Curator Delta Operation Contract

Prompt a Curator to emit only new, non-redundant playbook additions as typed JSON operations (type, section, content), never a regenerated playbook, leaving IDs and counters to deterministic system code.

The Curator in the generator–reflector–curator loop turns a Reflector diagnosis into incremental delta updates. Its prompt contract is what keeps the model from rewriting the whole context.

Inputs. The two published Curator prompts take different inputs:

  • AppWorld (Fig. 11): the question context, the current playbook, the Generator's final generated code, and the current reflections (passed as guidebook).
  • FINER (Fig. 14): the recent reflection, the current playbook, the question context, and a training-context block. That block gives the total token budget, training progress ("sample i of N"), and current playbook stats. The paper does not ablate it. Its presumable purpose is to let the Curator calibrate how much to add.

Instructions that matter.

  • Identify ONLY new insights, strategies, or mistakes that are missing from the current playbook. Add something only if it is a "perfect complement", not a restatement.
  • "Do NOT regenerate the entire playbook — only provide the additions needed."
  • Quality over quantity: a focused, well-organized playbook beats an exhaustive one.
  • Return an empty operations list when nothing is worth adding. A no-op must be a legal output.
  • The reflection used ground truth that "will NOT be available when the playbook is being used". Additions must help future predictions without that signal.
  • For coding tasks (AppWorld), carry API output-format and schema clarifications from the reflection into the playbook.
  • Respond with pure JSON, because code parses the output. The FINER prompt adds a "CRITICAL" line: valid JSON only, no markdown formatting or code blocks.

Output schema. reasoning plus operations, where each operation has type, section, and content, in both variants. In the published prompts the only operation type is ADD. Sections are named, so each addition lands in a stable region of the playbook. The AppWorld prompt shows strategies_and_hard_rules, apis_to_use_for_specific_information, and verification_checklist. The FINER prompt shows formulas_and_calculations.

Model writes content, system writes bookkeeping. The model never assigns bullet IDs or helpful/harmful counters. A deterministic merge step adds those. This keeps metadata trustworthy. Because deltas are itemized, the authors note that several can be merged in parallel. Their experiments use batch size 1. Redundancy is controlled in two layers. First, at the prompt level, the Curator is told to admit only insights missing from the current playbook and never restatements. Second, after the merge, grow-and-refine maintenance runs an embedding-based de-dup pass that catches near-duplicates the Curator let through.

Sources

  • ACE (Zhang et al.), §3, §3.2, App. F, Figs. 11, 14