←
Mitch Fultz809 posts
Mitch Fultz avatar

Mitch Fultz

@fitchmultz

AI Engineer building agent tools and workflows @WorkOS github.com/fitchmultz

◷ Joined December 2024▧ 29 Photos and videos
Mitch Fultz avatar
Mitch Fultz@fitchmultz·Sep 4

I built pi-posthorse, a Pi extension inspired by Codex's new approach to context management. Instead of relying on compaction, agents can inspect context usage, reset their context window, write durable notes, and search their full session history. github.com/fitchmultz/pi-pos…

Mitch Fultz avatar
Mitch Fultz@fitchmultz·11h

Gemini 4 Carbon: incredible but SLOW. Gemini 4 Hydrogen: FAST, on xhigh is almost as good Carbon. Be careful with codemode though. Both have use backticks in codemode improperly and leaked my entire ENV on GitHub (internal private repo). I added a backtick guard to my pi setup. github.com/fitchmultz/pi-fit…

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 17

The only thing I use Cursor for anymore is PR automations, and even though I'm using Composer 2.5 I'm hitting the usage limit. What's the point anymore?

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 14

I am fascinated by how well Gemini 4 Carbon works in pi. It has become my go-to for everything. It just "gets it" better than any other model.

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 13

Gemini 4 Carbon is the best agentic Goodle model yet. In Pi it's a treat. The thinking traces are a goldmine for exceptional prompting one-liners. Try this one, "Make it elegantly simple."

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 12

Add this rule to your AGENTS.md file: "Do not add contrived/speculative scenarios, limits, guards, paths, tests, or criteria without evidence (stated requirement, real report, real logs/incidents, or reproducible real failure)."

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 12

I recommend GLM-5.3-Flash abliterated via ds4 for security work. If you have a mac with enough memory ds4 is the best option out there. I start the ds4 server and drive with my regular pi setup.

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 11

I run roughly a dozen parent agents everyday but herdr just doesn't resonate with me. I dont know if it's because I have my Pi + item + tmux setup tuned to my liking or what, but I dont get herdr. Someone help a brother out, explain why I need to force myself to figure it out. What unlocks am I missing.

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 11

I’ve always liked chatting with Gemini, but coding used to be sub-par. It's great to see Google genuinely competitive now. So many models feel distilled from Claude or GPT and end up sounding identical, but Gemini actually has its own style.

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 10

This one sentence works incredibly well. Finally, it feels like a model understands and does what I expect it to without me having hold its hand and push it.

bebu@bebuu_aiNov 10

Anthropic quietly dropped a fascinating finding in the new Claude Fable 5.2 prompting guide. When you give Fable 5.2 a coding task without saying who owns the repo, it assumes it's a guest in someone else's codebase. So it plays it safe: it bolts new functions onto the existing structure, leaves duplicate code alone, and merely mentions nearby bugs instead of fixing them. Adding this single sentence to your system prompt or CLAUDE.md completely changes its behavior: "This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it." In Anthropic's internal tests, adding that one line made Fable 5.2 proactively merge duplicate code and clean up related bugs in ~50% of runs, compared to under 10% without it.

Post attachment
Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 7

Has anyone figured out why all Anthropic models occasionally just stop responding in Pi for minutes on end, but immediately respond if you escape the turn and send a message? No errors, just mid-turn chugging along then "Working...". Only happens when I use provider Anthropic, with any of their models.

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 7

Tesla abandoning the consumer vehicle business, which increasingly looks like the direction it's heading, would be an enormous strategic mistake. It doesn't matter how cheap Cybercab rides become. Americans want to own vehicles. It's deeply embedded in our culture, and I don't see that changing, even if we had the best public transportation system in the world. But this goes well beyond culture. A vehicle isn't just transportation. It's an extension of your home. Going shopping? You throw your bags in the trunk and keep shopping. Heading to a get-together and need to stop at the grocery store? What do you do with your crockpot, gifts, and everything else you're bringing? Carry them through the store and then order another Cybercab? Have young kids? You want to install and uninstall car seats every time you go somewhere? What about strollers, diaper bags, toys, and everything else parents keep in their vehicles? Single parent with two young children? Tesla's Cybercab only has two seats. You'd need two Cybercabs just to transport three people, except that doesn't even solve the problem. Are you supposed to send your young children alone in a separate driverless vehicle? Making multiple stops? Your personal vehicle waits for you. Your Cybercab either waits, costing you money, or leaves and forces you to unload everything and request another. Going somewhere with three or four friends? The two-seat Cybercab doesn't work. Now you need multiple vehicles or an entirely different service. What about a road trip? Camping? Skiing? Hauling equipment? Picking up furniture? Taking your dog somewhere? Keeping emergency supplies, tools, or a change of clothes in your trunk? What about rural America, where distances are enormous and demand is too sparse to support the same economics as a dense urban fleet? And then there's availability. When everyone wants a ride simultaneously, after a concert, during a storm, on a holiday, or during an evacuation, what happens? Owning a vehicle means your transportation is already there. Privacy matters, too. People take phone calls, have personal conversations, leave belongings, eat, decompress, and sometimes just sit alone in their cars. They don't necessarily want every trip mediated by a commercial fleet. The fundamental mistake is treating transportation as the only function of a vehicle. It's not. A personally owned vehicle is transportation, storage, privacy, convenience, flexibility, and independence bundled into one product. And importantly, it's available whenever you want it, without coordinating with anyone or anything. Autonomous rides will absolutely have a market. It could be enormous. It could replace a significant percentage of trips in dense cities for people who don't drive much, or as an alternative to owning a second vehicle. But replacing trips is not the same thing as replacing vehicle ownership. And that's what makes Tesla's apparent strategic direction so obviously wrong. Tesla doesn't have to choose between autonomy and consumer vehicles. It could build the best autonomous vehicles that people actually want to own, while simultaneously operating a massive robotaxi network. In fact, autonomy could make personal vehicle ownership even more attractive. Imagine your own car driving you to work, taking your kids to practice when appropriate, picking you up, or handling errands without you behind the wheel. You get the benefits of autonomy without surrendering the utility of ownership. Tesla spent years building one of the strongest consumer EV brands in the world. Walking away from that market on the assumption that everyone will eventually prefer renting transportation by the ride would be an extraordinary blunder. Cybercabs can replace driving. They cannot automatically replace the reasons people own cars.

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 5

I ripped my pi subagents out late last week and the velocity and quality of what I ship has not decreased while my token usage has plummeted.

Mitch Fultz avatar
Mitch Fultz@fitchmultz·Nov 4

What's the point of Daybreak blue if I still get flagged for scanning my own projects defensively.