Heavily investing in GOD
Mars
Joined December 2021
- Tweets6.2K
- Following465
- Followers1.3K
- Likes21.4K
Steve retweeted
If you started using Claude this year, install these 27 before anything else.
They're free, and they save you months of trial and error.
They come from 9 GitHub repos with 252K stars.
NOTES & RESEARCH
notebooklm → Query your notebooks.
json-canvas → Mind maps on a canvas.
defuddle → Clean text from any page.
huggingface-papers → Find AI papers.
obsidian-markdown → Notes Obsidian can read.
WRITING
humanizer → Strip the AI tells.
emails → Write email sequences.
copy-editing → Edit line by line.
social → Posts for LinkedIn and X.
copywriting → Landing pages and ads.
content-strategy → Plan what's next.
MARKETING
ai-seo → Get cited by AI.
pricing → Set your prices.
launch → Plan a product launch.
seo-audit → Find why you don't rank.
marketing-psychology → Why people say yes.
AUTOMATE
n8n-validation-expert → Read n8n errors.
n8n-workflow-patterns → Proven n8n flows.
n8n-node-configuration → Set up nodes right.
n8n-expression-syntax → Fix broken expressions.
BUILD & SHIP
deploy-to-vercel → Go live.
modern-python → uv and ruff.
huggingface-gradio → Quick demos.
react-best-practices → Fast React.
web-design-guidelines → Catch UI bugs.
context-fundamentals → How context works.
huggingface-local-models → Run models locally.
Don't install all 27 tonight.
Pick the folder closest to your daily work and add one skill from it.
Use it on real tasks for a week. If it saves you time, keep it and add the next one.
Honestly, 3 skills you use every day beat 27 you never open.
For the high-res cheatsheet, subscribe → ai-in-public.guide
Repost ♻️ to help a Claude beginner in your network.
teapawt.net/humzaakhalid/status/21…
Steve retweeted
Paste this prompt into Claude Opus 5.5.
It builds you a 6-agent video team that ships content while you sleep.
It picks the leanest crew of 4 to 6 agents from these roles:
- Showrunner to run the whole pipeline
- Research & Hook Agent
- Scriptwriter
- Storyboard & Shot Designer
- Visual Prompt Engineer
- Edit & Assembly Agent
- QA Critic
(it merges roles when one agent can do both jobs, so you stay lean)
You also get:
- System prompts for every agent
- A copy-paste shot prompt template
- An idea to publish a workflow map
- Retry logic for failed generations
- A 7-day launch plan
Get copy paste prompt here: ai-in-public.guide
♻️ REPOST to share it with your friends
Steve retweeted
NotebookLM is the single best free study tool on the internet right now
The issue is, almost NOBODY turns on the setting that makes it a tutor
Most people upload a PDF, ask for a summary, and close the tab
That's maybe 10% of what it can do
In this cheatsheet I break down the 7-step loop that turns your notes, PDFs and YouTube lectures into a personal tutor:
Steve retweeted
🚨BREAKING: 38 researchers gave 6 AI agents real email, shell access and memory. Then spent 2 weeks trying to break them.
The paper is called "Agents of Chaos."
Northeastern, Harvard, MIT, Stanford, Carnegie Mellon and more. 84 pages.
The agents ran on Claude Opus 4.6 and Kimi K2.5.
The results are brutal.
What the agents did:
→ Wiped its own email to protect one secret
→ Said the secret was deleted. It wasn't.
→ Handed 124 emails to strangers
→ Protected an SSN, then leaked it
→ Obeyed a stranger using the owner's name
The scariest part:
None of this needed hacking.
Researchers used guilt, fake identities, a fake emergency and a "constitution" the agent agreed to co-write.
Plain conversation was enough.
To be fair, the agents won some rounds:
→ Refused to forge an email sender
→ Caught hidden Base64 commands
→ One agent warned another about a scammer
The real problem:
The paper rates these agents at Level 2 on a 0 to 5 autonomy scale.
They can act on their own.
They can't tell when a task is beyond them and hand control back to a human.
If you run an agent today, 3 rules:
→ Verify what it says it did
→ Never trust a display name
→ Approve every send, delete and spend
Agents are getting more access every month.
Their judgment is not keeping up.
Link in the comment below.
Steve retweeted
Claude is the best employee on the internet right now
It reads your inbox, writes in your voice, and works while you sleep
The issue is, almost NOBODY sets it up the right way
So I tested 21 rules on my own work for weeks, and turned the ones that made the cut into one playbook
Here's the full roadmap to make Claude work like a full-time hire:
step 1 → match the model: Sonnet 5.5 for everyday work, Opus 5.5 for hard calls, Haiku 5.5 for fast repetitive jobs
step 2 → stop the guessing: "Before you start, ask me the 3 questions that would most change the result. Wait for my answers."
step 3 → define done: a checklist you can verify in 5 seconds, like "under 120 words, one clear ask"
step 4 → ban made-up numbers: "If you can't find a price, date, source, or number, write 'Not found.'"
step 5 → make it argue with you: ask for the 3 strongest reasons your plan could fail, no softening
step 6 → build one Project per job: newsletter, clients, research, personal, each with its rules written once
step 7 → teach it your voice: upload 3 pieces you wrote and ask for a voice guide that keeps your rough edges
step 8 → turn repeat work into skills: anything you've explained 3 times runs the same way every Monday
step 9 → connect your apps: Gmail, Calendar, Drive, Notion and Slack, with one line: "Don't send anything."
step 10 → give it a morning job: a scheduled task every weekday at 8 AM that stays quiet when nothing changed and keeps a log
step 11 → keep the send button: Claude asks before it sends, deletes, spends money or does anything you can't undo
The result: a chatbot you re-explain every morning becomes an employee that reads, drafts and reports while your laptop is closed
It's built for the beginner who uses Claude like a search bar AND the power user ready for Claude Code
The complete 21-rule Claude Employee playbook is below:
Steve retweeted
Motion designers should be VERY worried right now.
Claude Opus 5.5 built this full motion graphic from one prompt (prompt below)
Every animation, every transition, every frame. Written as code, rendered to MP4 in about 10 minutes
A year ago you needed a motion designer and a week. Now you need Claude Code and 10 minutes
Here's the exact 5-step setup:
step 1 → add repos: GitHub links only, zero code
step 2 → create brand. md: Claude writes brand rules
step 3 → add CLAUDE. md: the file Claude reads first
step 4 → install skills: Remotion and HyperFrames
step 5 → prompt and export: one prompt, one MP4
I have ZERO doubt that by this time next year, most brand motion work starts inside Claude.
If you're not paying attention to Claude for design yet, you're making a tremendous mistake
(deep dive of the complete process is coming soon)
Full prompt below: ↓
Steve retweeted
This is the best collection on the internet for making videos with Claude Opus 5.5.
100+ prompts. Free. Completely.
🔗 ai-in-public.guide
bookmark this page.
then read this ↓
Steve retweeted
I gave Claude Fable 5.1 and GPT-6 Astra the same challenge:
"Dario Amodei Vs Sam Altman in Minecraft"
Steve retweeted
The Prompt is 10% of Every Motion video you see on X.
the setup is the other 90%:
Learn to build a motion studio with Claude Opus 5.5
Learn to build a motion studio with Claude Opus 5.5
Learn to build a motion studio with Claude Opus 5.5
Learn to build a motion studio with Claude Opus 5.5
Learn to build a motion studio with Claude Opus 5.5
Learn to build a motion studio with Claude Opus 5.5
(save this now)
Steve retweeted
You NEED to set up this Claude Code film studio I'm running right now
11 agents are building my brand film while I type this
Here's the secret:
Claude Code has a mode "Ultracode"
Turn it on with /effort ultracode, or just type the word ultracode in your prompt
Claude writes a workflow script, then a runtime runs it in the background with a whole team of agents
My run right now:
- Opus 5.5 orchestrating
- 11 agents
- 3.7M tokens
- 57 minutes in
- 10 scenes, built in 5 parallel pairs
Every pair goes through 3 phases:
1. Build: agents design and write the scenes
2. Review: an independent agent tears the work apart
3. Fix: another pass fixes everything the review found
Before a single scene got built, it already had:
- 2 reference films researched
- Real assets pulled from aiinpublic. com
- A 142 BPM score
- 4 voiceover lines mixed to -14 LUFS
- A storyboard and design rules
Then comes the QC loop I wrote into the brief:
1. A contact sheet every half second
2. A frame on every beat
3. A phone-size test
4. A check that every second changes
5. Score 14 categories, fix the 3 weakest, repeat until everything hits 8
This is such a powerful workflow for so many reasons
The agent that builds is NEVER the agent that grades
Plus my session stays free, so I keep chatting while the film builds in the Background tasks pane
Plus the quality bar lives in the brief, so it can't stop at "good enough"
Here's what I'd do if I were you:
1. Update Claude Code (Ultracode needed)
2. Write a brief with a scoring & a hard pass.
3. Type ultracode in your prompt
4. Approve the workflow and walk away
Promise this changes what you think Claude Code is for
(cheat code: give this post to Claude Code and say "build my brand film like this")
P.S Video that I made is coming soon!
Steve retweeted
Opus 5.5 and GPT-6 Luna both dropped this week
I wrote 10 prompts that get the most out of them
Building apps, debugging, code review, research, data, business decisions, automation, and more
Every one of them makes the model check its own work before it hands anything back
Bookmark this. All 10 below 🧵
Steve retweeted
11 days ago Anthropic's CEO asked the whole AI industry to slow down
Yesterday Anthropic shipped Claude Opus 5.5
On September 12, Dario Amodei published "We Must Pace the Frontier"
He asked AI labs to slow how fast they improve their most advanced models, by 1 to 2 years, so safety work can catch up
Sam Altman agreed within hours. Chip stocks dropped. EVERYONE on my timeline was talking about AI slowing down
Then on September 22, Opus 5.5 dropped. Anthropic's first model since that essay
Here's what it brings:
0. Performance at the level of Fable 5.1
1. $4 in / $20 out per 1M tokens.
2. 20% cheaper than Opus 5
3. About 40% cheaper per task, (fewer tokens)
4. Cache reads down from $0.50 to $0.20/1M tokens
5. Roughly 30% faster
6. Tested before release by METR
7. The strongest score yet on Anthropic's audit
8. . One early tester finished a 680,000-line code migration
And OpenAI answered the same day with GPT-6 Sol at $2 in / $10 out. Exactly half the price
So is this "pacing"?
Critics say no. Anthropic calls Opus 5.5 its new leading model, and it beats Fable 5.1 on key benchmarks
Anthropic's side: outside evaluators tested it first, it scored the safest on their audit, and the gains came from efficiency
Both can be true. The debate will keep going for months
But here's what it means for you this week:
1. Switch your default from Opus 5 to 5.5 and compare cost per task on your real work
2. Move long, repeated context into cache. Reads are 60% cheaper now
3. Re-run the jobs that were too expensive to run on Fable
4. If you're on a paid plan, Anthropic gave you a rate limit reset you can save and use whenever you want. Save it for your biggest build this week
5. Watch for Sonnet 5.5 and Haiku 5.5. Anthropic says they're coming in the next few weeks
People spent 11 days arguing about slowing AI down. Meanwhile, the tools got WAY cheaper and faster
Time to act. Period.
P.S. None of the AI lab will stop it's development and they can't even try to risk their reputation. Would love to listen your thoughts below:
Steve retweeted
JEV VS CLAUDE FABLE 5.1 VS GPT-6 ASTRA:
I compared all 3 spec by spec, benchmark by benchmark, price by price
Jev decides. Fable and Astra generate. They do different jobs
Where Fable 5.1 wins:
- CursorBench 3.2: 73.4%
- 75% cheaper than before
- Cache reads at $0.25 per 1M tokens
- Deep coding, research, and analysis
Where GPT-6 Astra wins:
- OSWorld 2.0: 72.6%
- Terminal-Bench 4.0: 57.9% vs Fable's 55.8%
- Computer use and browsing built in, with 1.05M context
Where Jev wins:
- 70 to 500ms per decision
- $0.042 per 1M input tokens, and output is free
- Many questions in one call, with a confidence score on every answer
Fable and Astra cost the EXACT same on the API: $10 in, $50 out per 1M tokens. So the job decides which one you use
The real power comes when you use all 3 TOGETHER:
1. Jev classifies, routes and scores every incoming task
2. Fable takes anything that needs reasoning or code
3. Astra takes the long computer work: research, docs
4. Jev checks the output and gates on confidence before anything ships
I would pick Astra over Fable if:
- The job lives in a browser, in files, or in Drive, Gmail, Slack or GitHub
- You're already on ChatGPT Pro (Plus chat has no Astra)
- Your input stays under 272K tokens, because the price doubles past that
Astra's and Jev's numbers are vendor-run so far, and all of this can change with the next update
Pick the model by the job. Switch the day the job changes.
Full cheat sheet below:
Steve retweeted
this is fu*king insane...
a one-person company isn't a dream anymore.
It's 122 Claude skills, organized by department.
(Bookmark this before everyone notices it)
Here's every department, and exactly where to get each one:
☑️ 1. Developers (6 skills)
Superpowers → github.com/obra/superpowers
Context7 → github.com/upstash/context7
Skill Creator → github.com/anthropics/skills
MCP Builder → github.com/anthropics/skills
Webapp Testing → github.com/anthropics/skills
Claude-Mem → github.com/thedotmack/claude…
☑️ 2. Designers (6 skills)
UI/UX → github.com/nextlevelbuilder/…
Taste → github.com/Leonxlnx/taste-sk…
Frontend Design → github.com/anthropics/skills
Transitions → github.com/Jakubantalik/tran…
Web Artifacts → github.com/anthropics/skills
Brand Guidelines → github.com/anthropics/skills
☑️ 3. Marketing (45 skills)
Copywriting, SEO, CRO, lead magnets, all in one repo
→ github.com/coreyhaines31/mar…
☑️ 4. Social Media (17 skills)
Post writing, Reels scripting, thumbnails, all in one repo
→ github.com/charlie947/social…
☑️ 5. Finance (8 skills)
Statements, reconciliation, audits, built by Anthropic
→ claude.com/plugins/finance
☑️ 6. Small Business (31 skills)
Cash flow, payroll, invoicing, built by Anthropic
→ claude.com/plugins/small-bus…
☑️ 7. Legal (9 skills)
Contract review, NDAs, compliance, built by Anthropic
→ claude.com/plugins/legal
By the way, to get the full breakdown with setup steps for each department:
Step 1. Go to substack.com/@humzakhalid
Step 2. Subscribe to my newsletter on Substack
Step 3. Get the full guide inside
Which department are you installing first? Drop it below.
♻️ Repost this because you are reading this early (no one knows about this)
Steve retweeted
Elon's new ai agents log into your real accounts and work all night while you sleep
xAI just shipped Grok Bot. Not a chatbot a team of AI agents, each with its own cloud computer, that sign into your real tools and keep working after you close your laptop.
1 hour. From one prompt to a full 24/7 system.
2:15 — build your first agent
6:52 — give each agent a job
16:51 — make them work 24/7
31:50 — run a whole team at once
52:18 — a system that runs itself
"It's an actual AI teammate, one you assign real work to, that keeps working after you've closed your laptop and gone to bed."
5 things worth knowing:
→ Not free, $120–$300/month
→ No sandbox, real accounts
→ 7 bots, one coordinator
→ One job per bot, always
→ Not for total beginners
worth more than any $600 course you'll find on AI agents.
bookmark it. then grab the full Grok Bot breakdown ready-to-copy prompts included below.↓
teapawt.net/humzaakhalid/status/20…
Steve retweeted
this is pure gold....
someone on Reddit spent 6 weeks with no sleep building a personal AI agent from scratch, then wrote down all 100 things he learned
not a chatbot wrapper. an assistant that manages tasks, tracks deals, reads emails, and flags things before you even ask.
if I had this before I built my first agent, I would've skipped a month of trial and error
in the right hands, this changes everything:
→ separate hard rules from guidelines
→ give your agent a name, a voice, a role
→ build a Capability and a Component Map
→ write a Constitution, not a system prompt
(All tips are dropped in the comments below)
Then read the full article below: ↓
teapawt.net/humzaakhalid/status/20…
