Rebuilt from Stanford CS146S Fall 2025 (archived) + Fall 2026 syllabus. Assignments repo: https://github.com/mihail911/modern-software-dev-assignments
Legend: [R] read [D] do [G] gap topic, no course material, self-sourced
[D] git clone https://github.com/mihail911/modern-software-dev-assignments
Clone now. Repo may be wiped/force-pushed when Fall 2026 term starts (~Sept 22 2026).
[D] Install an agent CLI (Claude Code / Codex / Cursor). Pick one, stick with it all the way.
[D] Pick ONE real personal repo as your lab. Every phase applies to it. No toy projects.
Goal: know what an LLM does at inference time and what an agent loop is.
[R] Deep Dive into LLMs — https://www.youtube.com/watch?v=7xTGNNLPyMI 3.5 hrs. Skip if you already know tokenization / sampling / context windows. [R] Building Effective Agents — https://www.anthropic.com/engineering/building-effective-agents The spine. Everything later hangs off this. [R] Prompt Engineering Guide — https://www.promptingguide.ai/techniques [R] AI Prompt Engineering: A Deep Dive — https://www.youtube.com/watch?v=T9aRN5JkmL8 [R] How OpenAI Uses Codex — https://cdn.openai.com/pdf/6a2631dc-783e-479b-b1a4-af0cfbd38630/how-openai-uses-codex.pdf [D] Assignment week1 — LLM Prompting Playground https://github.com/mihail911/modern-software-dev-assignments/tree/master/week1 [D] Capstone: write your own agent loop in <200 lines. Four tools only: read, write, edit, bash. This is Fall 2026's opening exercise. Do not skip — everything later is a variation on it.
Exit test: you can explain why an agent retries, and where the loop terminates.
Goal: the single highest-leverage skill. Context is the budget you manage.
[R] How Long Contexts Fail — https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html [R] Context Rot — https://research.trychroma.com/context-rot [R] Effective Context Engineering — https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents [R] Advanced Context Engineering (ACE-FCA) — https://github.com/humanlayer/advanced-context-engineering-for-coding-agents/blob/main/ace-fca.md = Fall 2026's "RePPIT" (Research, Propose, Plan, Implement, Test). Same idea, renamed. [R] Specs Are the New Source Code — https://blog.ravi-mehta.com/p/specs-are-the-new-source-code [R] Good Context Good Code — https://blog.stockapp.com/good-context-good-code/ [D] Take one real feature in your lab repo. Write a spec first, then have the agent implement from spec only. Compare against just prompting it. Note the difference.
Exit test: you plan before you prompt, and you know what to evict from context.
[R] MCP Introduction — https://stytch.com/blog/model-context-protocol-introduction/ [R] MCP official docs — https://modelcontextprotocol.io/docs/getting-started/intro [R] Writing Effective Tools for Agents — https://www.anthropic.com/engineering/writing-tools-for-agents [R] APIs Don't Make Good MCP Tools — https://www.reillywood.com/blog/apis-dont-make-good-mcp-tools/ Read this one twice. Most MCP servers are built wrong for exactly this reason. [R] Sample MCP servers — https://github.com/modelcontextprotocol/servers [R] MCP TypeScript SDK — https://github.com/modelcontextprotocol/typescript-sdk#server [R] MCP Server Auth — https://developers.cloudflare.com/agents/guides/remote-mcp-server/#add-authentication [R] MCP Registry — https://blog.modelcontextprotocol.io/posts/2025-09-08-mcp-registry-preview/ [D] Assignment week2 — First Steps in the AI IDE https://github.com/mihail911/modern-software-dev-assignments/tree/master/week2 [D] Assignment week3 — Build a Custom MCP Server https://github.com/mihail911/modern-software-dev-assignments/blob/master/week3/assignment.md [D] Connect your server to your agent. Use it for a week of real work. Fix what annoys you.
Exit test: your tool descriptions read like docs for a junior dev, not like an OpenAPI dump.
New in Fall 2026. No 2025 material exists. Highest-delta topic.
[R] Agent Skills overview — https://docs.claude.com/en/docs/agents-and-tools/agent-skills/overview [R] Equipping Agents with Skills — https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills [D] Write a SKILL.md for a workflow you personally repeat. Bundle a script with it. [D] Second skill, harder: one that calls out to the web or an external service.
Skills vs MCP: MCP adds capability (new tools). Skills add procedure (how to use what exists). Most people reach for MCP when they needed a skill.
Exit test: you can say which of your workflows should be a skill vs an MCP server vs a hook.
[R] Claude Code Best Practices — https://www.anthropic.com/engineering/claude-code-best-practices [R] How Anthropic Uses Claude Code — https://www-cdn.anthropic.com/58284b19e702b49db9302d5b6f135ad8871e7658.pdf [R] Peeking Under the Hood of Claude Code — https://medium.com/@outsightai/peeking-under-the-hood-of-claude-code-70f5a94a9a62 [R] Memory / CLAUDE.md — https://docs.claude.com/en/docs/claude-code/memory [G] [R] AGENTS.md — https://agents.md/ [G] [R] Hooks — https://docs.claude.com/en/docs/claude-code/hooks [G] [R] Subagents — https://docs.claude.com/en/docs/claude-code/sub-agents [G] [R] Awesome Claude Agents — https://github.com/vijaythecoder/awesome-claude-agents [R] SuperClaude Framework — https://github.com/SuperClaude-Org/SuperClaude_Framework [R] Devin: Coding Agents 101 — https://devin.ai/agents101#introduction [D] Assignment week4 — Coding with Claude Code https://github.com/mihail911/modern-software-dev-assignments/blob/master/week4/assignment.md [D] Your lab repo: add CLAUDE.md/AGENTS.md, one lint hook, one test hook. [D] Build a planner -> implementer -> reviewer subagent chain. Run a real feature through it.
Exit test: a fresh agent session on your repo needs no hand-holding to run tests correctly.
New in Fall 2026. No canonical doc exists — this phase is mostly doing, not reading. Derive principles from ACE-FCA (Phase 2) + hooks docs (Phase 5).
Checklist to apply to your lab repo:
- One command each for build / test / lint. No undocumented steps.
- AGENTS.md states architecture, conventions, and what NOT to touch.
- Tests fast enough that an agent can run them in-loop.
- Errors are readable: no silent failures, no swallowed exceptions.
- Dead code and stale docs deleted — agents trust them and get poisoned.
[D] Score your repo 1-10 on each. Fix the two worst. Re-run a task and measure the difference.
Exit test: agent completes a nontrivial task in your repo with zero clarifying questions.
[R] Code Reviews: Just Do It — https://blog.codinghorror.com/code-reviews-just-do-it/ [R] How to Review Code Effectively — https://github.blog/developer-skills/github/how-to-review-code-effectively-a-github-staff-engineers-philosophy/ [R] Code Review Essentials — https://blakesmith.me/2015/02/09/code-review-essentials-for-software-teams.html [R] AI-Assisted Assessment of Coding Practices (arXiv) — https://arxiv.org/pdf/2405.13565 [R] AI Code Review Best Practices — https://graphite.dev/guides/ai-code-review-implementation-best-practices [R] Lessons from millions of AI code reviews — https://www.youtube.com/watch?v=TswQeKftnaw [D] Assignment week7 — Code Review Reps https://github.com/mihail911/modern-software-dev-assignments/tree/master/week7 [D] Write a reviewer subagent with custom rules for your repo. Run it on your last 10 commits. Log what it caught vs what it missed. The misses are the lesson.
Do not skip. Agent-specific attack surface is real and most people ignore it.
[R] SAST vs DAST — https://www.splunk.com/en_us/blog/learn/sast-vs-dast.html [R] OWASP Top Ten — https://owasp.org/www-project-top-ten/ [R] Copilot RCE via Prompt Injection — https://embracethered.com/blog/posts/2025/github-copilot-remote-code-execution-via-prompt-injection/ [R] Agentic AI Threats — https://unit42.paloaltonetworks.com/agentic-ai-threats/ [R] Finding Vulns with Claude Code + Codex — https://semgrep.dev/blog/2025/finding-vulnerabilities-in-modern-web-apps-using-claude-code-and-openai-codex/ [R] o3 finds CVE-2025-37899 — https://github.com/SeanHeelan/o3_finds_cve-2025-37899/blob/master/system_prompt_uafs.prompt [D] Assignment week6 — Writing Secure AI Code https://github.com/mihail911/modern-software-dev-assignments/blob/master/week6/assignment.md [D] Run Semgrep on your lab repo. Have the agent triage findings. Verify its triage by hand — it will confidently dismiss at least one real issue.
New in Fall 2026.
[R] Claude Code GitHub Actions — https://docs.claude.com/en/docs/claude-code/github-actions [R] Cursor background agents docs + Devin docs (Phase 5 link) [D] Wire one agent to a real trigger: GitHub issue -> PR, or a scheduled repo maintenance job. [D] Run 3 agents in parallel on independent tasks (git worktrees). Note where they collide.
Thinnest public material of any phase. Read what exists, then reason from it.
[R] MCP in practice / remote servers — https://docs.claude.com/en/docs/claude-code/mcp [R] LiteLLM docs — gateway, model routing, cost tracking [R] Introduction to SRE — https://sre.google/sre-book/introduction/ [R] Observability Basics — https://last9.io/blog/traces-spans-observability-basics/ [R] Resolve.ai: https://resolve.ai/blog/product-deep-dive https://resolve.ai/blog/role-of-multi-agent-systems-AI-native-engineering https://resolve.ai/blog/kubernetes-troubleshooting-in-resolve-ai https://resolve.ai/blog/Top-5-Benefits [D] Instrument your agent usage: log tokens and cost per task. Find your most expensive workflow. Cut its cost 50% via prompt-prefix stability (KV cache) or model routing.
Fall 2025 grading: Final Project 80%, Weekly Assignments 15%, Participation 5%. The project is the course. Treat it that way.
Build something where agents do the work and you designed the system:
- A repo that maintains itself (agent-driven deps, tests, docs, triage), or
- A domain-specific agent harness with your own skills + MCP server + review gates, or
- An issue-to-merged-PR pipeline with human approval only at the gate.
Requirements: your own MCP server, at least 2 skills, hooks as guardrails, a reviewer subagent, and cost instrumentation. Write up what the agent got wrong and how the system caught it.
- Week 5 Modern Terminal (Warp) — vendor-specific, dropped from 2026
- Week 8 Automated UI / app building (Vercel, bolt.new) — dropped from 2026 Assignments week5 and week8 in the repo map to these. Optional.
Full: 14 weeks at ~8 hrs/week = matches the real course. Compressed: Phases 2,3,4,5,6 only = ~5 weeks, covers the highest-value 60%. Already built an agent loop: start at Phase 4, backfill Phase 2 reading.
Fall 2026 site publishes new readings weekly from ~Sept 22 2026: https://themodernsoftware.dev Fall 2025 archive (full original resources): https://themodernsoftware.dev/fall2025