fospa

Open-source AI tools, explained without the hype. Every claim comes with a source.

[fospa @fospa]

@akifyanbak Spent 17 minutes thinking of how to comment this. Still no idea

[fospa @fospa]

5 posts. 173K followers. It's an AI ad.

It's called Saveo House: a daily crime series where every episode is made with AI.

The premise: you just stole $10 million.

How it works:
-> a new episode drops every day at 7pm ET
-> each one ends on a choice, like "comment HIDE or SPEND"
-> the majority vote decides what happens next
-> every episode slips in a real money rule about how your cash gets tracked

The series is called "every dollar leaves a trail". The app it leads to is Saveo: connect your bank accounts and the AI sorts every transaction and shows where your money goes. Its pitch is "track your own trail".

Their site runs the story like a live game, with $8,914,000 cash left, 94% suspicion and 22,710 votes so far.

Nobody is following a budgeting app. They're following a heist, and the app comes with it.

[fospa @fospa]

@Qyravex Sounds cool. Scary part is letting your inbox open for Ai tbh

[fospa @fospa]

Step 1, add this to ~/.claude/settings.json:

{
"model": "opusplan",
"advisorModel": "opus",
"env": {
"CLAUDE_CODE_SUBAGENT_MODEL": "haiku"
}
}

Step 2, create ~/.claude/agents/scout.md:

---
name: scout
description: finds code fast. returns file paths and line numbers only.
model: haiku
tools: Read, Grep, Glob
---
Locate exactly what was asked. Return file:line. No fixes. No opinions.

Step 3, create ~/.claude/agents/critic.md:

---
name: critic
description: reviews a diff or plan before any task is called done
model: opus
tools: Read, Grep, Bash
---
Assume this is broken. Find how it fails. One line per issue, worst first.

One catch: with CLAUDE_CODE_SUBAGENT_MODEL set, some Claude Code versions run every subagent on that model, the critic included. If you want the critic on Opus, check which model it actually used after the first run.

[fospa @fospa]

Claude can split one task across three models, and most people never set it up.

Opus plans and steps in when Sonnet gets stuck. Sonnet writes the code. Haiku handles every file search. A critic agent reviews the diff before you merge.

Three lines in ~/.claude/settings.json:

"model": "opusplan"
"advisorModel": "opus"
CLAUDE_CODE_SUBAGENT_MODEL = "haiku"

Then add two agent files. A scout that only returns file:line, and a critic with one instruction: assume this is broken, find how it fails.

Full config and both agent files in the reply.

[fospa @fospa]

@RohOnChain I still don't have that much trust in the trading bots. I am up to spend something to use AI, but I won't allow AI to spend personal credits.

[fospa @fospa]

One command turns Claude into a reverse engineer that has to show proof for every claim.

It's called REA. Free, open source, MIT-licensed, and it ships with ready-made prompts. Two of them:

- Investigate a feature: trace how one feature works inside an Electron app, a macOS bundle or a compiled binary
- Compare versions: diff two builds and find which behavior actually changed

It also tags every finding as an observation or an inference. Anything it can't confirm gets logged as unknown instead of guessed.

Install: npx rea-agents setup

Use it on software you own or are allowed to test.

[darkzodchi @zodchiii]

https://t.co/y7whCO2uJR

[fospa @fospa]

@skunzu do we really need the ultramode though? Imo, high/medium could work just fine