tw93
Tw93 — Chinese-language tech writer at tw93.fun. Engineer at Alibaba’s Fliggy unit; image hosting on Fliggy’s CDN suggests affiliation (2026-04-27-claude-code-architecture-governance-engineering).
Sources in this wiki
- 2026-04-27-claude-code-architecture-governance-engineering — 你不知道的 Claude Code：架构、治理与工程实践. The seed essay; introduces six-layer-agent-architecture.
- 2026-04-27-agent-principles-architecture-engineering — 你不知道的 Agent：原理、架构与工程实践. Companion essay zooming out from Claude Code to general Agent engineering. Centers on harness over model, agent-computer-interface tool design, and a walk-through of openclaw as a reference implementation.
- 2026-04-27-llm-training-principles-paths-practices — 你不知道的大模型训练：原理、路径与新实践. Third in the series; jumps a layer up from runtime Agent engineering to the training stack itself (nine-stage pipeline, post-training, eval-grader-reward, reward-hacking, meta-harness, distillation). Marks his shift from “I run these systems” to “I read frontier-lab training reports and synthesize them for non-specialists”.
Work referenced in this wiki
- kaku — open-source terminal he built using Claude Code (Rust + Lua,
tw93/Kaku). Running case study for Skills, Hooks, and mixed-language project layout in the Claude Code essay. tw93/waza— open-source health-check Skill marketplace plugin. Provides/healthto audit a Claude Code project’sCLAUDE.md/ rules / Skills / Hooks /allowedTools. Not yet a wiki page; mentioned only on 2026-04-27-claude-code-architecture-governance-engineering.
Stance
Tw93’s writing is unusually engineering-shaped: he treats Agents and claude-code as systems to be governed rather than chatbots to be prompted. The six-layer-agent-architecture frame from the first essay, the “Harness > model” thesis from the second, and the “back-half-of-the-pipeline-decides-everything” frame from the third are his three central contributions in this wiki.
He runs two paid Claude Code accounts (~$40/mo each) and has 6+ months heavy use of Claude Code as of the first essay. The second essay broadens his observation surface beyond a single vendor — he reads OpenAI’s Harness engineering, Anthropic’s Demystifying evals, Cursor’s MCP file-backed-description A/B test, LangChain’s State-of-Agent-Engineering survey, and walks through openclaw (which he didn’t write). The third essay broadens it again — he stops being a primarily-runtime author at all, and synthesizes across DeepSeek’s training reports, Anthropic’s alignment research, OpenAI’s reasoning-model and Deliberative-Alignment posts, the Meta-Harness paper, and the Kimi / Cursor / Chroma Agent-training cases. The trajectory across three essays is “Claude Code user → Agent engineer reader → training-stack synthesizer” — single-vendor lens has fully loosened. Still worth flagging when his summaries of secondary sources conflict with primary docs (prompt-caching open question on defer_loading; the meta-harness open questions on the 6× spread number are similar cases for the third essay).