LLM Wiki

Appearance

Built 26/04/18 17:24commit 5ccb4ff

ShareAI Lab Learn Claude Code ​

中文 | English

Summary ​

shareAI-Lab/learn-claude-code is now represented in the vault as a GitHub-backed raw source family plus a bilingual source-page pair. The repository argues that agency comes from the trained model, while product engineering mostly builds the harness around the model: tools, context controls, task systems, skills, background execution, team protocols, and worktree isolation.

Source Family ​

  • Repository: https://github.com/shareAI-Lab/learn-claude-code
  • Raw files currently staged in the vault:
    • raw/github/shareai-lab/learn-claude-code/README.md
    • raw/github/shareai-lab/learn-claude-code/README.zh.md
    • raw/github/shareai-lab/learn-claude-code/README.ja.md
    • raw/github/shareai-lab/learn-claude-code/s_full.py
  • The current raw pair captures the repository's central editorial claim and the capstone full-reference implementation surface.
  • The raw README siblings are partial preserved excerpts for this corrective ingest pass, not yet a full repository mirror.

What This Source Adds ​

  • A forceful articulation of Agent Product = Model + Harness.
  • A staged harness curriculum that separates loop, tool dispatch, planning, subagents, skill loading, context compaction, task systems, background jobs, team protocols, and worktree isolation into explicit mechanisms.
  • A Claude-Code-centered argument against overbuilt prompt-plumbing systems that try to proceduralize intelligence instead of improving the model's operating environment.
  • A concrete teaching artifact for comparing Claude Code, Codex, OpenClaw, and other agent-first repository patterns.

Key Claims ​

  • The model is the real agent, and harness code is the environment layer that lets the model perceive and act.
  • Most practical “agent engineering” is actually harness engineering rather than intelligence creation.
  • Claude Code matters as a reference because it gives the model tools, context management, permissions, and execution surfaces without replacing the model's judgment with rigid workflow graphs.
  • Planning, compression, background execution, teammate coordination, and worktree isolation are best understood as modular harness layers, not proof that the surrounding code itself is the agent.

Structure Notes ​

The agents/ tree presents one progressive lesson per mechanism plus one all-in reference:

  • s01_agent_loop.py through s12_worktree_task_isolation.py form the staged curriculum.
  • s_full.py combines the earlier mechanisms into one full reference cockpit.

This layout makes the repository useful not just as a claim-heavy manifesto, but as a decomposed teaching specimen for harness architecture.

Relationship To Existing Wiki Topics ​

This source reinforces several existing wiki pages:

  • Agent Harness Design, because it turns loop, tools, planning, skills, compression, tasks, background work, teams, and isolation into a staged instructional sequence.
  • Agent-First Repositories, because it treats repo-local skills, task state, inboxes, and worktree lanes as parts of the agent's operating environment.
  • Codex Operating Practices, because it offers a Claude-Code-centered comparison point for why harness surfaces often matter more than one-off prompting tricks.

Uncertainty And Scope Notes ​

  • The repo's philosophical stance is stronger and more polemical than many engineering notes in this vault; it should be treated as a useful lens, not as an uncontested definition of all agent systems.
  • The current raw family is still partial. Future maintenance may want to preserve additional session files from agents/ if this source becomes central evidence for more pages.
  • The current source page is a correction over an earlier too-direct wiki write-up; this version now explicitly grounds itself in raw files preserved inside the vault.

Sources ​

  • Repository: https://github.com/shareAI-Lab/learn-claude-code
  • Raw: raw/github/shareai-lab/learn-claude-code/README.md
  • Raw translation: raw/github/shareai-lab/learn-claude-code/README.zh.md
  • Raw Japanese sibling: raw/github/shareai-lab/learn-claude-code/README.ja.md
  • Raw implementation reference: raw/github/shareai-lab/learn-claude-code/s_full.py