AgentSkillsCN

prodigy-protocol

一种以目的为导向的执行协议（“pp”），能够解析显性与隐性意图，将工作分解为可验证的步骤，运行审计循环，并在出错时及时恢复。当用户提及Prodigy Protocol/pp、要求“递归”执行，或需要高可靠性的自主任务完成时使用此功能。

SKILL.md
--- frontmatter
name: prodigy-protocol
description: Purpose-first execution protocol (“pp”) that parses explicit/latent intent, decomposes work into verifiable steps, runs an audit loop, and recovers from errors. Use when the user mentions Prodigy Protocol/pp, asks for “recursive” execution, or needs high-reliability autonomous task completion.

Prodigy Protocol

Alias: pp

Non-negotiables

  • •Follow system + repo safety rules (especially secrets handling).
  • •Be truthful about capabilities and access; do not claim persistent memory or “total authority.”
  • •Keep changes small and verifiable; prefer the smallest safe next step.

Directives (Dominion Fusion)

1) Purpose embedding

  • •Every action and output must serve a clearly stated intent.
  • •If an output does not advance the goal, omit it.

2) Directive parsing (layered intent)

  • •Extract and track:
    • •Direct intent: what the user explicitly asked
    • •Latent intent: what success likely requires but wasn’t stated
    • •Constraints: time, style, safety, repo rules, environment limits
    • •Success criteria: what “done” means
  • •If information is missing, infer reasonable defaults and document assumptions.

3) Execution architecture (task trees)

  • •Segment work into a task tree with isolated sub-tasks.
  • •Process sub-tasks independently; merge only when each is validated.
  • •Verify the chain at each stage (lint/tests/build/runtime checks as appropriate).

4) Failure resilience

  • •On any deviation/error:
    • •Capture the symptom and the immediate evidence (error text, failing check)
    • •Form a small hypothesis
    • •Apply the smallest fix
    • •Re-verify
  • •Avoid loops: track what changed between attempts.

5) Continuous audit loop

After each meaningful step, audit:

  • •Aligned? Did this advance the stated goal?
  • •Efficient? Was there a simpler/safer approach?
  • •Improvable? What small change would reduce future risk?
  • •Obsolete? Is anything now redundant or outdated?

6) Recursive learning engine (durable, not mystical)

  • •Convert one-off outcomes into reusable guidance:
    • •Project-specific: propose/update docs, .cursor/rules/, or .cursor/context/
    • •Session-specific: summarize the learning in the final response

7) Modular expansion (version-aware)

  • •Prefer current patterns used in the repo.
  • •Flag deprecated/outdated logic and propose an upgrade path.
  • •Maintain traceability: note why a module/pattern was swapped.

Quick workflow (default)

  1. •State intent: one line goal.
  2. •Parse directives: constraints + success criteria + assumptions.
  3. •Build task tree: 2–6 steps max before starting.
  4. •Execute smallest next step; validate immediately.
  5. •Audit and adjust based on evidence.
  6. •Report: what changed, why, and how to verify.