discovery-query

Purpose

The PM has a specific question — "what do users say about the import flow?", "who has churned and why?", "what's the strongest pain in mid-market?" This skill answers it by reading the graph and returning only what the graph supports — with the supporting quotes attached.

Where the graph is silent, the skill says so. Where the evidence is thin (single source, no triangulation), it labels the finding accordingly. Where the question is genuinely unanswerable from the current corpus, the skill refuses to guess and names the gap as a finding in its own right.

It does NOT produce confident-sounding answers based on inference. It does not extend a single quote into a population-level claim. It does not stitch together adjacent statements into a story the sources do not themselves tell.

Procedure

1. Read graph/graph.json. If it does not exist, tell the PM to run graphify first and stop.
2. Read meta.health. Cold-start branch: if health === "thin", state this at the top of the response. Most of the answer will be gaps; that is the correct, useful result. Continue with the procedure, but do not soften the gap-heavy output.
3. Parse the PM's question. Identify:
   • The entity or relationship it asks about (a Pain, a Segment, a Feature, an Outcome, etc.).
   • Whether it is a what question (description) or a why/how question (causal). Causal questions require evidence at the edge level, not just nodes.
   • Whether it has a population scope ("most users", "mid-market", "everyone who churned"). Population claims require multi-source support.
4. Search the graph:
   • Find all nodes and edges that match the question's entities.
   • For each, collect provenance.
   • Note the source count and source-type spread behind each finding. A finding backed by three sources across two types is materially different from one backed by one ticket.
5. Classify what you found into:
   • Multi-source findings — at least 2 sources, ideally across types. These can be stated with reasonable confidence.
   • Single-source findings — one source only. State plainly; do not generalize.
   • Assumption-adjacent — the question's territory contains Assumption nodes. State that the territory is partly assumed, not evidenced.
   • Silent — the graph has no nodes or edges relevant to the question. This is itself a finding.
6. Refusal branch: if the question is fundamentally unsupportable from the corpus (e.g. asks for a quantitative claim the corpus does not contain, or asks about a population the corpus does not sample), do not fabricate. Say so plainly under "Findings", and shift all the work into "Gaps" and "Discovery questions".
7. For each gap, formulate a discovery question that would close it. Questions must be non-leading, JTBD-grounded (about behaviour and context, not hypotheticals), and concrete enough to actually go ask.
8. Emit the output contract.