Skip to main content
The Ontora MCP server ships eight MCP prompts that walk an AI assistant through a complete, defensible campaign evaluation. Each step builds on the previous one’s output, and every prompt already encodes the counting and privacy rules that make the numbers trustworthy:
  • a person counts once per topic cluster, even when they raise the topic at several questions,
  • person counts and example counts are always reported separately,
  • no verbatim quotes from privacy-protected interviews (they still count in aggregates),
  • “nicht genannt” (not stated) always beats invention — no guessed numbers, no invented sources.
The prompt content is written in German; run the chain with any model that handles German well.

​
The eight steps

interview_id is the campaign UUID (from list_campaigns). question_ids lets you pin the exact question IDs for the closed questions (step 1) or the single-priority question (step 4); leave it out and the prompt instructs the assistant to map them via get_interview_guide itself. Steps 2 and 3 deliberately run the durable full-corpus analysis instead of semantic search: it processes every eligible interview exactly once, so the counts are population statements, not “the most relevant hits”. For extraction-heavy steps (1–4) use the strongest reasoning model you have; for the writing steps (6–7) the model choice barely matters.

​
Running the chain from Claude

Claude Code, Claude Desktop, and other MCP clients with prompt support discover the prompts automatically once the Ontora server is connected:
  1. Pick campaign_analysis_step_0 from the client’s prompt menu (in Claude Code, prompts appear as /ontora:campaign_analysis_step_0 style commands) and fill in the interview_id.
  2. Let the assistant finish the step and save the output before moving on — each step consumes the previous results, and a saved copy lets you restart cleanly if the context drifts.
  3. Continue through step 7 in order. Steps 6 and 7 only reshape earlier results, so run them in the same conversation as (or with pasted output from) steps 0–5.

​
Running the chain from LangDock

LangDock’s remote-MCP integration currently exposes tools; if your LangDock version does not list server prompts, fetch a prompt’s text once from any MCP client (or ask us for the text) and paste it as the message to your LangDock assistant — the Ontora tools it references are all available through the LangDock integration. The chain works identically; only the prompt delivery differs. Because steps 2 and 3 call start_full_corpus_analysis, Claude’s /mcp/directory connection requires an Ontora workspace admin. On the legacy /mcp surface, the API key needs full access (a read-only key covers every other step).