (except Q1, which is required for slug derivation).
• Do not use AI-coded vocabulary in the question text (no "surfaced", "leverage", "robust", "comprehensive" as filler, etc.). The questions below are the wording — use them as written.

Q1: Condition name

Ask: "What's the working diagnostic label, even if contested?"

• Free text. Required (cannot skip — needed to derive the directory slug).
• If user types skip or unknown, re-prompt: "I need a label to derive the project's directory name. Even a contested or working label is fine."
• Store as condition_name.
• Wait for the user's answer before moving to Q2.

Q2: Whose research

Ask: "Are you researching for yourself, or on behalf of someone else? Options: self / child / partner / parent / other."

• Enum. Default on skip: self.
• Store as whose_research.
• Wait for the user's answer before moving to Q3.

Q3: Project framing

Ask: "In one sentence, what does this project most need to crack? (e.g. a specific subtype hypothesis, a treatment to find, a refractoriness puzzle.)"

• Free text, one sentence.
• Skip behavior: emit a TODO sentence in the seed file rather than seeding it.
• Store as project_framing.
• Wait for the user's answer before moving to Q4.

Q4: History years

Ask: "How many years has this been going on? (integer, 0–100)"

• Integer 0–100.
• Skip behavior: omit the history-length sentence from the context block.
• Store as history_years.
• Wait for the user's answer before moving to Q5.

Q5: Refractory status

Ask: "Where are you on the treatment arc? Options: never-treated / partial / refractory-multi / dx-ongoing."

Brief gloss if helpful:

• never-treated: newly diagnosed, no trials yet.

• partial: some response to standard care; still seeking better.

• refractory-multi: failed multiple classes (the typical case for this skill).

• dx-ongoing: working label contested or diagnostic workup ongoing. The project organizes evidence and questions for clinicians who are still working a diagnostic case; it does not do diagnostic work itself.

• Enum. Default on skip: refractory-multi.

• Store as refractory_status.

• Wait for the user's answer before moving to Q6.

Q6: Specialist status

Ask: "Are you working with a specialist clinician for this? Options: yes / no / seeking."

• Enum. Default on skip: yes.
• Store as specialist_status.
• Wait for the user's answer before moving to Q7.

Q7: Tracker data

Ask: "Any existing tracker data we should know about? (e.g. tracker app X, 5y, tracker app Y, 18mo, none.)"

• Free text.
• Skip behavior: omit the {{tracker_data_note}} placeholder.
• Store as tracker_data.
• Wait for the user's answer before moving to Q8.

Q8: Terseness preference

Ask: "How terse do you want responses? Options: terse / balanced / fuller."

Brief gloss if helpful:

• terse: effect sizes, citations, no filler.

• balanced: concrete and specific where it matters; readable everywhere else.

• fuller: narrative explanation welcome.

• Enum. Default on skip: balanced.

• Store as terseness_pref.

• Wait for the user's answer before moving to Q9.

Q9: Disclaimer preference

Ask: "How should I handle medical disclaimers? Options: off / soft / standard."

Brief gloss if helpful:

• off: no routine "ask your doctor" reflexes (the safety floor still applies — see preamble).

• soft: include only when safety-relevant.

• standard: include routinely.

• Enum. Default on skip: off.

• Store as disclaimer_pref.

• Wait for the user's answer before moving to Q10.

Q10: Pushback preference

Ask: "How hard should I push back on weak hypotheses? Options: hard / gentle / defer."

Brief gloss if helpful:

• hard: challenge directly; argue both sides; surface confounders.

• gentle: surface concerns diplomatically.

• defer: state once and move on.

• Enum. Default on skip: hard.

• Store as pushback_pref.

• Wait for the user's answer before moving to Q11.

Q11: Primary literature comfort

Ask: "How deep do you read primary literature? Options: abstracts-and-methods / abstracts / summaries."

• Enum. Default on skip: abstracts.
• Store as primary_lit_comfort.
• Wait for the user's answer before moving to Q12.

Q12: Evidence rigor

Ask: "How strict should evidence-tagging be? Options: full / tags-only / informal."

Brief gloss if helpful:

• full: evidence tags + PMID/DOI round-trip required for every medical claim.

• tags-only: tags required; round-trip when the claim is load-bearing for a candidate.

• informal: surface the strongest evidence with citations where available.

• Enum. Default on skip: full.

• Store as evidence_rigor.

• Wait for the user's answer before m