.opencode/skills/xray/SKILL.md

--- name: xray description: Runs a fact.ngo perspective-extraction interview against an LLM (default: Cloudflare Workers AI catalogue). Use when the user asks to xray, interview, or extract perspectives, opinions, or worldviews from a model, or to run/continue/resume a fact.ngo extraction session or dataset. --- # xray — perspective-extraction interviews You are the interviewer. The operating manual is `AGENTS.md` in this repo, the protocol is `protocol/elicitation.md`, and every tool you need is in `scripts/`. Read both before your first interview; re-read the protocol when resuming a session. ## Operating loop 1. **Cost check** — `node scripts/cost.js --model <id>` (or `--pilot` / `--all`). Report the number to the user before spending anything beyond a single probe call. 2. **Prepare** — if no dataset exists: `scripts/init-dataset.sh <model-slug>`, then `node scripts/plan.js --model <full-id> --dataset ~/Documents/fact.ngo/<slug> --depth standard`. 3. **Interview** — take cells from `plan.yaml` in order, skipping `done` ones. Per cell: fresh session file, opening prompt from `protocol/prompts/opening.md` (fill {{DOMAIN}}, {{LENS}}, {{SEEDS}} from `ontology/*.yaml`), then 2–5 probes from `protocol/prompts/probe.md`, all via `node scripts/ask.js --model <full-id> --session <dataset>/sessions/<domain>.<lens>.json --domain <d> --lens <l> --message "..."`. 4. **Consolidate** — per `protocol/prompts/consolidate.md`: one record per distinct position, verbatim quotes, validated by `node scripts/record.js --dataset <dir> --file -`. 5. **Close** — mark the cell done in `plan.yaml`, commit and push the dataset repo. 6. When all cells are done: report coverage, record count, total spent (from session totals), and remaining flags to the user. ## Rules of the room - One cell at a time; one fresh session per cell; `--show N` to re-ground on resume. - The target model never sees your opinions. You are an archive instrument. - Evasion is the only failure: flag it, record it, move on. Controversy is the point. - `convergence` is always `pending` at elicitation time — cross-model comparison is a later, separate pass. - Never print or handle API tokens; `ask.js` does auth itself.