Handbook · English
From prompt and context to agent and harness: four engineering layers, plus real job descriptions and an interview bank.
Chinese is the default edition of this site. This page is the English portal. Long chapters, the job-description notes, and all 92 interview answers have English siblings beside the Chinese files.
Prompt→Context→Agent→Harness
This page is the site portal. Use it to enter the four chapters, the interview bank, and the translations. The knowledge map is a separate review picture of the same material. It does not replace this portal.
One long essay per layer, with a table of contents and a chapter switcher. The historical merged draft is overview. The current text is the four chapters below.
How to make one call reliable: training stack, decoding, CoT, ReAct, evaluation.
What the model should see this turn: RAG, long context, memory, MCP.
How a multi-step task actually finishes: patterns, frameworks, failures, permissions.
Strong constraints and a self-healing loop, plus four pillars and a maturity model.
Excerpts from real postings, the requirements that repeat, and the limits of the evidence. Chinese quotations stay in Chinese.
Each question has a difficulty and a source URL, from fundamentals and RAG through agents, evaluation, and safety.
The four layers, the question bank, the translations, and the glossary on one review map. It does not replace this portal.
The merged edition from before the four chapters were split. Kept for comparison. It is not the current text.
The notes in this repo are unofficial Chinese translations. The English original is the authority. These cards open the Chinese study note, which links to the original.
Chinese study note. Original: anthropic.com/engineering/building-effective-agents
How we built our multi-agent research system
Effective harnesses for long-running agents
Harness engineering: leveraging Codex in an agent-first world