996.fm perspective graph

Route a founder's dilemma to the debates that genuinely bear on it. Present the readings SEPARATELY — never merge them into one answer. Respect the question's altitude: scale-up material does not answer a pre-PMF dilemma. When coverage is thin it says so — advise from general knowledge rather than stretching a weak match. Returns ≤4 debates and ≤36 stances — a sample, not the full set; get_stances(debate_id) has a debate's complete stances, each with speaker, date, strength and a receipt.

{
  "type": "object",
  "properties": {
    "topic": {
      "type": "string",
      "description": "The founder's question or dilemma, with a line of context"
    }
  },
  "required": [
    "topic"
  ]
}

Every stance on one debate, grouped by reading — raw material for your own who-would-push-back read. No synthesis. Attribute and date every position (dates are episode publication dates); keep hedged claims hedged. The graph never says who opposes whom — work out yourself which positions push back on the founder's leaning, and name them.

{
  "type": "object",
  "properties": {
    "debate_id": {
      "type": "string",
      "description": "Debate id from find_perspectives or list_debates, e.g. deb-3f9c1a2b"
    }
  },
  "required": [
    "debate_id"
  ]
}

One speaker's takes in order of episode publication date (optionally on one debate) — reversals and re-assertions visible, never judged. Attribute and date every position (dates are episode publication dates); keep hedged claims hedged. A 2023 position is a 2023 position; if the same person later said otherwise, show both.

{
  "type": "object",
  "properties": {
    "person": {
      "type": "string",
      "description": "A speaker's full name as the graph records it — an ambiguous surname is refused, never merged"
    },
    "debate_id": {
      "type": "string",
      "description": "Optional: narrow the chronology to one debate"
    }
  },
  "required": [
    "person"
  ]
}

The transcript around one claim — the moment in context, with its receipt. Spoken: the episode's public transcript, windowed ±90s around the claim's timestamp. Written (an essay): the verbatim excerpt snapshotted at digest plus a link into the original; never the surrounding text, never a live fetch. Quote the receipt — episode + timestamp, or essay link + excerpt — so the founder can check the moment.

{
  "type": "object",
  "properties": {
    "claim_id": {
      "type": "string",
      "description": "Claim id from any tool that returned a card, e.g. clm-5p0ken01"
    }
  },
  "required": [
    "claim_id"
  ]
}

Browse the debate registry: id, question, readings, stance count, born (earliest publication date). Optional query: a debate matches when its question has every word of it, case-insensitive, word forms folded (hire/hiring, co-founder/cofounder). Browsing only — it does not rank against a question; use find_perspectives for that. Defaults to debates with 2+ stances, 120 rows a page — page with offset.

{
  "type": "object",
  "properties": {
    "query": {
      "type": "string",
      "description": "Words the debate question must all contain (case-insensitive; word forms folded)"
    },
    "min_claims": {
      "type": "integer",
      "minimum": 1,
      "description": "Minimum stances on the debate (default 2)"
    },
    "limit": {
      "type": "integer",
      "minimum": 1,
      "maximum": 120,
      "description": "Rows to return (default and max 120)"
    },
    "offset": {
      "type": "integer",
      "minimum": 0,
      "description": "Rows to skip — the tail line tells you the next offset"
    }
  },
  "required": []
}