Ballpoint 8-week AI programme: consolidated reading list

A condensed eight-week AI engineering curriculum has been published, cutting an original 13-week programme down to the sessions deemed most valuable. Each week pairs roughly 60-90 minutes of reading and video with a hands-on practical, covering context engineering, APIs, MCPs and CLI tools, with evaluation placed at week five ahead of the agents and loops material that depends on it.

A condensed rerun of the original 13-week programme, cut to the eight weeks that earned their place. Each week is roughly 2 hours total: 60-90 minutes of watching and reading, plus a practical. Optional extras sit outside the 2 hours. Weeks 3 and 4 are a deliberate pair "how data moves" then "where data lives" . Evals sits at week 5, just before the agents, compound engineering and loops weeks that depend on it. Deployment is folded into week 8. The claude.md file is the persistent context that shapes every session. Get this right and everything downstream improves. Read ~45 min : - Hannah Stulberg, "Claude Code for Everything: A CLAUDE.md Deep Dive" https://hannahstulberg.substack.com/p/claude-code-for-everything-the-best-personal-assistant-remembers-everything-about-you https://hannahstulberg.substack.com/p/claude-code-for-everything-the-best-personal-assistant-remembers-everything-about-you Practical ~45 min : Be an active learner as you read. Open Claude Code and build a claude.md for a real workflow you own. Test it. Bring it to the seminar, along with a view on how we should structure shared claude.md files across clients and functions. The shift from "write a good prompt" to "control what surrounds the prompt". Your claude.md files, project files, folder structure, and MCPs are the context. It has a hard limit, like RAM on an old computer. Watch ~15 min : - "How Super Mario Bros Was Made Into 40 Kilobytes" https://www.youtube.com/watch?v=vH-D6GdkyCw https://www.youtube.com/watch?v=vH-D6GdkyCw Read ~30 min : - Anthropic, "Effective Context Engineering for AI Agents" https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents Optional: - 3Blue1Brown, "Transformers, the tech behind LLMs" https://www.youtube.com/watch?v=wjZofJX0v4M https://www.youtube.com/watch?v=wjZofJX0v4M — how the model decides what to pay attention to - Firecrawl, "Context Engineering" https://www.firecrawl.dev/blog/context-engineering https://www.firecrawl.dev/blog/context-engineering Practical ~45 min : Audit your claude.md files from week 1, personal and project-level. For every line ask: is this earning its place, or wasting space? Rewrite them. Bring the before and after. Come with an answer to: What's the "clouds and bushes" of your claude.md? In Super Mario Bros, clouds and bushes are the same sprite in different colours. What in your context could be written once and reused? Building skills, agents and software means interacting with third-party services. There are three ways in, and they're all the same thing underneath: APIs are the raw connection, MCPs are pre-configured wrappers that make API calls for the model, and CLI tools are small, sharp programs that do one job brilliantly. Claude is fluent in all three. You need to know which to reach for. Watch ~30 min : - Exponent, "APIs Explained in 4 Minutes " https://www.youtube.com/watch?v=bxuYDT-BWaI https://www.youtube.com/watch?v=bxuYDT-BWaI — the 4-minute grounding - Greg Isenberg & Ras Mic, "Model Context Protocol MCP , clearly explained why it matters " https://www.youtube.com/watch?v=7j NE6Pjv-E https://www.youtube.com/watch?v=7j NE6Pjv-E ~23 min — the most-watched plain-English MCP explainer 1.3M views . Why MCPs exist, what problem they solve, why they're "APIs for AI" Read ~25 min : - IBM, "What Is Tool Calling?" https://www.ibm.com/think/topics/tool-calling https://www.ibm.com/think/topics/tool-calling ~10 min — how an AI model actually decides to call an API. This is the bridge between "what's an API" and "what Claude does with one" - Firecrawl, "What is a CLI and Why AI Agents Prefer It" https://www.firecrawl.dev/blog/why-is-cli https://www.firecrawl.dev/blog/why-is-cli ~15 min — why the command line fits how a model thinks, and why CLI tools often beat MCPs on tokens Optional: - MDN, "An overview of HTTP" https://developer.mozilla.org/en-US/docs/Web/HTTP/Overview https://developer.mozilla.org/en-US/docs/Web/HTTP/Overview — the layer underneath all of it Practical ~45 min , three parts: Part 1 — make one raw API call ~10 min : 1. In your terminal: curl "https://api.open-meteo.com/v1/forecast?latitude=51.5&longitude=-0.12¤t weather=true" — that's London, and you've just made an HTTP GET request 2. Break it on purpose: curl "https://api.open-meteo.com/v1/nonsense" — read the error 3. Ask Claude to get the weather in Paris the same way. Is its request the same as yours? Part 2 — connect an MCP server to Claude Code ~15 min : Follow the official guide: "Connect Claude Code to tools via MCP" https://code.claude.com/docs/en/mcp https://code.claude.com/docs/en/mcp . Pick one MCP server you'd actually use Notion, Slack, Google Drive . Add it, restart, and ask Claude to do something real through it. Then ask Claude: "what tools did that MCP give you?" You'll see the API underneath. Part 3 — feel the power of CLI tools ~20 min : Pick one and ask Claude to drive it: - yt-dlp — "download this YouTube ad as an mp4" - ffmpeg — "convert it to a sub-10mb mp4" or "pull a thumbnail from 3 seconds in" - imagemagick — "batch-resize these ad creatives to 1080x1350" - jq — "pull just the temperature out of that weather response" The lesson: single-purpose tools plus precise instructions are absurdly powerful. "Convert these videos using ffmpeg to sub-10mb mp4s" beats "make these files smaller". Come with: one thing you made a CLI tool do, and an answer to "when would you choose an MCP over just letting Claude run a CLI?" Week 3 covered how data moves. This week is where it lives, and why the format matters. The core premise: markdown is AI's native language. Knowing why, and when to use markdown vs JSON vs a database, is the highest-leverage AI skill you can build. The centrepiece is Karpathy's LLM wiki: a 400,000-word knowledge base maintained entirely by an LLM, all in plain markdown. Watch ~40 min, in order : - DevXOps Tech, ".md Format — Why AI Loves It So Much" https://www.youtube.com/watch?v= 220GNTbevU https://www.youtube.com/watch?v= 220GNTbevU ~6 min - ritvikmath, "RAG Explained in 8 Minutes" https://www.youtube.com/watch?v=HREbdmOSQ18 https://www.youtube.com/watch?v=HREbdmOSQ18 ~8 min - "Andrej Karpathy Just 10x'd Everyone's Claude Code" https://www.youtube.com/watch?v=sboNwYmH3AY https://www.youtube.com/watch?v=sboNwYmH3AY ~18 min — the main event - Computerphile, "JSON, not Jason" https://www.youtube.com/watch?v=A0hoqSkyY7o https://www.youtube.com/watch?v=A0hoqSkyY7o ~6 min Read ~40 min : - Karpathy's LLM Wiki gist https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f ~20 min — the source material. Read the whole thing. The compiler analogy is the key: raw sources are the source code, the LLM is the compiler, the wiki is the executable - SuperGeekery, "Designing Better Markdown for LLMs" https://supergeekery.com/blog/designing-better-markdown-for-llms https://supergeekery.com/blog/designing-better-markdown-for-llms ~15 min, July 2026 — the fresh follow-up to "why AI loves markdown": how to structure markdown so models read it well Optional: - Anthropic, "Introducing Contextual Retrieval" https://www.anthropic.com/engineering/contextual-retrieval https://www.anthropic.com/engineering/contextual-retrieval — for anyone who wants the full RAG picture - "Why Every AI Tool Writes in Markdown" https://unmarkdown.com/blog/why-every-ai-tool-writes-in-markdown https://unmarkdown.com/blog/why-every-ai-tool-writes-in-markdown - Starmorph, "How to Build Karpathy's LLM Wiki" https://blog.starmorph.com/blog/karpathy-llm-wiki-knowledge-base-guide https://blog.starmorph.com/blog/karpathy-llm-wiki-knowledge-base-guide Practical ~30 min : Open clients.json in the Ballpoint repo. 1. How many clients are in the file? How is the list structured? 2. Pick a client. What platforms are they on? What are their target metrics? 3. Find the Slack channel IDs for a client. What's the difference from internal to client-facing? 4. How is this different from describing the same client in a sentence? Then build your own wiki, properly: pick one scattered area of knowledge you own a client, a process, a craft and gather at least 30 raw sources. Call transcripts, Slack threads, Notion pages, briefs, reports, emails, whatever you've got. Dump them in a raw/ folder and have Claude compile them into structured, cross-linked markdown notes plus an index file. Don't tidy the sources first, that's Claude's job. Then interrogate it: "what patterns do you see across all of this?" At 2-3 sources a wiki is a party trick. At 30+ it starts answering questions you couldn't answer yourself, which is the whole point of Karpathy's system. Come with: your wiki, the most surprising thing it told you, and one answer to "what area of Ballpoint knowledge is currently scattered and could benefit from being structured?" AI by default sounds correct but is often wrong. You need a way of deciding whether output is actually good. That's an eval: define success, run the task, check the result against that standard. This week isn't about eval code, it's about the habit of checking AI work before trusting it. Read ~60 min : - Aakash Gupta, "AI Evals Explained Simply" https://www.news.aakashg.com/p/ai-evals-explained-simply https://www.news.aakashg.com/p/ai-evals-explained-simply ~30-40 min read, or ~60 min watch — the primer - Anthropic, "Demystifying evals for AI agents" https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents — the intro, "The structure of an evaluation", and "Why build evaluations?" ~20 min Optional: - Dario Amodei on Dwarkesh Patel, first 29 min https://www.youtube.com/watch?v=n1E9IZfvGMA https://www.youtube.com/watch?v=n1E9IZfvGMA Practical ~40 min : Pick one real AI task you already do at work. Before running it, write a tiny scorecard: what must a good answer include, what would make it misleading or dangerous, what would you still verify manually? Run the task. Then ask Claude: what assumptions did you make, what are the three most likely ways this is wrong, what would a sceptical reviewer challenge? Mark the output against your scorecard: passed, partial, failed. Bring all three to the seminar. Come with an answer to: What's one AI task where you've been relying on "that looks about right"? What would a lightweight eval for it look like? The next level up from shipping skills is knowing when to leave something as a skill, when to chain it into a workflow, and when to give Claude the keys and call it an agent. The field has moved fast: the 2026 consensus is that the value isn't in the agent, it's in the system around it. Skills, verification, and the human in the outer loop. Watch ~45 min : - "Building Great Agent Skills: The Missing Manual" — AI Engineer World's Fair 2026 https://www.youtube.com/watch?v=UNzCG3lw6O0 https://www.youtube.com/watch?v=UNzCG3lw6O0 ~21 min, June 2026 — how skills actually work under the hood user-invoked vs model-invoked, the description as a context pointer and what separates a good skill from skill hell - Philipp Schmid, Google DeepMind, "Don't Ship Skills Without Evals" — AI Engineer World's Fair 2026 https://www.youtube.com/watch?v=0vphxNt4wyk https://www.youtube.com/watch?v=0vphxNt4wyk ~22 min, July 2026 — the week 5 lesson applied directly to what we ship: most published skills have no evals, AI-generated skills can hurt performance, human-written beats generated, keep skill files under 500 lines Read ~35 min : - Anthropic, "Building Effective Agents" https://www.anthropic.com/research/building-effective-agents https://www.anthropic.com/research/building-effective-agents ~17 min — still the canon: prompt chaining, routing, orchestrator-worker, evaluator-optimiser. Most "agents" should be workflows - Latent Space, "5 Trends That Defined AI Engineering at World's Fair 2026" https://www.latent.space/p/aiewf26trends https://www.latent.space/p/aiewf26trends ~15 min, July 2026 — where the field is right now: systems around agents, the outer loop, loop engineering a taster for week 8 Optional: - Anthropic, "The 2026 State of AI Agents Report" https://resources.anthropic.com/hubfs/The%202026%20State%20of%20AI%20Agents%20Report.pdf https://resources.anthropic.com/hubfs/The%202026%20State%20of%20AI%20Agents%20Report.pdf August 2026 — skim for what other teams are actually deploying - Lilian Weng, "Harness Engineering for Self-Improvement" https://lilianweng.github.io/posts/2026-07-04-harness/ https://lilianweng.github.io/posts/2026-07-04-harness/ — deep but defining the current thinking - Simon Willison, "I think 'agent' may finally have a definition" https://simonwillison.net/2025/Sep/18/agents/ https://simonwillison.net/2025/Sep/18/agents/ ~5 min Practical ~40 min : Audit one of your existing skills against this week's material. Is it a prompt, a workflow, or an agent? Is it under 500 lines? Does it have any check on its own output? Map it to one of Anthropic's patterns and reshape it. Come with: your reshaped skill, and a position on "where should the human sit in the loop for the work you run?" The end-to-end AI engineering workflow: brainstorm, plan, execute, review, and make every failure compound into a system improvement. This is how you get consistently good output rather than vibe-coding and hoping. Watch ~30 min : - "Compound Engineering: the AI coding workflow that actually learns" https://www.youtube.com/watch?v=4hLJ62m3OqI https://www.youtube.com/watch?v=4hLJ62m3OqI Read ~55 min : - Every, "Compound Engineering: How Every Codes With Agents" https://every.to/chain-of-thought/compound-engineering-how-every-codes-with-agents https://every.to/chain-of-thought/compound-engineering-how-every-codes-with-agents ~45 min - Every, "Stop Coding and Start Planning" https://every.to/source-code/stop-coding-and-start-planning https://every.to/source-code/stop-coding-and-start-planning ~10 min Optional: - "How I Vibe Coded a Recipe App using Claude Code Full Build + Marketing " https://www.youtube.com/watch?v=9atd5lczG2k https://www.youtube.com/watch?v=9atd5lczG2k — a full build end to end, useful to see the contrast with the CE method Before the seminar: install the Compound Engineering plugin in Claude Code https://github.com/EveryInc/compound-engineering-plugin https://github.com/EveryInc/compound-engineering-plugin . Practical ~40 min : Run the full loop on one real spec: /ce-brainstorm → /ce-plan → /ce-work → /ce-code-review → /ce-compound. Aim for a working v1. No longer do we prompt AI to do stuff: we give it parameters to go away and build big projects for us. Loops combine everything so far. Evals become critical, and so do clarity of thought, purpose, and good scoping. And a loop that only runs with your laptop open is half a loop, so this week also covers deployment. Watch ~1h 20min : - Anthropic workshop, "Build agents that run for hours" https://www.youtube.com/watch?v=mR-WAvEPRwE https://www.youtube.com/watch?v=mR-WAvEPRwE ~1h, the big one - "How to use Claude Routines better than 99% of people" https://www.youtube.com/watch?v=LLE6oMh7SxE https://www.youtube.com/watch?v=LLE6oMh7SxE ~21 min — the deployment layer: getting your loop running on a schedule without you Read ~15 min : - Every, "Loops for Non-Coders" https://every.to/context-window/loops-for-non-coders https://every.to/context-window/loops-for-non-coders - Claude Code Routines docs https://code.claude.com/docs/en/routines https://code.claude.com/docs/en/routines — skim Pondering: - Boris Cherny on loops https://x.com/bcherny/status/2071379474277613732 https://x.com/bcherny/status/2071379474277613732 Optional: - "The Great Loops Debate" — AI Engineer World's Fair 2026 https://www.youtube.com/watch?v=c35YoMdnI78 https://www.youtube.com/watch?v=c35YoMdnI78 — can fully autonomous agents manage their own loops yet? Both sides argued live - Addy Osmani, "Loop Engineering" https://addyosmani.com/blog/loop-engineering/ https://addyosmani.com/blog/loop-engineering/ - "Loop Engineering: Why Every AI Engineer is Learning It" https://open.substack.com/pub/emergingai/p/loop-engineering-why-every-ai-engineer https://open.substack.com/pub/emergingai/p/loop-engineering-why-every-ai-engineer Practical: Design one loop for a real task you own: the goal, the success criteria your week 5 scorecard , the stopping condition, and where it should run locally or as a routine . Bring the design, and if you got it running, the result.