Skip to content

CLAUDE.md

Latest commit

 

History

History
75 lines (48 loc) · 9.13 KB

CLAUDE.md

File metadata and controls

75 lines (48 loc) · 9.13 KB
Raw

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Repository Overview

This is a ChatGPT and GitHub Copilot training course repository for a 4-hour O'Reilly hands-on workshop (July 2026 delivery). The course teaches professionals how to use AI tools across developer, IT ops, data science, and information worker roles.

Course Architecture

The repository follows a 4-hour teaching cadence:

  • Hour 1: ChatGPT fundamentals (tiers, Projects, Tasks, Custom GPTs)
  • Hour 2: Advanced ChatGPT + agentic AI (Codex, Images 2.0, data analysis, Personal Finance preview)
  • Hour 3: GitHub Copilot core (tiers, models, chat modes, custom instructions)
  • Hour 4: Enterprise + agentic coding (MCP, Copilot CLI, coding agent, governance)

The current course plan lives at course-plan-july-2026.md. A detailed news timeline is at whats-new-july-2026.md. Previous plans are archived in docs/archive/.

Filename convention: master files are dated per delivery (course-plan-{month}-{year}.md, whats-new-{month}-{year}.md, warner-chatgpt-github-copilot-{month}-{year}.pptx). Each new delivery cycle, these are renamed via git mv and the README header links + repo tree get updated to match. Don't create a new file when refreshing -- rename the existing one so history follows.

Key File Relationships

  • course-plan-july-2026.md - Master course plan with detailed timings and content
  • warner-chatgpt-github-copilot-july-2026.pptx - Presentation deck (binary, must match course plan; single canonical deck at repo root)
  • docs/instructor/mcp-teaching.guide.md - MCP demo setup walkthrough
  • Teaching artifacts (do NOT treat as project rules): .github/copilot-instructions.md, .github/agents/, .github/instructions/ exist as live examples students inspect during Segment 3. The content is generic Copilot-instruction guidance, not directives that govern work in this repo.

Running Demo Code

There is no build / lint / test framework in this repo. Each script under demos/ is standalone and runs on demand during teaching. Don't search for a unified test runner or package.json at the root -- none exists.

OpenAI API examples (run from PowerShell on Tim's primary box; bash shown for cross-platform):

export OPENAI_API_KEY="your-key"
python demos/chatgpt/api-examples/call_openai_api.py

Weather MCP server (Node.js, the canonical MCP demo): mock data, no API key, so it always runs on stage.

cd demos/mcp/weather-server && npm install && node server.js

It uses the modern MCP TypeScript SDK (^1.29.0): McpServer + server.registerTool with zod input schemas, StdioServerTransport. The old low-level Server + setRequestHandler pattern is deprecated, so keep any rewrite on the high-level API. Logging goes to stderr because stdout carries the JSON-RPC channel. Point the Inspector at it with npx @modelcontextprotocol/inspector node server.js.

Pre-delivery linter (run before every class): pwsh scripts/utilities/preflight-check.ps1 asserts every datasets/... path referenced in the plan/labs resolves on disk, flags stale model/version/command tokens (it exempts dated historical-timeline rows), and blocks committed OpenAI keys. Update its $StaleTokens map each refresh cycle. Copy .env.example to .env for local keys; .gitignore excludes .env.

Standalone Python MCP servers (calculator-mcp.py, weather-mcp.py) are referenced by docs/instructor/mcp-teaching.guide.md as build-it-live exercises -- they're authored on stage from the teaching guide, not committed. Use npx @modelcontextprotocol/inspector to point at any running server.

Vulnerable demo (Docker):

docker build -f demos/vulnerable-code/Dockerfile -t chatgpt-demo .
docker run -p 5000:5000 chatgpt-demo

Files That Look Broken But Are Not

  • Intentionally vulnerable: demos/vulnerable-code/ and demos/security-scanning/ contain insecure code and outdated dependencies for security-education demos. Do not "fix" these unless explicitly asked. The requirements.txt in demos/vulnerable-code/ has known CVEs by design.
  • Live-completion stubs: demos/copilot/inline_suggestions.py, main.py, and test.py are deliberately incomplete (a function signature with no body, a missing colon). Tim types these live and Copilot completes them on stage, so they will not compile. Leave them incomplete. When compile-checking the repo, exclude these three plus demos/vulnerable-code/bad-python.py.

Editing the Deck

The .pptx is a binary that must match the course plan. Slides fall in three zones: scaffold (~1-11) the reusable class opener (title, bio, schedule, materials), teaching core (~12-41) the four segments, and archive (~42+) supplemental pull-outs (a slide titled "Archive / Supplemental" marks the boundary). Preserve the scaffold. Edit content-slide text in place at the run or paragraph level; never add, remove, move, or resize a shape or slide. Prove layout preservation with a semantic shape-geometry fingerprint compared before vs after across all slides (zero geometry changes is the pass bar). For speaker notes, use the pluralsight-speaker-notes skill (edits the notes text frame only) and route spoken-word notes through the tim-warner-voice skill.

Verifying links means fetching each URL and confirming HTTP 200 to the correct target, not just well-formed syntax. openai.com and github.com return 403 to WebFetch and plain curl bot-checks, so use curl -sSL -o /dev/null -w "%{http_code}" -A "<real UA>" <url> or gh. A 401 on https://api.githubcopilot.com/mcp/ is expected (auth-gated), not a break.

Technology Landscape (July 2026)

When updating course content, these are the current-gen references:

ChatGPT / OpenAI: GPT-5.6 family (Sol flagship / Terra balanced / Luna cost-efficient), launched July 9, 2026 across ChatGPT, Codex, and API (Codex context 272K); GPT-5.5 Instant remains the everyday-chat default (May 5, 2026; 52.5% fewer hallucinations than 5.3); GPT-5.5 / 5.5 Pro (xhigh reasoning), GPT-5.4 Thinking / 5.4 Pro (1M context in API); $100 and $200 ChatGPT Pro tiers; ChatGPT + Codex merged into one desktop app (macOS / Windows) plus new "ChatGPT Work" agent, July 9; ChatGPT Images 2.0 (April 21, text-in-image); Memory Sources (all plans); Personal Finance GA on Plus ($20) as of June 30, 2026 (was US Pro preview May 15, via Plaid); Tasks (Plus 40/mo, Pro unlimited); ChatGPT Agent (browser automation via /agent); Deep Research; Canvas (out of beta). Sora 2 discontinued April 26; models + Videos API fully shut down Sept 24, 2026. GPT-4o/4.1/o4-mini retired Feb 13. GPT-5.3 Instant retiring ~August 2026.

GitHub Copilot: VS Code current v1.103 (July 2026) (June v1.102); recent highlights include agentic browser tools GA, 1M-context support for compatible Anthropic/OpenAI models, agent-session grouping with cost visibility, MCP trust layer, semantic codebase search + grep across orgs, /chronicle chat-history queries, agent terminal read/write, remote CLI control from github.com/mobile. Copilot CLI (GA Feb 25; package @github/copilot, current v1.0.71 as of July 16, 2026; new terminal interface GA June 23) with plan/autopilot/remote modes and auto model routing (July 1). Claude Sonnet 5 GA in Copilot June 30 (Pro/Pro+/Max/Business/Enterprise), Claude Opus 4.8 available, Fable 5 in preview; Auto model selection is the default paradigm (enterprises can set Auto as org default, July 1) rather than a single pinned model; other models include GPT-5.6-Codex, GPT-5.3-Codex, Gemini 3 Pro/Flash. Grok Code Fast 1 deprecated May 15, 2026. Net-new since May: Agent Finder GA (June 17, implements Agentic Resource Discovery / ARD spec), Claude as agent provider in JetBrains (June 22), C++ modernization agent GA (June), GitHub Copilot app GA to all (July 7). Enterprise AI Controls GA (Feb 26); Copilot Memory on by default for Pro/Pro+; Mission Control + Agents tab in every repo.

MCP: Donated to Agentic AI Foundation (Linux Foundation) Dec 2025; 2,000+ servers; new spec (2026-07-28) in Release Candidate - largest revision since launch: stateless HTTP core, MCP Apps + Tasks as formal extensions, OAuth/OIDC-aligned auth, formal 12-month deprecation policy. The prior Nov 2025 spec added async Tasks, OAuth 2.1, statelessness; MCP Apps (SEP-1865) formalized April 2026 (interactive UIs via ui:// scheme). TypeScript SDK current v1.29.0 (modern McpServer + server.registerTool API; the low-level Server + setRequestHandler pattern is deprecated). The npm package @modelcontextprotocol/server-github is deprecated; use GitHub's hosted MCP server (https://api.githubcopilot.com/mcp/) or the github/github-mcp-server binary. Azure MCP Server preview.

Claude / Anthropic: current lineup is the Claude 5 family (Fable 5), Opus 4.8, Sonnet 5, Haiku 4.5. Opus 4.8 has a fast-mode preview. Sonnet 5 replaced Sonnet 4.6 as the efficient everyday choice; Opus 4.7 is superseded by 4.8. Claude Code 2.x (skills, auto-memory, /teleport, custom agents, hooks, wildcard tool permissions).