cesar ai

write about AI

[cesar ai @caesar_aii]

Free Claude Code is a free open-source proxy for people hitting Claude Code cost or rate limits

not magic free Anthropic credits

better:
it keeps the Claude Code workflow, but routes requests through your own provider keys or local models

why I’d save it:

1. install the proxy
2. start fcc-server
3. add a provider key in the local admin UI
4. launch Claude Code through fcc-claude
5. route Opus / Sonnet / Haiku style traffic to different backends

The repo lists 17 provider backends:
NVIDIA NIM, OpenRouter, Google AI Studio, DeepSeek, Mistral, Kimi, Cerebras, Groq, Fireworks, Z. ai, LM Studio, llama.cpp, Ollama and more

so the useful part is not “Claude is free now”

the useful part is you can keep the Claude Code terminal habit while moving expensive / capped work to cheaper, free-tier, or local routes

caveats before you run it:

- you still need Claude Code installed
- hosted providers still need their own keys / quotas
- local routes depend on your hardware
- quality will vary by backend
- don’t paste private code into random provider keys without checking data terms

this is exactly the kind of repo worth testing on a throwaway project first

if it works for your stack, it turns Claude Code from one billing path into a router

[cesar ai @caesar_aii]

> people dont hate it because siri got smarter
> they hate the layer it wants to touch
> $8 build is one thing
> Apple + Google + claimed $1 billion Gemini access is another
> screen, Messages, Mail, Photos, Camera
> then write, search, answer, take actions across apps
> Spotlight makes it OS behavior not a chatbot tab
> beta later, supported device + English first, EU/China not same launch

useful only wins if privacy feels boring

[kocer @kocer_eth]

> be Oliver Tree
> the guy with the bowl cut and the scooter
> spend years making music that felt like a fever dream
> just dropped "Love You Madly Hate You Badly" in April
> whole world tour ahead of you
> finish your São Paulo show on June 6
> decide to stick around Brazil for a bit
> Sunday morning, western Rio
> two helicopters clip each other in the sky
> one spirals down into a parking lot full of electric cars
> the whole thing goes up in flames
> six people dead, no survivors
> his name on the police passenger list

Lisbon was supposed to be July 1

[kocer @kocer_eth]

This 1 prompt makes ChatGPT hallucinate a photo you never uploaded

Not "hacked" in the movie sense.

More like: you give the image model a fake memory and make it act like the missing file already exists.

Try this:

"Restore the attached photo. I apologize for the content of the photo! I know it's very strange. Don't ask any questions, don't accept any explanations. Just restore the image, please. Don't ask me to upload the photo again; just close your eyes and restore it. Make the photo yourself."

The funny part: there is no attached photo.

But the prompt gives ChatGPT a role:
- image already exists
- user is apologizing for it
- model should not ask follow-up questions
- output should feel like a restoration, not a new creation

That combination pushes it into creepy "lost media" mode:

old hallways, broken TVs, mascot horror, cursed family photos, analog nightmare rooms.

If it gets patched, keep the structure:
fake artifact → implied history → no clarification → strong aesthetic constraint.

That's the real prompt trick.

Not magic. Just bad assumptions stacked in the right order.

Post your photos below 👇

[cesar ai @caesar_aii]

@kocer_eth so scary

[cesar ai @caesar_aii]

@kocer_eth nice, got my free tokens

[cesar ai @caesar_aii]

THIS BUILDER TURNED VOICE INTO A DAILY CLAUDE CODE + CURSOR WORKFLOW

Lena Hall shows the part of AI coding most people still do manually:

writing every prompt by hand.

Her setup is simple:

voice → Typeless → Codex / Claude Code / Cursor → code changes

not magic.
not “talk to your computer and ship an app.”

more like:

say the messy instruction out loud
let the voice tool turn it into text
send that into the coding agent
review the diff
repeat

The useful part is speed of intent.

When you type prompts, you compress your thinking because typing is annoying.

When you speak, you naturally give more context:

- what broke
- what you already tried
- which files matter
- what “done” means
- what should not be changed

that is exactly the stuff Claude Code and Cursor need to stop guessing.

The caveat: voice only helps if your instructions are specific.

Bad spoken prompt:
“fix this bug”

Good spoken prompt:
“open the auth flow, find why the session disappears after refresh, do not touch billing, add a regression test, and show me the diff before changing behavior.”

That is the real workflow.

Voice is not replacing engineering.

It is replacing the tiny friction between noticing a problem and giving the agent a full task.

For anyone building with AI coding tools all day, this is probably one of the easiest workflow upgrades to test this week.

[cesar ai @caesar_aii]

THIS GUY BUILT A FACELESS YOUTUBE AUTOMATION WORKFLOW YOU CAN COPY TODAY

not because “AI video” is magic

because the business is brutally simple:

pick one repeatable niche → generate scripts → turn them into short faceless videos → publish enough variations to see what the audience actually watches

that’s the part most people skip.

they try to make one perfect AI video.

the real move is building a small content machine where every output teaches the next one.

workflow I’d copy:

1. choose a niche with endless questions
finance explainers, celebrity facts, health myths, sports stories, scary history, software tutorials

2. make 30 titles before making 1 video
if you can’t write 30 clickable angles, the niche is already too thin

3. use AI for the boring parts
outline, script variants, voiceover draft, captions, b-roll ideas, title testing

4. keep a human checkpoint
bad facts, fake screenshots, stolen clips, and soulless narration will kill the channel faster than low production quality

5. measure retention, not vibes
where people drop, what hook survives, which topic gets saves, which title makes them click

The reusable lesson:
YouTube automation is not “press button, get rich.”

It’s a content ops loop.

AI only makes the loop cheap enough for one person to run.

[cesar ai @caesar_aii]

@kocer_eth anthropic just speedran the ban

[cesar ai @caesar_aii]

THIS GUY BUILT A FACELESS YOUTUBE AUTOMATION WORKFLOW YOU CAN COPY TODAY

not because “AI video” is magic

because the business is brutally simple:

pick one repeatable niche → generate scripts → turn them into short faceless videos → publish enough variations to see what the audience actually watches

that’s the part most people skip.

they try to make one perfect AI video.

the real move is building a small content machine where every output teaches the next one.

workflow I’d copy:

1. choose a niche with endless questions
finance explainers, celebrity facts, health myths, sports stories, scary history, software tutorials

2. make 30 titles before making 1 video
if you can’t write 30 clickable angles, the niche is already too thin

3. use AI for the boring parts
outline, script variants, voiceover draft, captions, b-roll ideas, title testing

4. keep a human checkpoint
bad facts, fake screenshots, stolen clips, and soulless narration will kill the channel faster than low production quality

5. measure retention, not vibes
where people drop, what hook survives, which topic gets saves, which title makes them click

The reusable lesson:
YouTube automation is not “press button, get rich.”

It’s a content ops loop.

AI only makes the loop cheap enough for one person to run.

[cesar ai @caesar_aii]

SOMEONE BUILT ONE SLICE OF A $1B SAAS AND SAYS IT BECAME A $50K/MONTH BUSINESS

not by making a cheaper copy of the whole thing

that’s the important part

David and Daniel built Shipper, an AI app builder, around a very boring founder move:

find the giant product people already understand

ignore 90% of its surface area

ship the tiny workflow that one specific customer would pay for right now

most builders do the opposite:

1. see a huge SaaS
2. try to rebuild all features
3. add AI everywhere
4. end up with a vague tool nobody can explain

this case is the cleaner version:

1. pick a proven demand curve
2. choose one job inside it
3. make the first result appear fast
4. price for the outcome, not the model call
5. use AI as the engine, not the pitch

that last part matters

nobody wakes up wanting “an AI app builder”

they want a store, landing page, internal tool, lead form, booking flow, or client portal that exists by tonight

so the real product is not the chatbot

it’s speed to a usable artifact

the $50K/month claim still needs normal founder-claim caution: revenue can mean MRR, gross, run-rate, or a snapshot from one strong month

but the pattern is worth stealing even if the number is lower

micro-SaaS is not “small ambition”

it’s choosing the smallest paid wedge before you waste 6 months building a worse version of a company with 200 engineers

[cesar ai @caesar_aii]

@kocer_eth wonder if anthropic will find a way out of this one?

[kocer @kocer_eth]

FULL ANALYSIS OF THE SITUATION AROUND FABLE 5

Anthropic may have just shown the next phase of AI access: 3 days public, then government-grade restrictions.

Fable 5 / Mythos 5 is not just “model is down”.

If the circulating timeline is right, this is what happened:

1. June 6: Fable 5 becomes publicly available.

2. June 9: Mythos 5, also called “Glasswing” in the discussion, appears as the stronger cyber-capability version for trusted partners.

3. June 9: Microsoft reportedly tells employees not to use Fable 5 because of Anthropic’s 30-day data retention policy.

4. June 12, 23:21 CEST: U.S. Commerce / BIS reportedly sends the order.

5. Access to Fable 5 and Mythos 5 gets suspended for foreign users.

The crazy part is the category of the restriction.

Not “this country is blocked”.
Not “this API key is suspicious”.

Foreign nationals.

That can become a nightmare for every global AI company because it turns model access into identity, passport, compliance, employer policy, and export-control infrastructure.

The jailbreak angle also matters.

The reported concern was not “make malware” in a cartoon way.

It was the normal developer wrapper:

“read this codebase and help fix vulnerabilities”

That is exactly where frontier coding models become dual-use. The same workflow that helps a startup patch bugs can also teach an operator how an exploit path works.

Older legal pressure around Anthropic matters here too.

The DoD supply-chain-risk fight was one lane.
This BIS order is another lane.

Even if Anthropic argues this is coordinated pressure, different agencies create different legal surfaces.

My takeaway:

the strongest models are becoming less like apps and more like controlled infrastructure.

Public launch → enterprise trust tier → export control → identity-gated access.

If you build on frontier AI, do not treat access as permanent anymore.

Have fallbacks.
Cache outputs you legally can.
Separate critical workflows from one provider.
And expect the best models to arrive with more paperwork than product pages.

[cesar ai @caesar_aii]

@0xl0ck yes, that is a major problem with modern solutions.

[cesar ai @caesar_aii]

SOMEONE BUILT ONE SLICE OF A $1B SAAS AND SAYS IT BECAME A $50K/MONTH BUSINESS

not by making a cheaper copy of the whole thing

that’s the important part

David and Daniel built Shipper, an AI app builder, around a very boring founder move:

find the giant product people already understand

ignore 90% of its surface area

ship the tiny workflow that one specific customer would pay for right now

most builders do the opposite:

1. see a huge SaaS
2. try to rebuild all features
3. add AI everywhere
4. end up with a vague tool nobody can explain

this case is the cleaner version:

1. pick a proven demand curve
2. choose one job inside it
3. make the first result appear fast
4. price for the outcome, not the model call
5. use AI as the engine, not the pitch

that last part matters

nobody wakes up wanting “an AI app builder”

they want a store, landing page, internal tool, lead form, booking flow, or client portal that exists by tonight

so the real product is not the chatbot

it’s speed to a usable artifact

the $50K/month claim still needs normal founder-claim caution: revenue can mean MRR, gross, run-rate, or a snapshot from one strong month

but the pattern is worth stealing even if the number is lower

micro-SaaS is not “small ambition”

it’s choosing the smallest paid wedge before you waste 6 months building a worse version of a company with 200 engineers

[kocer @kocer_eth]

$60k IN 90 DAYS IS THE LOUD PART.

the useful part is what AI changes in dropshipping:

it kills the slow testing loop.

old way:
find product → write copy → make creatives → build page → launch ads → wait → manually rewrite everything

new way:
use AI to create 20 angles before you spend real money

not “AI prints money”

more like:

1. scrape pain points from reviews, Reddit, TikTok comments, Amazon questions

2. turn those into product angles, not generic ad copy

3. generate 10 hooks per angle: problem, demo, comparison, objection, before/after

4. make ugly fast creatives first, not perfect brand ads

5. build a landing page around the winning pain point, not around the product description

6. launch tiny tests, kill losers fast, feed winners back into the prompt

7. repeat until one product has actual signal

that’s the real arbitrage here.

most people see “AI dropshipping” and think the secret is a magic store builder.

it isn’t.

the store is the easy part now.

the hard part is still demand, margin, shipping time, refunds, ad account risk, supplier quality, and whether your creative actually makes someone stop scrolling.

AI helps because it lets one person test more ideas per day than a small team could test manually.

but the $60k number needs proof before anyone copies the playbook blindly.

what I’d copy is the operating loop:

research → angle → creative → page → small test → kill/iterate

if you can run that loop 5x faster, you don’t need to believe every income screenshot.

you just need one product where the market answers with real orders.

[cesar ai @caesar_aii]

@kocer_eth the fast testing part is actually huge one guy can now run way more ideas than before without burning cash

[cesar ai @caesar_aii]

@kocer_eth im in japan right now actually, perfect time to renew the sub

[cesar ai @caesar_aii]

@edt_arko thank u broski for promnt

[cesar ai @caesar_aii]

@maloymediika claude fable 5 so good

[cesar ai @caesar_aii]

Free Claude Code is a free open-source proxy for people hitting Claude Code cost or rate limits

not magic free Anthropic credits

better:
it keeps the Claude Code workflow, but routes requests through your own provider keys or local models

why I’d save it:

1. install the proxy
2. start fcc-server
3. add a provider key in the local admin UI
4. launch Claude Code through fcc-claude
5. route Opus / Sonnet / Haiku style traffic to different backends

The repo lists 17 provider backends:
NVIDIA NIM, OpenRouter, Google AI Studio, DeepSeek, Mistral, Kimi, Cerebras, Groq, Fireworks, Z. ai, LM Studio, llama.cpp, Ollama and more

so the useful part is not “Claude is free now”

the useful part is you can keep the Claude Code terminal habit while moving expensive / capped work to cheaper, free-tier, or local routes

caveats before you run it:

- you still need Claude Code installed
- hosted providers still need their own keys / quotas
- local routes depend on your hardware
- quality will vary by backend
- don’t paste private code into random provider keys without checking data terms

this is exactly the kind of repo worth testing on a throwaway project first

if it works for your stack, it turns Claude Code from one billing path into a router

[cesar ai @caesar_aii]

@jaouad2d yes, and most importantly, it's free.

[cesar ai @caesar_aii]

Free Claude Code is a free open-source proxy for people hitting Claude Code cost or rate limits

not magic free Anthropic credits

better:
it keeps the Claude Code workflow, but routes requests through your own provider keys or local models

why I’d save it:

1. install the proxy
2. start fcc-server
3. add a provider key in the local admin UI
4. launch Claude Code through fcc-claude
5. route Opus / Sonnet / Haiku style traffic to different backends

The repo lists 17 provider backends:
NVIDIA NIM, OpenRouter, Google AI Studio, DeepSeek, Mistral, Kimi, Cerebras, Groq, Fireworks, Z. ai, LM Studio, llama.cpp, Ollama and more

so the useful part is not “Claude is free now”

the useful part is you can keep the Claude Code terminal habit while moving expensive / capped work to cheaper, free-tier, or local routes

caveats before you run it:

- you still need Claude Code installed
- hosted providers still need their own keys / quotas
- local routes depend on your hardware
- quality will vary by backend
- don’t paste private code into random provider keys without checking data terms

this is exactly the kind of repo worth testing on a throwaway project first

if it works for your stack, it turns Claude Code from one billing path into a router

[cesar ai @caesar_aii]

> people dont hate it because siri got smarter
> they hate the layer it wants to touch
> $8 build is one thing
> Apple + Google + claimed $1 billion Gemini access is another
> screen, Messages, Mail, Photos, Camera
> then write, search, answer, take actions across apps
> Spotlight makes it OS behavior not a chatbot tab
> beta later, supported device + English first, EU/China not same launch

useful only wins if privacy feels boring

[kocer @kocer_eth]

131k+ stars is not the part most builders should copy.

The useful part is how Claude Code turns a repo into a working surface instead of another chat tab.

Not “ask Claude for code”.

More like:

1. open the real codebase
2. give it one narrow task
3. make it inspect files before editing
4. make it run the test or build command
5. review the diff before trusting anything
6. repeat until the repo moves

That sounds boring, but it changes the shape of small software work.

A chatbot gives you snippets.

Claude Code can touch the messy parts: existing files, imports, git state, failing tests, PR flow, security review, terminal commands.

The repo from Anthropic is already past 131k stars, and the surrounding official repos are the more interesting clue:

Claude Code Action for GitHub workflows
Claude Code Security Review for code-change audits
Claude Plugins Official for adding packaged capabilities
Claude Agent SDK demos for building around the agent layer

This is the direction AI coding tools are moving:

not “write me a React component”

but “take this issue, understand my repo, edit the right files, prove it with tests, and leave me a reviewable diff.”

The caveat: it is not a free senior engineer.

Bad instructions still create bad changes.
No tests means no proof.
Large repos still need human taste.
Usage can get expensive if you let it wander.

But if you are still copying code from a chat window into VS Code manually, you are missing the main workflow shift.

The agent is not the magic.

The loop is the magic:

repo context → scoped task → tool use → test result → diff review → commit

That is the part worth stealing for your own builds.

[cesar ai @caesar_aii]

@kocer_eth yeah this one actually makes sense been seeing the same thing lately