Context Engineering 入门

Overview

Who it's for

Anyone who has used Claude Code or similar coding assistants but finds it hard to get the AI to reliably finish whole tasks; beginners can follow the template too.

What you'll learn

1. 01Core idea: what Context Engineering is

   The repo contrasts levels of AI coding: Vibe Coding is 'casually chat and generate code', Prompt Engineering is 'get things done with clever wording', and Context Engineering is designing the entire context as a system — including docs, examples, rules, patterns and validation. It likens prompt engineering to handing someone a sticky note, and context engineering to writing a full screenplay with every detail spelled out.

2. 02Why it's more reliable: most agent failures are context failures

   The core argument: most agent failures are not model failures but 'context failures' — the AI never got the information needed to finish the job end to end. With the right context, the AI follows your project's patterns and conventions (consistency), handles multi-step features, and self-corrects through validation loops. That's what '10x better than prompt engineering, 100x better than vibe coding' means.

3. 03Template structure: CLAUDE.md + INITIAL.md + examples/ + PRP

   The template is a clone-and-go project skeleton: CLAUDE.md holds global rules (code structure, file size limits, testing requirements, style conventions); INITIAL.md is where you write the initial feature request; examples/ holds reference code patterns (called 'critical'); PRPs/ stores generated implementation blueprints; and .claude/commands/ defines the /generate-prp and /execute-prp slash commands. It's Claude Code-first, but the ideas port to other coding assistants.

4. 04The PRP workflow: from request to validation-gated implementation

   The core loop is two slash commands: write INITIAL.md, run /generate-prp so the AI researches your codebase patterns, finds relevant docs, and produces a complete PRP; then run /execute-prp to break the blueprint into a plan, implement step by step, run tests and lint, and iterate on failures. A PRP is like a PRD but crafted specifically to instruct an AI coding assistant — each one carries validation gates to raise the odds of getting it right the first time.

5. 05Best practices: be specific, feed plenty of examples

   Key takeaways: write INITIAL.md explicitly — FEATURE should not be 'build a scraper' but 'build an async scraper with BeautifulSoup, handle rate limiting, store in PostgreSQL'; fill examples/ with reference patterns (the AI performs dramatically better with patterns to follow, and examples should show both what to do and what not to do); add validation gates (test commands that must pass); leverage official docs and MCP resources; and customize CLAUDE.md per project.

Practice tips

• Clone the template, adapt CLAUDE.md to your project's rules and conventions, and run one real session to see the difference.
• Write a feature you want in INITIAL.md — try both a one-line request and a fully filled-out one, then compare the resulting PRPs.
• Drop your most typical module into examples/ and ask the AI to mimic its structure for new code.
• Run one small feature through the full /generate-prp → /execute-prp loop and watch how validation gates force the AI to self-test.

Companion workflows

Related resources