A practitioner's guide to shipping production software with Claude Code, Cursor, and AI-powered development frameworks. From foundational concepts to team-wide enablement.
Whether you use Claude Code, Cursor, Copilot, or Codex — every productive developer converges on the same meta-pattern. Internalize this before touching any tool.
The #1 mistake developers make with AI coding tools is letting the AI jump straight into writing code. Addy Osmani (Chrome team lead at Google) summarized the right approach: treat the LLM as a powerful pair programmer that requires clear direction, context, and oversight — not autonomous judgment. At Anthropic themselves, approximately 90% of the code for Claude Code is written by Claude Code itself, but always within structured workflows.
The paradigm shift in 2026 is from "prompt engineering" to context engineering — giving the AI the right information at the right time in the right format. Boris Cherny (Staff Engineer at Anthropic, creator of Claude Code) runs 10–15 sessions simultaneously with only ~100 lines in his CLAUDE.md file. His golden rule: "Anytime we see Claude do something incorrectly, we add it to CLAUDE.md so it doesn't repeat next time."
As a conversation grows, the AI accumulates prior messages, code snippets, and tool outputs that fill its context window. Quality degrades — the AI starts cutting corners. At 50% context usage, quality drops noticeably. At 70%+, hallucinations increase. At 90%+, responses become erratic. The fix: use subagents for heavy tasks, start fresh sessions for new work, and never dump information "just in case."
His 6-step task loop: (1) Give Claude a task → (2) Claude creates a plan → (3) Review the plan → (4) Claude executes → (5) Review the diff → (6) If Claude made a mistake, have it write a rule in lessons.md. Over time, Claude literally teaches itself to be better at your specific project.
A terminal-first AI coding agent that reads files, runs commands, makes edits, and iterates. It's the backbone of the most productive AI-augmented engineering teams in 2026.
Claude Code is not autocomplete. It's an execution loop: it receives an instruction, pulls context from the codebase, chooses tools, makes changes, reads results, and decides what to do next. The prompt is one line; execution can be five or more steps. Understanding this loop is the key to using it effectively.
Claude Code's power comes from its extensibility system. There are six main customization layers, each serving a different purpose:
| Layer | What It Does | Key Insight |
|---|---|---|
| CLAUDE.md | Project-level memory loaded every session | Advisory — Claude follows it ~80% of the time. Keep it lean. |
| Skills | On-demand expertise loaded when relevant | Skills are folders with SKILL.md, templates, scripts. Loaded dynamically. |
| Subagents | Isolated specialist assistants with fresh context | Run in own context with own tools. Keeps your main session clean. |
| Hooks | Deterministic scripts that run on events | 100% reliable — unlike CLAUDE.md which is advisory. Use for linting, formatting. |
| Commands | Reusable prompt templates (/slash-commands) | Markdown files in .claude/commands/. Automate repetitive workflows. |
| Plugins | Bundled packages of skills + hooks + agents | Installed via /plugin from the marketplace. Community-shareable. |
His CLAUDE.md is only ~100 lines / ~2,500 tokens. He runs 10–15 parallel sessions (5 terminal tabs, 5–10 web, a few mobile). Each session gets its own git worktree. His team's rule: the file is checked into git, shared by the whole team, and updated multiple times a week. See the full breakdown.
Plan Mode, CLAUDE.md, subagents, skills, hooks, MCP servers, parallel sessions, agent teams, CI/CD integration. The starting point for everything.
Hook recipes (PostToolUse formatter, PreToolUse safety blocks), agent teams, /loop scheduling, statusline scripts, context compaction strategies, and more.
Plan-first with Shift+Tab, TDD approach (70% fewer production bugs), custom slash commands, /rewind for safe experimentation, multi-session parallelism.
Systems-level understanding: context ingestion, tool selection, iteration loops, failure modes, and scaling across a team. For engineering leads.
The single highest-leverage investment you can make. This file transforms Claude from a generic AI into a teammate who understands your codebase.
Claude Code has a layered configuration system. Understanding it prevents duplication and conflicts:
All levels combine — they don't replace each other. More specific rules override on conflicts.
Claude Code's system prompt already contains ~50 instructions. That's nearly a third of the ~150-200 instruction limit models can reliably follow. Your CLAUDE.md must be lean — every line competes for attention.
| ✓ Include | ✗ Exclude |
|---|---|
Build commands: npm run test, npm run build | Personality instructions: "Be a senior engineer" |
| Non-obvious project conventions and architecture | Formatting rules — use a linter/formatter hook instead |
| Common pitfalls specific to your codebase | @-mentioning docs that embed entire files every session |
| Testing commands and verification scripts | Duplicate rules already in global or parent CLAUDE.md |
| Pointers to where to read more (just-in-time context) | Step-by-step instructions for rare tasks (use Skills instead) |
CLAUDE.md is advisory (~80% compliance). Hooks are deterministic (100%). If something must happen every time without exception (formatting, linting, security checks), make it a hook. If it's guidance Claude should consider, CLAUDE.md is fine. Never send an LLM to do a linter's job. See Writing a Good CLAUDE.md for the full analysis.
Next.js/TypeScript, Python/FastAPI, and generic templates. Based on official guidance and Boris Cherny's setup. Includes the "do's and don'ts" guide — the must-read companion file.
Instruction limits, progressive disclosure, leverage diagrams. The most analytical treatment of what belongs in CLAUDE.md vs. hooks vs. skills.
Working demo of the full Command → Agent → Skill architecture pattern with weather example, sound hook system, and comprehensive CLAUDE.md configuration.
The extensibility system that transforms Claude Code from a chatbot into a configurable engineering platform.
Skills are folders with a SKILL.md file that give Claude specialized instructions for specific tasks. Unlike CLAUDE.md which loads every session, skills load dynamically only when relevant — keeping your context window clean.
references/, scripts/, examples/ subdirectories for progressive disclosure!`command` in SKILL.md to inject dynamic shell output into the prompt at invocation timeSubagents run in their own context window with their own tools. They're defined in .claude/agents/ as markdown files with YAML frontmatter specifying name, description, tools, and model. Key properties include skills: (preloaded skills), mcpServers:, hooks:, isolation: "worktree", and effort: level.
Hooks run scripts automatically at specific lifecycle points. Available events include: PreToolUse, PostToolUse, UserPromptSubmit, Notification, Stop, SubagentStart/Stop, PreCompact, SessionStart/End, PermissionRequest, TaskCompleted, and more. Use hooks for anything that must happen 100% of the time.
Build Skills from scratch, write frontmatter, distribute via plugins, deploy organization-wide, wire into subagents. The official course from Anthropic.
Hands-on: skills across Claude.ai, Claude Code, API, and Agent SDK. Combine skills with MCP and subagents for complex workflows. Taught by Elie Schoppik.
The Claude A / Claude B iteration pattern, progressive disclosure, conciseness testing, and the "complete → identify → package → iterate" skill development cycle.
Full reference: SKILL.md format, frontmatter options, directory structure, subdirectory discovery for monorepos, scripts, progressive disclosure patterns.
The spec-driven development framework with 45.6K+ GitHub stars that makes AI coding reliable at scale. Supports 8 AI runtimes including Claude Code and Cursor.
Vibecoding has a bad reputation — you describe what you want, AI generates code, and you get inconsistent garbage that falls apart at scale. GSD fixes this by providing the context engineering layer that makes AI coding reliable. It forces planning before coding, spawns fresh context windows for each task, and verifies everything before shipping.
GSD sits between you and your AI agent, structuring every interaction through four layers: Commands (your entry points), Workflows (orchestrators that load context and spawn agents), Agents (15 specialists with focused roles), and State (persistent .planning/ directory). The system uses 37 commands, wave-based parallel execution, and goal-backward verification.
GSD v1 (get-shit-done, v1.28.0, 45.6K ⭐) is the prompt-based framework — installed as slash commands into your agent's configuration. Works across Claude Code, Cursor, Gemini CLI, Codex, Copilot, Windsurf, and Antigravity.
GSD v2 (gsd-2, newly released) is a standalone CLI built on the Pi SDK with direct TypeScript access to the agent harness. It can actually clear context between tasks, inject files at dispatch time, manage git branches, track costs, detect stuck loops, and auto-advance through entire milestones without human intervention.
The original prompt framework: 15 agents, 37 commands, multi-wave parallel execution. Trusted by engineers at Amazon, Google, Shopify, and Webflow.
The evolution: standalone CLI on Pi SDK with real control over context, sessions, git, costs, and auto-recovery. 20+ model providers supported.
Build a real Expense Splitter app while learning GSD. Covers context rot, fresh subagent contexts, atomic verifiable builds. Hands-on walkthrough.
How GSD works under the hood. Spec-driven dev patterns, workflow files, Q&A requirement definition systems. Understand fundamentals to build your own.
The AI-first IDE built on VS Code. Over 90% of Salesforce developers and 40,000+ engineers at major companies now use it. Cursor 2.0 shipped agents, hooks, and the Composer model.
Create a .cursor/rules/ directory with .mdc files for project-specific rules. Key rules to add immediately: "Never replace code with placeholders", "Always provide complete code", "Break problems into smaller steps", "Always provide a PLAN with REASONING before making changes."
Plan Mode, rules, agent skills, review workflows (Stop → Review → Agent Review), background agents, and the "iterate on setup" philosophy.
.cursorrules files, .mdc rule files, Composer vs Agent mode, instructions.md, context management, .cursorignore, and the full rule priority system.
The 10-line .cursorrules that fixed common problems, context-referencing technique, morning planning / afternoon execution workflow split.
Product Requirement Documents + Cursor's Cloud Agents + Interactive UIs + Multi-model support. Connecting product thinking to code generation.
Anthropic Academy launched 13 free courses on March 2, 2026. All free, all with certificates, no paid subscription needed. Plus additional resources from DeepLearning.AI and Coursera.
Hosted on Skilljar at anthropic.skilljar.com, these courses are taught by the engineers who built Claude. Every completed course awards an official certificate you can add to LinkedIn. The full catalog takes approximately 15-20 hours to complete.
21 lessons: AI coding assistant architecture, tool system, context management, custom commands, MCP integration, GitHub automation, hooks, and the SDK. Best ROI course in the entire Academy.
Build Skills from SKILL.md frontmatter, organize with progressive disclosure, distribute via plugins, deploy organization-wide, wire into subagents. Troubleshooting included.
Use and create subagents to manage context, delegate tasks, and build specialized workflows that keep your main conversation clean and focused.
First conversations, prompting, desktop app, Projects, Artifacts, Skills, enterprise search, Research Mode, and role-specific use cases. The entry point for non-developers.
Build MCP servers and clients from scratch in Python. Three core primitives: tools, resources, prompts. Advanced course covers sampling, notifications, transport mechanisms for production.
The deepest course: 84 lectures, 10 quizzes. API authentication, multi-turn conversations, prompt engineering, tool use, streaming, extended thinking, RAG, and production patterns.
Live and recorded deep dives directly from the Anthropic engineering team — covering practical workflows you won't find in the docs.
The fastest way to go from zero to productive with Claude Code. Ideal for team workshops — watch together, then immediately apply to your own projects.
Deep dive into subagents, MCP server integration, and scaling to real codebases. Covers the advanced patterns that separate beginners from power users.
Andrew Ng × Anthropic. Build skills across Claude.ai, Claude Code, API, and Agent SDK. Combine with MCP and subagents. Create a research agent from scratch.
Same content as Anthropic Academy version but with Coursera Coach AI assistant, additional quizzes, assessments, and interactive dialogues for reinforcement.
Build a complete online radio station web app from scratch. Hands-on project covering the full dev lifecycle, limitations of AI, and when human insight is essential.
5 interactive notebook courses: API fundamentals, prompt engineering tutorial, real-world prompting, prompt evaluations, and tool use. All code-along format.
The curated collection of community resources, tools, and references that accelerate your AI coding workflow.
Skills, hooks, slash-commands, agent orchestrators, CLAUDE.md files, plugins, tooling, and educational resources. Updated daily. Bookmark this and check weekly.
Interactive deep-dive into Claude Code's internals mapped from the source: the agent loop from keypress to response, 40+ built-in tools, multi-agent orchestration, slash commands, and unreleased features. 929 HN points. Essential for understanding how it actually works.
A comprehensive collection of Claude Code resources, guides, tools, and community projects in one place. Great as a discovery starting point alongside awesome-claude-code.
Beginner to power user. Production templates, security auditing (24 CVEs, 655 malicious skills tracked), context management rules, quizzes, cheatsheet, and case studies.
Progressive learning with diagrams and copy-paste templates. Install your first slash command in 15 minutes. Includes quizzes per module. Perfect for onboarding.
Install agents, commands, hooks, MCPs, and settings via a single CLI. Real-time session monitoring dashboard. Complete dev stack in one command.
Copy-paste starter templates based on official guidance and Boris Cherny's setup. Lean, battle-tested configs with the do's-and-don'ts companion guide.
Plugin marketplace: dev-workflows (backend), dev-workflows-frontend, discover (product discovery), metronome (prevents shortcut-taking), governance (signoff checkpoints).
Template repo with kk plugin (9 skills, 4 commands, hooks). Team-first: project-scoped settings shared via git. Template sync workflow for updates.
Skills vs plugins vs hooks in practice, real-world CLAUDE.md approach, when to use subagents, MCP integration decisions, and the "harness heavy-carries" insight.
AI doesn't replace QA — it shifts the role from writing repetitive test code to designing testing strategy. Claude Code + Playwright is the most powerful QA automation combination available today.
Claude integrates directly with frameworks like Playwright, Selenium, and Cypress — going from generated test cases to running them inside your CI/CD pipeline with minimal manual translation. Playwright now ships with three specialized agents (planner, generator, and healer) that work as Claude Code subagents specifically built for test automation.
OpenObserve's QA team built 8 specialized AI agents powered by Claude Code that automate their entire E2E testing pipeline. Feature analysis dropped from 45-60 minutes to 5-10 minutes, flaky tests reduced by 85%, and test coverage grew from 380 to 700+ tests. The key insight: specialization over generalization — early iterations tried using one "super agent" and failed.
claude "read the auth module and generate Playwright E2E tests for login, registration, and password reset"Create a QA persona that runs on every pull request via GitHub Actions using anthropics/claude-code-action with Playwright MCP. The agent tests the app like a real user — clicks buttons, fills forms, resizes the browser, checks mobile layouts — and writes detailed bug reports. Setup guide: Building an AI QA Engineer with Claude Code and Playwright MCP
How 8 specialized Claude Code agents automated E2E testing. 85% reduction in flaky tests, caught production bugs while writing tests. Full implementation details.
Playwright's 3 specialized agents (planner, generator, healer) as Claude Code subagents. How the planner works through scenarios like a QA engineer would.
Complete setup: MCP server config, Three Amigos workflow integration, requirement-based test generation, failure analysis, and automated test healing.
From raw codebase to fully populated test suite in minutes using TestCollab MCP Server (17 tools). Generate suites, write cases, manage plans, log execution results.
Real QA workflow with Claude: recorder-to-code pipeline, Page Object Model generation, common pitfalls (fragile selectors, hardcoded values), and the human/AI handoff points.
Structured evaluation: Claude scored 9.5/10 on test generation quality. Covers positive/negative test cases, Jest suites, CI/CD integration, and edge case coverage.
The design workflow is being redefined. Designers at Jane Street now design with Claude more than Figma. The Claude Code + Figma MCP bridge is eliminating the translation layer between design and code.
In February 2026, Figma officially launched Claude Code to Figma — an MCP integration that lets you capture a rendered UI from a browser (production, staging, or localhost) and convert it into editable Figma frames. Not images. Not screenshots. Fully editable layers with proper naming and auto layout. You can push changes back and forth between code and canvas.
Figma's CEO says the traditional "product → design → code" sequence is dead. Modern workflows start anywhere and go everywhere. The ultimate skill is now deciding what to build, not how to build it.
Describe what you want → Claude Code builds it → Send to Figma for refinement → Push changes back. A Jane Street designer reports that prototypes built this way have "felt better in almost every way compared to mockups and docs." Claude gave unlimited iteration — unbothered when they changed their mind for the 50th time.
Build Claude Code skills that encode your design system: principles, component specs, content strategy, motion guidelines. Then create capability skills that compose these into pipelines: PRD → design ticket → user flow → generate design → review → QA signoff.
Generate user flow diagrams, Gantt charts, system architecture diagrams, and decision trees directly from PRDs or conversations with Claude — all as editable FigJam diagrams for team collaboration.
The official launch announcement. How code-first exploration meets canvas-based divergence. Capture browser UI as editable Figma frames, iterate between code and design.
A designer's experience building prototype features in the actual codebase. "Workflow improvements that would have taken days of engineering back-and-forth, or more likely would never have happened."
The complete skill architecture for designers: reference skills (principles, specs), capability skills (design review, generation, prototyping), and MCP tool connectors. Full pipeline examples.
"After 100 days of vibe-coding as a designer" — the smoothest workflow for 5x faster designs. Figma MCP setup, token-first workflow (C), and one-shot game built from FigJam flowchart.
Turn AI conversations into editable FigJam diagrams. User flows from PRDs, system architecture from code, Gantt charts, decision trees. Visual collaboration for PMs, designers, and engineers.
PMs face increasing pressure to move beyond writing requirements and waiting for sprints. Claude Code enables non-engineers to build functional prototypes, investigate user issues, and analyze codebases directly.
Product managers don't need to become engineers. But learning to work with Claude Code unlocks capabilities that were previously impossible without developer help:
The most common PM mistake is the "spec dump" — feeding a complete feature specification and expecting a miracle. The baby steps framework breaks features into 3-5 focused tasks with clear verification criteria at each step. Build → verify → add complexity → verify → integrate → verify.
Connect Claude Code to your product tools: GitHub MCP (create PRs, update issues, run Actions), Jira MCP (update tickets, add comments, estimate story points), Figma MCP (bridge design and implementation), Slack MCP (notifications, fetch messages), Database MCP (query data, analyze schemas).
5 real workflows with step-by-step examples. Installation, agents, MCP servers, the "baby steps" framework, cost management, and avoiding the "spec dump" anti-pattern.
How to write PRDs that AI agents can actually execute. Cursor's Cloud Agents, Interactive UIs, multi-model support. Connecting product thinking to code generation.
The 5 meta-skills that matter now: problem shaping, context curation, product "taste," agent orchestration, and knowing when NOT to use agents. The new PM skillset.
The starting course for PMs: prompting, Claude desktop app, Projects, Artifacts, Skills, Research Mode, and role-specific use cases. Certificate for LinkedIn.
A phased approach to rolling out AI coding workflows across your engineering team. Designed for technical leads and engineering managers.
npx get-shit-done-cc --claude --global.cursor/rules/ with basic rules — commit to repoMost teams operate at only 15-25% flow efficiency — 75-85% of lead time sits in queues and delays. If AI accelerates coding but review queues grow, overall delivery won't improve. Measure: