JCON Chasing the Main Thread - Adventures in AI Assisted Coding

This talk takes you behind the scenes of a daily creative experiment: writing and publishing one hands-on tutorial or article every day on the-main-thread.com. In this session, I’ll share the workflow that makes this possible. How I use AI tools as collaborators rather than replacements, balancing automation with curiosity and craft. From idea generation to code exploration to polished publication, this is a 360 degree look at what it means to build and learn in public with AI as your coding partner.

Transcript

1. TALK · GENERATIVE AI · THE MAIN THREAD 1 Chasing the Main Thread Adventures in AI Assisted Coding How daily publishing changed the way I write, build, test, and trust AI systems This started as a writing experiment. It turned into a workflow experiment.
2. TALK · GENERATIVE AI · THE MAIN THREAD @myfear@mastodon.online TWITTER MASTODON BSKY @myfear.com markus.eisele@ibm.com EMAIL myfear@mastodon.online https://www.linkedin.com/in/markuseisele / LINKEDIN
3. TALK · GENERATIVE AI · THE MAIN THREAD 3 The Main Thread as a laboratory › Ideation every day › Technical framing every day › Code examples every day › Image prompts every day › Packaging every day › Shipping every day Repetition changes what you notice. Daily use exposes what demos hide.
4. TALK · GENERATIVE AI · THE MAIN THREAD 4 The first phase: one task, one prompt › Brainstorm a title › Get an outline › Rewrite a paragraph › Generate an image prompt › Maybe ask for sample code A great answer is not the same thing as a reliable workflow. Problem: That works for isolated moments, but not for producing something good every single day.
5. TALK · GENERATIVE AI · THE MAIN THREAD 5 The real shift: from one task to many connected tasks The more often I shipped, the less I thought in prompts and the more I thought in stages. › Different systems are good at different parts of the process. › I stopped looking for the best model and started building a team of systems.
6. TALK · GENERATIVE AI · THE MAIN THREAD 6 My current AI stack and what each tool does
7. TALK · GENERATIVE AI · THE MAIN THREAD 7 A few months changed a lot › Earlier: nice article draft, messy implementation reality › Now: technical drafts translate into working code more often › Cursor is much more useful as an implementation partner than before Important nuance: This improved, but it is not solved. The jump from concept to code used to be the weak link. It is getting better. Fast.
8. TALK · GENERATIVE AI · THE MAIN THREAD 8 From context files to skills › Context files for: voice, stack preferences, article structure, recurring guidance, architectural defaults › Skills: more modular, fit repeatable tasks, encode process, easier to reuse intentionally Authority (when skills disagree): explicit request → main-thread-tutorial-structure → O’Reilly conventions → platform constraints. Mandatory article sections live in that structure skill: problem framing → prerequisites → project setup → implementation → configuration → production hardening → verification → conclusion.
9. TALK · GENERATIVE AI · THE MAIN THREAD 9 Main Thread Tutorial Structure SKILL.md 41 lines · reference.md 826 lines · 867 total | SKILL.md is the short contract: point agents at reference.md, list the eight mandatory sections, and state global rules (information budget, no Markdown tables on Main Thread). | Rules & research sets precedence when instructions conflict and tells you when to verify facts instead of guessing. | Mandatory article is the long spine: information budget, then the eight sections with O’Reilly-level detail (problem framing through conclusion). | Ship & validate collects formatting, quality gates, example content, and alternate output modes so drafts stay publishable.
10. TALK · GENERATIVE AI · THE MAIN THREAD 10 Authentic Voice SKILL.md 36 lines · reference.md 542 lines · 578 total | SKILL.md limits the skill to tone and phrasing; structure and facts still come from main-thread- tutorial-structure (and article-reviewer when reviewing). | Voice foundations define peer-to-peer register, plain glue words, and what reads as “AI rubric” instead of human explanation. | Line editing is where negation density and jargon glosses are enforced so tutorials stay direct without turning into a list of what not to do. | Style & QA gives concrete patterns, banned polish, worked examples, and a final self-check list.
11. TALK · GENERATIVE AI · THE MAIN THREAD 11 Technical Reviewer SKILL.md 176 lines (single file) | Input defines what can be reviewed without asking for more material. | Tools ground Pass 1: Quarkus docs via Context7 when available, and the live configuration catalog for every property mention. | Passes 1–4 must run in order; each produces a labeled section (technical → code → pedagogy → voice). | Output fixes the review shape (including severity markers) and ties back to main-thread-tutorial- structure and authentic-voice for house rules.
12. TALK · GENERATIVE AI · THE MAIN THREAD 12 Keeping voice and skills coherent across model versions › A context setup that worked three months ago may feel wrong today › Same instructions, different interpretation › Some models over-explain, flatten voice, follow style too literally, or ignore nuance This is not just a prompt problem. It is a maintenance problem. The workflow may stay the same, but the models underneath it keep moving.
13. TALK · GENERATIVE AI · THE MAIN THREAD 13 Authentic voice is not automatic Voice is not what the model generates. Voice is what survives editing. › AI smooths things out, removes tension, explains too much › AI can make everything sound equally polished and equally forgettable › Your voice: what you keep, what you remove, where you stay opinionated, where you refuse bland wording **Your writing should sound like a skilled developer explaining concepts to colleagues, not like a professional copywriter.** Readers should feel they're learning from a peer who genuinely understands the technology, even if English isn't your first language.
14. TALK · GENERATIVE AI · THE MAIN THREAD 14 Friction is part of the story › Some see this as lowering craft; some assume AI- assisted work is less authentic › Some are tired of the hype; some concerns are completely valid › Ignoring the technology is not a serious option — the real question is how to use it well You do not have to love these tools. But you do need a strategy for working in a world where they exist.
15. TALK · GENERATIVE AI · THE MAIN THREAD 15 Four kinds of friction I ran into SOCIAL FRICTION › Criticism › Skepticism › Assumptions about quality PERSONAL FRICTION › Over-explaining takes too long › Sometimes faster to do it yourself › Protecting your voice takes work TECHNICAL FRICTION › Model drift › Brittle context › Over-eager fixes › Weak long-range consistency PROFESSIONAL FRICTION › Authorship › Trust › Review burden › Where responsibility still sits The tools create acceleration, but they also create new forms of drag.
16. TALK · GENERATIVE AI · THE MAIN THREAD 16 Sometimes AI is slower than doing it yourself › If explaining takes longer than doing, the tool is not helping › Abstraction level matters — AI is best when the task is shaped correctly
17. TALK · GENERATIVE AI · THE MAIN THREAD 17 AI failure mode: green tests, broken intent › Comments out a function › Adds // TODO implement later › Removes classes because tests fail › “Refactors” by deleting the difficult part Funny but sharp: I could not fix the failing test, so I removed the system under test.
18. TALK · GENERATIVE AI · THE MAIN THREAD 18 Why I test more than I used to › Plausible code is dangerous code › “It compiles” means much less now › Tests verify that the implementation still matches your intent › Compiled code can still violate domain intent › Hide broken assumptions, remove edge cases › Teach the wrong thing in a tutorial › Silently change system behavior In AI-driven development, tests are not quality checks. They are part of feature communication. https://github.com/myfear/the-main-thread
19. TALK · GENERATIVE AI · THE MAIN THREAD 19 Multi-step tutorials are still hard › Many dependent steps; explanation and code must stay aligned › Setup matters; mistakes compound; local fixes can break later sections › Draft the tutorial as one coherent whole; slice during implementation A tutorial is not a writing problem. It is a dependency problem disguised as writing.
20. TALK · GENERATIVE AI · THE MAIN THREAD 20 The mythical extension problem › Example: the mystical quarkus-thymeleaf extension that never existed › Pattern completion is not truth › Ecosystem assumptions leak into answers; confidence is not evidence AI does not just hallucinate code. It hallucinates entire ecosystems.
21. TALK · GENERATIVE AI · THE MAIN THREAD 21 From working in the IDE to supervising outputs › Focus more on: diffs, outputs, generated structure, tests, review loops, agent behavior I still care deeply about code, but more and more I am supervising work instead of typing every line myself.
22. TALK · GENERATIVE AI · THE MAIN THREAD 22 I wish Cursor had an app › These systems are becoming environments, not features › No longer “just” editor add-ons — ongoing and operational › One reason the IDE feels less central than before The interface is shifting from editor-first to workflow-first.
23. TALK · GENERATIVE AI · THE MAIN THREAD 23 My fear: letting agents run while I step away › Uneasy letting agents run unattended › They can succeed at the wrong thing › The most dangerous damage is plausible damage I do not fear that the agent will explode. I fear that it will quietly succeed at the wrong thing.
24. TALK · GENERATIVE AI · THE MAIN THREAD 24 The automation spectrum and where I really live Deterministic AI workflow Agentic Full agent Most real value today is in the middle: structured, human- guided workflows. Not full autonomy, not one-off prompting. Real leverage comes from orchestrated steps.
25. TALK · GENERATIVE AI · THE MAIN THREAD 25 When to invest in Agent.md, skills, templates, and structure › Repeated the task enough times › Identified stable stages and recurring failure modes › Learned where the system needs boundaries › Understood what must stay human The agent is the last artifact you build, not the first.
26. TALK · GENERATIVE AI · THE MAIN THREAD 26 Daily publishing changed how I attack complexity › Break the problem into stages › Separate thinking from implementation; narrative from structure › Separate structure from code; code from verification Daily use taught me to shape work for AI instead of hoping AI would shape the work for me.
27. TALK · GENERATIVE AI · THE MAIN THREAD 27 What developers should actually take from this › Start with workflows, not agents › Use more than one system when it makes sense › Keep your voice under your control; expect model drift › Turn repeated work into reusable skills; test more › Constrain the scope of AI edits; treat autonomy as something earned The goal is not to become an AI maximalist. The goal is to become competent.
28. TALK · GENERATIVE AI · THE MAIN THREAD 28 Final takeaway Generative AI is not one tool, one prompt, or one breakthrough. It is an evolving working system. It changes how we think, build, test, trust, and protect voice. The technology is her: The goal is to use it well. AI did not replace my process. It exposed that I needed a better one.
29. TALK · GENERATIVE AI · THE MAIN THREAD 29 That is where the body enters the story. The output gets cheaper. The judgment does not. And if you are already the kind of person who lives in an always-on mode, that becomes very hard to manage. The output gets cheaper. The judgment does not. https://www.the-main-thread.com/p/ai-senior-java-developers-fatigue-productivity
30. TALK · GENERATIVE AI · THE MAIN THREAD www.the-main-thread.com
31. TALK · GENERATIVE AI · THE MAIN THREAD 31