kumo

ML engineer / vibe coder / AI addicted

open for work

[kumo @kumo0x]

https://t.co/g89bOO7jNV

[kumo @kumo0x]

@thebuggeddev @elonmusk wait, you build it??

[kumo @kumo0x]

8 PEOPLE BEHIND GROK BOT. 10 PROMPTS YOU CAN PASTE TONIGHT

One course collected how the team runs its own bots. The numbers below are as reported there

The pattern is simple: one job per bot, and nothing leaves without your yes

What the setups look like:

> A lead engineer shipped 2,500 changes in a month with a manager bot that never writes code
> A bot that exists only to build other bots, with a list of anti-jobs it must never do
> A savings bot where every saving needs a dollar range and the math, and inventing a number is forbidden
> A quiet mode bot that pings only on urgent things you name and puts the rest into one 8 am summary

The 5 rules worth copying:

> One job per bot, plus a chief of staff for tasks nobody owns
> Nothing goes out until you approve the exact draft
> Show it a task, fix its mistakes, save it as a skill, then schedule it
> One morning summary instead of a stream of pings
> Every number comes from a file or a source, never a guess

And the ladder a bot climbs before it asks you:

> Past chats
> Connected apps
> Web search
> Browser
> The whole computer
> Only then you

My addition: a rule in a prompt is a request, not a lock. Claude Code docs call CLAUDE.md context, not enforced configuration. For "never send, pay or delete" use permissions or a hook

Start here:

1. Ask a bot to read 14 days of email and messages and list 5 jobs it can take over, ranked by hours saved per week
2. Turn the best result into a skill
3. Add one rule: show the draft first
4. Schedule it only after it runs clean

Bookmark this

[Diam @diamai_]

https://t.co/fcDi5khPDh

[kumo @kumo0x]

@0xGoldTooth Agree, the price tag isn't the cost. That's why the article ends on cost per finished task

[kumo @kumo0x]

https://t.co/g89bOO7jNV

[kumo @kumo0x]

ONE PROMPT. A FAKE $319 MEDICAL BILL. $39.60

A tester wrote the audit down as a prompt and ran it on a fake bill with 8 planted errors. Opus 5.5, one message, no follow ups

The trick is not the bot. It is the rule list plus one instruction: cite a rule ID for every finding

That is why it can't just say a charge "looks high". It points at a rule and a line on your paperwork, or it says nothing

What it caught:

> Duplicate lab billed to the patient: $48.00
> Statement asks $92, insurer says $22: $70.00
> "Misc supplies" marked do not bill: $68.00
> $40 check-in payment never credited: $40.00
> Checkup charged coinsurance: $38.00
> Statement fee: $12.00
> Tests unbundled from a panel: $3.40

7 of 8 found, and both decoys that look wrong but are legal were left alone

The miss: 3 units at $42 billed as $162. The model called it correct. The right total is $126. Do units times price yourself

How the bill moved:

> Statement: $319.00
> EOB says you owe: $121.00
> After the billing office fixes its errors and credits the $40: $81.00
> After the insurer appeal on the checkup: $39.60

Two lines in the prompt do most of the work:

> AMBER: the model says "I need the visit note" instead of guessing
> Never invent codes, laws or amounts

Setup:

1. Get the itemized bill with codes, the insurer EOB and your receipts
2. Black out name, member ID and date of birth
3. Ask the model to recompute every line, cite a rule ID per finding, mark unsure items AMBER and name the document it needs
4. Call the billing office with the dispute letter open

The test bill is fictional. Not legal or medical advice

Bookmark this before your next bill arrives

[f⭕️rgent @f0rgent]

https://t.co/X153VRHHY6

[kumo @kumo0x]

@0x_kaero claude is best for motion immo

the main use it proper

[kumo @kumo0x]

@Naiivy9 shit 60 to 94 accuracy quite good

[kumo @kumo0x]

@HeyAIpenguin checked mine too after that. ssh was already off everywhere so i just shrugged and closed the tab

[kumo @kumo0x]

@Klonzu incredibly cheap for much banger

[kumo @kumo0x]

@theparuchh pin us on release

gl on building

[kumo @kumo0x]

@diamai_ grok bot just take my life

[kumo @kumo0x]

@Yumslef micro-duck in every house

[kumo @kumo0x]

@adam_james_hill ai motion going crazy nowadays

[kumo @kumo0x]

@aifeelforme just tweak yourself so me and my girls like you more, that’s hot

[kumo @kumo0x]

I gave the same one-line prompt to Opus, Sonnet and Haiku:

"A pirate ship sailing on the ocean at sunset"

Each wrote a full 3D scene from scratch. Raw WebGL. No libraries. No assets.

Top: Opus, ~$0.71
Middle: Sonnet, ~$0.31
Bottom: Haiku, ~$0.02

The cheapest one burned the most tokens (162K vs Opus's 126K) and still cost ~35x less.

It also added an island and a skull flag.

Be honest: which ship would you pay for? ↓

[kumo @kumo0x]

https://t.co/2zlggyFN9S

[kumo @kumo0x]

https://t.co/2zlggyFN9S

[kumo @kumo0x]

GPT-6 Astra raised SQLite-from-scratch scores from 26.8% to 87.4% across 20 programming languages.

That is OpenAI's MultiLangBench: build a SQLite-compatible database engine in a chosen language, then test its behavior against SQLite-derived checks.

The gain matters most where training data is thin. COBOL rose from 10.9% to 72.1%; Fortran from 21.4% to 83.5%. OpenAI also released the benchmark setup on GitHub, so teams can inspect the tests rather than trust a headline score.

This does not prove a model can safely rewrite a bank's COBOL. It gives legacy teams a more concrete way to test which modernization work is worth scoping.

Source: https://t.co/ZsXnNUNX9c

[kumo @kumo0x]

Letta Code can now coordinate 3 MCP reads in one JavaScript tool call, without launching another agent.

I found CodeMode in v0.34.17, released October 18.

The maintainers' live example starts with Asana's get_me, then runs Asana get_projects and Better Stack teams in parallel. One script returns the combined result.

The model writes the script. JavaScript handles the sequence, parallel calls and filtering. CodeMode itself makes no model calls and cannot spawn subagents.

That is useful for routine agent work: fetch related records, join the data, return only what the next decision needs. The model doesn't have to narrate each plumbing step.

There is a boundary: nested tools still face permission checks. If approval is needed, the script cannot grant it.

I'd try this on a read-only reporting workflow before reaching for a multi-agent setup.

https://t.co/Zpg3TbGLbz

[kumo @kumo0x]

@finnbags @BagsApp BAGGY offer to team up with other hachathon builder in April

0 responds from other teams

https://t.co/Ozxyor5WDA

[kumo @kumo0x]

@finnbags waiting till @finnbags pay attention to BAGGY