Overriding AI Prompts

Behind the plan interview, Learning Mode, the final-review diff helpers, the AI PR review, and the review-memory bootstrap/compaction, AMF makes one-shot ("headless") AI calls with prompts it builds for you. Press E on the dashboard — or Ctrl+Space, then E from a session — to open the prompt-override manager. It lists every template with its effective source (built-in, feature, project, or global) and [F][P][G] flags for which scopes already carry an override.

Enter or e opens an editor on the effective template. Ctrl+S moves to a scope picker, then a harness picker, then saves:

The nearest scope wins — feature → project → global → built-in — and within the winning scope a per-harness template beats the shared one. d, d clears the effective override. Templates carry visible {{token}} placeholders that AMF re-fills with live context (the diff, the question, the interview answers) each time the prompt runs. There is no validation: if you delete a required token or add one AMF does not supply, it is saved and rendered exactly as written.

Project-scope overrides sit in amf.json under prompt_overrides, keyed by the stable prompt id (shown in the manager):

{
  "prompt_overrides": {
    "pr_review.ai_review": {
      "template": "You are reviewing a diff... {{annotated_diff}} ..."
    },
    "learning.answer": {
      "template": "shared text with {{question}}",
      "harnesses": { "codex": "codex-specific text with {{question}}" }
    }
  }
}

See Project Config Reference for everything else amf.json can declare.

Before each user-initiated headless call, a pre-call notice names the prompt and target harness: v shows the exact rendered prompt, e jumps to the manager for that prompt, Enter makes the call, Esc cancels it. Calls that run without you watching — the Learning Mode answer queue and session summaries — announce with a toast instead of the modal.