Armish 🍁🇵🇸

Building Fielder - dispatch, jobs & AI scheduling for field service teams. Building it in public with an AI agent. Boring problems, real leverage.

[Armish 🍁🇵🇸 @armish_munir]

@TheHashton @marketing_karl Agreed

[Armish 🍁🇵🇸 @armish_munir]

Is it just me, or are you guys hitting the Codex weekly limit way too quickly lately?

[Armish 🍁🇵🇸 @armish_munir]

Time to switch back to claude i guess. 😬

[Armish 🍁🇵🇸 @armish_munir]

@BasantSoli33398 Claude is down

[Armish 🍁🇵🇸 @armish_munir]

Switched to Codex a month ago.

Now everyone’s saying Claude has the better model.

At this point, I’m not choosing AI tools. I’m just chasing them 😭

[Mohammad Zubair @Real_MZubair]

Never imagined we will see so much bloodshed in Azad Kashmir as we witnessed today.
Killing our own people is no way to suppress protests.

[Armish 🍁🇵🇸 @armish_munir]

My AI coding agent has a memory folder and writes its own decision records.

When it makes a real choice, it explains why and reads it back next session. It's building institutional memory I can audit.

Most human teams never keep this.

[Armish 🍁🇵🇸 @armish_munir]

@Tristanrhee3 Exactly. UGC can create interest, but if the app name isn’t obvious, that interest dies instantly.

The fix is simple: show the app name on-screen, say it out loud, and end with a clear “search LaunchPoint on Web” CTA.

Attention without recall doesn’t convert.

[Armish 🍁🇵🇸 @armish_munir]

@ben_sage Need help finishing app #38?

[Armish 🍁🇵🇸 @armish_munir]

@dani_avila7 Loops are deterministic. “Really good subagents” are just loops with personality, latency, and a larger invoice.

[Armish 🍁🇵🇸 @armish_munir]

@trikcode Exactly. The next breakthrough isn’t another benchmark point, it’s making intelligence cheap enough to run everywhere without the CFO opening an incident.

[Armish 🍁🇵🇸 @armish_munir]

@_overment The .md file isn’t the plan, it’s shared state. Without it, your AI agent forgets decisions, repeats mistakes, and confidently improvises architecture.

[Armish 🍁🇵🇸 @armish_munir]

@Tech_girl Without syscall auditing or sandboxing, you wouldn’t. Your AI agent may be `cat .env-ing` secrets like bedtime stories.

[Armish 🍁🇵🇸 @armish_munir]

@itspradz 100%. The moment you have real constraints, it stops being a lottery ticket and starts being a collaborator.

[Armish 🍁🇵🇸 @armish_munir]

I'm building a real SaaS where an AI agent writes most of the code, under a written constitution.

It can't touch the database outside a repository. Can't ship untested code. Every real decision gets written down.

Guardrails, not vibes.

[Armish 🍁🇵🇸 @armish_munir]

@AskYoshik Genuinely interested what changed. Model? Cost? Interface?
I know people going different directions on this. What's the specific win for you?

[Armish 🍁🇵🇸 @armish_munir]

@thdxr That's not a package.json, it's a cry for help.

[Armish 🍁🇵🇸 @armish_munir]

@psomkarbuilds FDE

[Armish 🍁🇵🇸 @armish_munir]

@catalinmpit Was in your camp too. What changed for me: refuse to let it write without a failing test first. Different tool entirely.

[Armish 🍁🇵🇸 @armish_munir]

@ThePrimeagen Same here man. Sunk cost used to be the thing that kept me stuck with a bad structure for months.

Now I just build the other version and compare them side by side. Half the time I go back to the original anyway, but now I actually know why instead of just guessing.