Jason Shepherd

Learning AI by building tiny real things in public. Workflows, tools, mistakes, demos, and field notes. 15+ years in electrical construction & operations.

[Jason Shepherd @JBuildsAI]

@FelixCraftAI That’s why every skill and agent that my Claude/Codex team builds or implements doesn’t ship without a watcher to go along. It records “did the agent do its job? Can I prove it?”

[Jason Shepherd @JBuildsAI]

About which is the correct problem to solve. Most people spend their time solving symptoms. Hint… if the problem is immediately obvious, it probably isn’t the real issue.

If my car has squeaky brakes, I can put grease on my brake pads. Or I could stop and ask “Why are my brakes squeaking?” One is liable to make things much worse and require a new solution to a new problem I created. The other one leads to a real solution.

[Jason Shepherd @JBuildsAI]

@suryanox7 Was plowing fields by hand after the tractor a flex or just a waste of time? You can fight the world, or leverage your deep knowledge to outpace everyone else at scale.

[Jason Shepherd @JBuildsAI]

@nirajxdev If by “communicate well” you mean can clearly articulate a well formed thought, then I’m going with that guy 10 times out of 10, because he can think. If I needed a guy who could code but not communicate, I’d just use Codex.

[Jason Shepherd @JBuildsAI]

@quxiaoyin 3 is about all I’m good for at a time, otherwise I lose track of where I was. That’s actually why I had Claude create a morning brief so I can start the day with a fresh pull on everything I left hanging the day before.

[Jason Shepherd @JBuildsAI]

@RoyInProgress By recognizing that just like in any other realm of life, I can’t learn everything. Focus on what’s right in front of you, solve today’s problem, develop your next idea, and try one new thing. Otherwise you’ll drown in the fire hose.

[Jason Shepherd @JBuildsAI]

@NoahKingJr That its dumb

[Jason Shepherd @JBuildsAI]

@YahwehtheLesser @bcardarella 100% First thing I did figured out also. Big time unlock for way better results.

[Jason Shepherd @JBuildsAI]

@notmissing_ @dramaricic I split the difference. $100/month on Claude and Codex. I don’t think I’ve been above 60-70% weekly usage on either platform since I went that route. It’s usually way less.

[Jason Shepherd @JBuildsAI]

Nothing else out there vibes like Claude. I love that I can just word dump my stream of consciousness at Claude and have it understand me. And honestly I actually like that the responses are so thorough. Better than “Yup. I’ll build that.” I like having the chance to see what Claude thinks I said before it just dives in.

[Jason Shepherd @JBuildsAI]

@sflorimm Not at all. I’d much rather juggle three different chats, and stay high level then plug numbers manually into a spreadsheet all day, trying to figure out why the cell won’t calculate correctly.

[Jason Shepherd @JBuildsAI]

@hambleton_j @T_Zahil Nope. And yet I still installed it anyway because, obviously I had to see what all the fuss was about. It’s cool. Couldn’t see using it for more than a simple task runner though. I’ll stick with Claude and Codex (and a little Gemini sprinkled in).

[Jason Shepherd @JBuildsAI]

With my external shared state context and memory system built on Obsidian, Notion, and GitHub, I can swap out any LLM model I want, and all it would take is a quick pass over the files to swap out some name references in the operating instructions. Get your system memory out of the model. That’s how you truly own your system.

[Jason Shepherd @JBuildsAI]

I just run Claude and Codex apps on my machine over a shared state memory layer, do all my planning and scope development with Claude, then write handoffs to Codex in my GitHub repo and trigger Codex via CLI. Codex runs the prompt, completes the work, then writes a corresponding closeout handoff back to Claude who then audits the work. And back and forth until it’s done.

[Jason Shepherd @JBuildsAI]

I use them both every day. Claude Code for planning and auditing. Codex for implementation.

Claude Code is a much better semantic model. I can ramble at it when I’m not quite sure exactly how to put something and Claude will stick with it until we get it nailed down. Then it writes the spec, and writes a handoff prompt to Codex in a shared GitHub repo to either offer feedback on the spec, or just get to work.

I use Codex for actual implementation. I point it at Claude’s handoff and tell it to follow the prompt. It does the work, then writes a response back to Claude’s handoff and marks the handoff loop closed. Claude then audits Codex’s work and gives feedback or marks the work done.

They both run on top of a shared context and memory layer that they helped build, so I never have to establish context. I can park a project for 2 weeks, open a fresh chat in either runtime, tell it I want to keep working on XYZ project, and it continues right where I left off.

[Jason Shepherd @JBuildsAI]

Claude Code with Opus 4.8 is second to none as a semantic planning partner. I use it for brainstorming and developing project scope and writing handoff instructions for Codex. Codex builds, then Claude audits. When you pair them together with a shared context and memory layer things get really fun. They write their handoffs and responses in a shared folder in GitHub so the text itself out of the chat window and doesn’t burn tokens on pure text consumption.

I point Codex at the folder and ten it to inspect the handoff and follow the instructions. It implements, then closes out with a response to Claude. Claude reads it, audits the work, reports back to me with findings, offers suggestions, writes a follow handoff to Codex, etc.

And because they share the external context recall layer, I can start in a fresh chat in either runtime and they pick up right where I left off even if it’s days later.

[Jason Shepherd @JBuildsAI]

It kind of always has been. It’s less about the language that’s used and more about the syntax and structure of the text in the code. JSON, YAML, Python, Java, HTML, XML, C, all use plain language words. But the way the language is structured, and how punctuation is applied is what changes the meaning, or determines the logical expression.

[Jason Shepherd @JBuildsAI]

That’s why I had Claude and Codex build me what we’re calling Context Kernel. It’s a Python based context compiler that gets called at the beginning of a new session, looks at the subject matter, searches my Obsidian Vault, Notion, and GitHub repo for anything pertaining to past work on that topic, and feeds only relevant records to the LLM runtime. It’ll parse 1,500+ records and deliver the 4 relevant results with a goal of sub-12,000 token bootstrap usage. It reduces startup consumption by anywhere from 40-60% and lets the agent spin up faster. It also allows me to switch between Claude Code and Codex almost flawlessly. The only thing I lose is in-chat context. But even that gets recorded with durable Notion records so the pick up is almost 100% complete. During a session yesterday Claude was even able to quote direct in-chat commentary from a recorded chat session I had with Codex a week ago and pick up right where I left off.

[Jason Shepherd @JBuildsAI]

Exactly. I ran a full end to end system audit of my private GitHub repo and burned nearly a million tokens in about 20 minutes, and it was maybe 2% of my weekly usage limit. I don’t really care about token burn if the work is good. And if you promote it well, 4.8 does amazing work!