AI 原生开放教材

外观

AI 工程从零到一（中文）源码研读·中文原文 ↗原件 ↓仓库 ↗

AI 工程从零到一（中文） ​

Given a problem class (policy-bound workflow vs code optimization vs open-ended task), pick a planner and produce a correct scaffold.

Decision:

  1. Does the problem have hard preconditions / policy / scheduling constraints? -> HTN (ChatHTN).
  2. Does the problem have a deterministic, machine-checkable fitness function? -> Evolutionary (AlphaEvolve).
  3. Neither? -> Reach for ReAct (Lesson 01) or ReWOO (Lesson 02) instead.

For HTN, produce:

  1. Operator type with preconditions, effects_add, effects_remove.
  2. Method type with task, preconditions, subtasks.
  3. A planner that tries methods first, falls back to LLM decomposition, and caches successful LLM decompositions.
  4. A validation step that rejects LLM decompositions referencing unknown operators or methods.

For Evolutionary, produce:

  1. A seed population of candidate programs.
  2. A deterministic evaluator returning a scalar fitness.
  3. A mutation operator (LLM-driven or rule-based).
  4. A selection loop (keep top-k, mutate, repeat) with early stopping.

Hard rejects:

  • ChatHTN where LLM output is applied directly without operator-schema validation. The soundness claim fails.
  • AlphaEvolve where the evaluator calls an LLM judge. Fitness must be deterministic; LLM judges introduce stochastic noise the loop cannot recover from.
  • Either pattern for open-ended tasks ("write a blog post"). No evaluator, no preconditions -> use ReAct.

Refusal rules:

  • If the domain has no clear operator schema, refuse ChatHTN. Suggest ReWOO or plain ReAct.
  • If the domain has no machine-checkable fitness, refuse AlphaEvolve. Suggest Self-Refine (Lesson 05).
  • If the user wants "planner + LLM makes final call," refuse. The split between symbolic correctness and LLM exploration is load-bearing.

Output: operators.py, methods.py, planner.py (HTN) or evaluator.py, mutator.py, loop.py (evolutionary), plus README.md with the decision rationale. End with "what to read next" pointing to Lesson 25 if debate-style verification fits the problem, or Lesson 02 if the task is actually ReWOO-shaped after all.