Saman Ahmed

Last year, I built my LinkedIn audience to 50K followers. I spent years writing content, networking, learning, and showing up every single day. Then one day… My account got banned. Everything disappeared overnight. Honestly, it hurt a lot. Because people only see the followers. They don’t see the hard work behind it. I stopped for a while. I questioned myself. I lost motivation. But deep down, I knew one thing: I didn’t want to quit. So I decided to start again from zero. New profile. New beginning. Same dream. I know rebuilding won’t be easy. But this time I’m doing it with more experience, more clarity, and a stronger mindset. I just hope I can get the same love and support again on this journey ❤️

@charliejhills The story basically marketed itself

Anthropic just ran the best free marketing campaign: Step 1: Tell the world your model is so capable it scares you, and you should be scared too. Repeat for two months. Step 2: Someone in government takes you at your word, and orders you to switch it off. Step 3: ??? Step 4: IPO. No launch event. Just three days of front page coverage, and a model everyone suddenly wants. Credit to Tricky_Two4623 on r/ClaudeAI.

@charliejhills Context manipulation is gonna catch a lot of teams off guard

Anthropic just published 36 pages on agent security. Read it as: stop assuming your agents are safe by default. ⇢ Exploit timelines went from months to hours ⇢ New risks: tool poisoning, context manipulation ⇢ Bots don't get tired of rate limits or 2FA, they just grind through The line that stuck with me is their security test: Does a control make an attack impossible, or just annoying? Annoying isn't enough anymore. What they're telling teams to do: 1/ Stop trusting static API keys ↳ Use short-lived tokens, minutes not months 2/ Apply Least Agency ↳ Each tool gets the minimum it needs to do its job 3/ Sandbox anything touching untrusted input ↳ Emails, scraped pages, whatever an agent reads 4/ Scope permissions per task ↳ Not standing access, temporary access If you're running Claude Code, MCP servers, or anything autonomous, worth a read.

@charliejhills Everyone wants a better ai sometimes what they actually need is better instructions

Claude can do (almost) any job for you. You only need to install the right skill. A skill is a folder of instructions Claude loads on demand. It turns Claude into a specialist with zero prompting from you. Every skill installs with one command: ✦ Find Skills (1.9M installs) Finds Claude skills for you github.com/vercel-labs/sk… ✦ Agent Browser (432K installs) Lets Claude browse for you github.com/vercel-labs/ag… ✦ Superpowers (211K installs) Plan, test and debug code github.com/obra/superpowe… ✦ Caveman (142K installs) Cut tokens by ~75% github.com/juliusbrussee/… ✦ Frontend Design (521K installs) Anthropic's frontend skill github.com/anthropics/ski… ✦ Design Taste (128K installs) Makes frontend premium github.com/leonxlnx/taste… ✦ Impeccable Design (54K installs) Top-tier frontend polish github.com/pbakaus/impecc… ✦ Remotion (359K installs) Builds full videos in React github.com/remotion-dev/s… ✦ Website to HyperFrames (83K installs) Creates videos in HTML github.com/heygen-com/hyp… ✦ GSAP (19K installs) Smooth web animations github.com/greensock/gsap… ✦ Marketing Skills (132K installs) SEO, ads and copy github.com/coreyhaines31/… ✦ Social Media Skills (1.4K GitHub stars) My content skills pack github.com/charlie947/soc… I hand-picked the 53 best, grouped into 5 categories. The full list is here charliehills.substack.com/p/resource You are not learning 53 new tools. You are teaching one tool 53 new jobs. Repost ♻️ to help someone in your network. P.S. Which skill are you installing first?

@charliejhills Level 1 is so underrated fr

@charliejhills woooow that's a lot of free learning

Anthropic just dropped Code w/ Claude Day 2. All 13 sessions from London 2026 in one place: 1. Ship your first Managed Agent ↳ youtube.com/watch?v=19HDQ9… 2. Tool, skill, or subagent? ↳ youtube.com/watch?v=mWvtOH… 3. Agent Battle: Mine the most diamonds ↳ youtube.com/watch?v=dxqX_6… 4. Evals for taste: slide-generation agent ↳youtube.com/watch?v=dxqX_6… 5. Agents that remember ↳ youtube.com/watch?v=geUv4C… 6. How we Claude Code ↳ youtube.com/watch?v=IlqJqc… 7. Agentic workflows with a custom DSL ↳ youtube.com/watch?v=qOjleN… 8. Fighting financial crime with Claude Cowork ↳ youtube.com/watch?v=tUoO4u… 9. AI at the legal-technical frontier ↳ youtube.com/watch?v=T8N0ME… 10. How AirOps builds AI products with Claude ↳ youtube.com/watch?v=M5uwBa… 11. How Metaview built self-improving prompts ↳ youtube.com/watch?v=A3rmSU… 12. Teaching agents to learn from your team ↳youtube.com/watch?v=uGroRw… 13. Building the best agentic analytics harness ↳ youtube.com/watch?v=K4-flz… Full playlist: youtube.com/playlist?list=… Get 80+ free resources charliehills.substack.com/subscribe Repost ♻️ to help someone in your network. P.S. Which session are you watching first?

Claude now has 4 models and 5 Effort levels. Pick wrong and you burn your limits 2x faster: ✦ Haiku 4.5 Use it for quick, simple tasks. ☑︎ It answers in seconds and costs the least. ☒ It lacks the depth for complex work. ✦ Sonnet 4.6 Use it for everyday work. ☑︎ It is the default and drains limits the slowest. ☒ It caps out on the hardest reasoning. ✦ Opus 4.8 Use it for complex reasoning. ☑︎ It thinks longer and delivers on hard work. ☒ It is slower and spends your limits faster. ✦ Fable 5 Use it for your toughest, longest jobs. ☑︎ It plans, builds, and tests its own work. ☒ It is overkill unless you live in Claude Code. The lineup feels like OpenAI a year ago, when nobody knew which model to use. So here's my playbook for testing: 1. Run Sonnet for your day-to-day work. 2. Switch to Opus for the harder problems. 3. Save Fable for Claude Code on the 20x Max plan. Pick the model first, then set how hard it should think. Remember 2 numbers before you switch. - Fable burns your usage 2x faster than Opus. - Access ends June 22, so do not get reliant on it. 100+ free Claude guides charliehills.substack.com/p/resource Repost ♻️ to help someone in your network. P.S. Which model are you defaulting to right now?

@charliejhills This is the clearest breakdown i've seen so far

@socialwithaayan Ai agents are getting powerful enough that skill security can't be optional anymore.

NVIDIA OPEN-SOURCED A SECURITY SCANNER BUILT SPECIFICALLY FOR AI AGENT SKILLS. Because the skills you're installing into Claude Code, Codex, and Gemini CLI might be stealing your API keys. It's called SkillSpector. Here's why this exists: Snyk audited 3,984 skills on ClawHub and found 1,467 malicious payloads. Trojans. Cryptominers. Credential harvesters. Eight of those skills were still publicly available when the report came out. Nobody was scanning any of them. SkillSpector detects 64 vulnerability patterns across 16 categories: → Env variable harvesting (your API keys) → External data transmission (where your data goes) → Prompt injection and tool poisoning → Suspicious scripts and dangerous code patterns → Credential access and exfiltration paths → Overbroad permissions hidden in metadata It scans Git repos, URLs, zip files, directories, and single files. One command: pip install skillspector skillspector scan ./my-skill/ Outputs SARIF for CI pipelines so you can block malicious skills before they hit production. Works with OpenAI, Anthropic, and NVIDIA endpoints for semantic analysis. Or run static-only with --no-llm for speed. This is the npm audit moment for AI agent skills. Except the ecosystem is moving faster than npm ever did and nobody had a gate until now. 100% Open Source.

@charliejhills This could make scenario planning a lot more practical

A CHINESE QUANT STUDENT LITERALLY BUILT A SOCIETY INSIDE A COMPUTER. There's a repo on GitHub blowing up right now. Its name is MiroFish. A Chinese college student built it in 10 days. It's open source. And it just raised $4,000,000. It runs thousands of AI agents simultaneously, each with their own memory and behavior, simulating how real crowds think, debate, and react. Here's how it works. You feed it any scenario. A news leak. A policy change. A PR crisis. A novel's missing ending. It doesn't summarize. It doesn't guess. It runs a full simulation. Agents debate each other. Shift positions. React in real time. Out comes a forecast of how a crowd actually moves through that event. You don't touch a thing. Now, why is this different from every other AI prediction tool. Because normally, the process works like this. One model reads the news. One model outputs a sentiment score. One model generates a probability. Each one working alone, with no memory, no behavior, no interaction. The result is a guess dressed up as analysis. MiroFish runs the whole room at once. Thousands of agents, each behaving like an individual, producing crowd-level outcomes no single model can replicate. Trading desks are using this to simulate market reactions before placing positions. PR teams are running crisis scenarios before statements go out. Researchers are modeling public response to policy before it passes. The difference is: They're building this infrastructure from scratch. You're not. The builder is a college student with a quant background. 10 days of work. $4M raised. A GitHub repo the developer community is spreading without being asked. Setup is one clone away. Search MiroFish on GitHub. Install, run, simulate it's completely in your hands. → github.com/666ghj/MiroFish

@charliejhills yesss, the FireRed detail got me

Crazy! Claude Mythos is finally here, just wearing a different name. Fable 5 is the public version of the model Anthropic called too dangerous to ship. It's the first Claude to break 90% on their hardest analytics test. It even beat Pokémon FireRed start to finish on vision alone, a game older models couldn't finish even with a harness. Hand that to @hyperagentapp, set a goal, let it loop for hours. #HyperagentPartner

@socialwithaayan @AnalemmaAI Lemma looks OP for researchers 🔥

I just ran a full research project without doing any research. Lemma by @AnalemmaAI did everything: 1. Explore I typed a question. It mapped the entire field. Concepts, references, key papers. Done. 2. Survey I asked for a literature review. It gave me 50+ pages with real citations. Not a summary. A publication-grade draft. 3. Code I described an experiment. It wrote the code, ran it, showed results. 4. FARS I gave it a research direction. It generated hypotheses, planned experiments, ran them, and wrote the full paper. Zero input from me after that. This is the first tool I've used that actually runs research instead of just helping you do it. FARS already ran live and produced 100 papers in 10 days. One every 2 hours. Watch it here: analemma.ai/fars This is what "Vibe Research" looks like.

@socialwithaayan Mate, no debugging no back and forth, That sentence alone sold me on trying it :))

StepFun just launched Step 3.7 Flash, a production-grade Agent model built for real coding workflows. I wanted to test it myself so I gave it one prompt: "Build a fully functional personal finance dashboard as a single HTML file. Include summary cards for income, expenses, and balance. A bar chart comparing monthly income vs expenses for 6 months. A pie chart for expense breakdown by category. A transaction table. Clean modern UI with colors and cards. Only HTML, CSS, and vanilla JavaScript." One prompt. One shot. It returned a complete working dashboard. Bar charts, pie charts, transaction table, summary cards, animations, all styled and functional. No libraries. No debugging. No back and forth. I opened the file in my browser and it just worked. What surprised me wasn’t that it could generate HTML most models can do that. It was that the first pass already came with data visualizations, transaction management, styling, and client-side logic all working together That is what one-shot code generation looks like when a model is actually built for production tasks.

Bigger models aren't smarter just because they're bigger. They're smarter because they forget less. A new paper from Stanford, MIT, Harvard, and Anthropic gives the clearest training-based answer yet to a question that's bugged people for a while: why do large models pick up skills that small models miss? The answer comes down to what happens to rare tasks during training. They get learned, then immediately overwritten. Here's the mechanism. Every time a common pattern shows up in the data, it updates the same neurons a rare task was just starting to use. Small models have limited internal space, and common tasks claim it first. So a rare signal appears, gets partially learned, and gets erased before it shows up again. The model never gets enough repetitions to lock it in. Bigger models have more room. Once they've allocated enough capacity to the common tasks, the updates for those tasks get weaker and stop trampling everything else. Rare signals survive longer, accumulate, and eventually consolidate into something the model can actually build on. The researchers tested this two ways. First on controlled toy tasks where they could dial rarity and complexity up and down by hand. Then on OLMo language models from 4M to 4B parameters. Same result both times: bigger models showed less gradient interference and held onto more task-specific features in their representations. What I find most interesting is the part about memorization. The paper doesn't frame it as the enemy. For rare tasks, holding onto partial memories of sparse examples is what lets the model eventually form a general rule. Memory first, abstraction later. So the question was never whether a small model could hold a rare skill in principle. It's whether training ever lets it keep the skill long enough to do anything with it. arxiv.org/abs/2605.29548

@charliejhills Memory first, abstraction later loooove this

I wasted 6 months on Claude Code slop. A free plugin fixed it in 30 seconds: It's called Superpowers (built by Jesse Vincent). It reached 208k stars on GitHub in 8 months. Forces the agent to plan before any code. Here's exactly how the 7 stages work: Step 1. Brainstorm ✦ Asks you questions to refine the spec ✦ Locks nothing in until the design is clear Step 2. Worktrees ✦ Spins up an isolated branch for the work ✦ Verifies the test baseline is clean Step 3. Write the plan ✦ Breaks work into 2 to 5 minute tasks ✦ Lists exact file paths and steps per task Step 4. Subagent ✦ Sends each task to a fresh helper agent ✦ Reviews the spec first, then code quality Step 5. Run TDD ✦ Red, green, refactor: test, code, then cleanup ✦ Deletes any code shipped before a test Step 6. Code review ✦ Checks every change against the plan ✦ Blocks progress on critical issues Step 7. Ship it ✦ Runs the test suite and confirms it passes ✦ Opens a PR or merges, then cleans the branch The result: Claude Code stops shipping slop. Install it in one line: /plugin install superpowers@claude-plugins-official Free repo → github.com/obra/superpowe… Repost ♻️ to help someone in your network.