AI Learnings - January 19, 2026
Claude Code Power User Techniques
The Claude Code community is pushing the boundaries of what's possible with AI-assisted development, sharing increasingly sophisticated workflow optimizations.
Smart Forking: Never Lose Context Again
Zac (@PerceptualPeak) shared a breakthrough technique called "Smart Forking" that leverages your entire Claude Code history:
"Why not utilize the knowledge gained from your hundreds/thousands of other Claude code sessions? Don't let that valuable context go to waste!!"
The system works by embedding your prompt, cross-referencing it against a vectorized RAG database of all previous chat sessions, and returning the top 5 most relevant historical sessions with relevance scores. You can then fork from any of these sessions, starting new work with all that accumulated context intact.
Infinite Sessions: Solving Context Collapse
Evan Boyle (@_Evan_Boyle), who Scott Hanselman notes is "leading the charge on the GitHub Copilot CLI," revealed work on "infinite sessions" to address a fundamental problem:
"When you're in a long session, repeated compactions result in non-sense. People work around this in lots of ways. Usually temporary markdown files in the repo that the LLM can update - the downside being that in team settings you have to juggle these artifacts."
The solution promises "one context window that you never have to worry about clearing, and an agent that can track the endless thread of decisions."
Cross-Tool Skill Sharing
Jeffrey Emanuel (@doodlestein) offered a practical one-liner for syncing Claude Code skills to Codex:
``bash
mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills" && rsync -a "$HOME/.claude/skills/" "${CODEX_HOME:-$HOME/.codex}/skills/"
``This kind of interoperability between AI coding tools suggests we're moving toward a more unified ecosystem.
3D Game Development Revolution
Min Choi (@minchoi) announced that "3D game dev is about to change forever" with MCPs (Model Context Protocols) that let Claude communicate directly with Unity, Unreal, and Blender:
"Claude can now talk directly to Unity / Unreal / Blender... so you can build crazy 3D scenes + game with just prompts."
This represents a significant expansion of what "vibe coding" can accomplish - moving from traditional software development into creative 3D workflows.
The Multi-Agent Abstraction Layer
Guillermo Rauch (@rauchg), CEO of Vercel, highlighted an API that manages multiple coding agents:
"An API that abstracts over and manages every major coding agent for you. If you're looking to build coding AI into your products (think: auto-fixing, code review, testing, …), I'd start here first."
This meta-layer approach could simplify building AI-powered developer tools by handling the complexity of different agent implementations.
The Philosophy of Local LLMs
Ahmad (@TheAhmadOsman) made a compelling case for running local models, framing it as "cognitive security." A commenter (@0xCanaryCracker) pulled out this striking quote:
"If you can steer a model, you can recognize when one is steering you. If you can't, you're just another uncalibrated endpoint in someone else's reinforcement loop."
This perspective reframes the local vs. cloud LLM debate from a privacy concern to one of cognitive autonomy and understanding.
The Meta Commentary
Of course, no AI digest would be complete without acknowledging the recursive nature of all this optimization. As near (@nearcyan) observed:
"men will go on a claude code weekend bender and have nothing to show for it but a 'more optimized claude setup'"
A fair critique - though arguably these optimizations compound over time into genuine productivity gains. The trick is knowing when to stop configuring and start building.
Key Takeaways
1. Context management is the frontier - Both Smart Forking and Infinite Sessions address the same core problem: preserving valuable context across AI interactions
2. Tool boundaries are dissolving - Skills sync between Claude and Codex; MCPs connect Claude to 3D software; APIs abstract across all agents
3. The local LLM argument is evolving - Beyond privacy, it's now about understanding and controlling the systems that increasingly mediate our thinking
4. Self-awareness is healthy - The community can laugh at its own tendencies toward infinite optimization loops
Source Posts
M
I'm always saddened when people ask me where I go to church, because I know I can't tell them.
The church and congregation would be targeted by the Left, like this and it wouldn't be fair to them.
If you are barging into churches screaming at people, you're not the good guy. https://t.co/8eKVnYmpjK
D
The ease with which this works is amazing.
I gave it 10mb of logs and asked it to figure out the most common failure modes. Just worked.
i
isaac 🧩
@isaacbmiller1
The dspy.RLM module is now released 👀 Install DSPy 3.1.2 to try it. Usage is plug-and-play with your existing Signatures. A little example of it helping @lateinteraction and I figure out some scattered backlogs: https://t.co/Avgx04sNJP
C
NEW: Don Lemon loses it after a much smarter man shuts down his gaslighting attempts at the Minnesota church that was stormed by far-left agitators.
Lemon was seen harassing the man and following him around as he tried to leave.
"They've emptied a house of worship. Everybody has gone home. Their point has been proven worthless, and so in the end, I think they lose," the man said.
"But more importantly, if I were to break into any of their houses uninvited, cause derision and upset according to their values, I would be kicked out."
A
Don Lemon is now claiming that the MAGA administration is making stuff up, and that he didn’t storm the church.
On his livestream Don Lemon stated “people need to feel terrified. This is what protesting is all about”— after seeing children running out of the church scared.
Lock him up.
R
I’ve figured out a new way of working that’s unlocked my speed of iteration massively.
Here’s how it works:
I have a simple cron job that runs every night at midnight.
It gathers information from my database on user activity, marketing stats, and a couple other data points that are important.
It then feeds that data into Opus 4.5 and asks for one important action item that I should take based on this data, and then emails me.
It also creates a markdown file with the recommendation, which is then stored in my reports folder in the GitHub repo. (This means I can fire up Amp anytime and chat either all of the historical recommendations whenever I want - learning about patterns.)
I then look at this email every morning and decide whether or not to take action on it.
Almost every time it surfaces something really valuable for me to iterate.
So I just open Amp, tell it to action idea, and then ship it.
Obviously, the next iteration of this is just to have Amp autonomously implement the suggestion by itself, and then I'll wake up to a PR instead of an email.
Right now, though, I like the Human-In-The-Loop version of this.
And as soon as we iterate enough like that, I'll probably just set it up to automatically take the suggestion, create the PR, and then I'll have a look at it.
Obviously, you can take this loop even further by having many parts of your business evaluated this way.
What's interesting to me is that this is what I used to rely on my VP of Marketing, my VP of Engineering, or my VP of Sales to do, but it happens automatically for about $0.15 per day.
A
this is huge.
ollama is now compatible with the anthropic messages API. which means you can use claude code with open-source models.
think about that for a second. the entire claude harness:
- the agentic loops
- the tool use
- the coding workflows
all powered by private LLMs running on your own machine.
R
Claude researched sprite-making best practices. Gemini 3 Pro generated the actual pixel art. 13 cents per character. I can't draw at all but now I have full sprite sheets for my multiplayer soccer game. https://t.co/mavbez0INl
W
Anthropic is developing a new feature for Claude Cowork called Knowledge Bases (KBs) persistent, topic-specific memory containers that Claude will automatically reference and update.
These KBs are designed to store user preferences, decisions, lessons, and facts essentially acting as long-term, self-maintaining knowledge repositories that enhance Claude’s ability to reason over context-rich workflows.
Internal instructions hint that Claude will proactively use and grow these KBs during conversations.
T
TestingCatalog News 🗞
@testingcatalog
BREAKING 🚨: Anthropic is working on "Knowledge Bases" for Claude Cowork. KBs seem to be a new concept of topic-specific memories, which Claude will automatically manage! And a bunch of other new things. Internal Instruction 👀 "These are persistent knowledge repositories. Proactively check them for relevant context when answering questions. When you learn new information about a KB's topic (preferences, decisions, facts, lessons learned), add it to the appropriate KB incrementally."
L
@DavidOndrej1 Summary; Most people will miss the AI revolution not because they lack intelligence or time, but because they choose the wrong problems and business models instead of building truly valuable AI-driven products.
M
AgentWealth: a personal finance command center where AI agents manage your money like you have a family office.
Watch them negotiate your bills, pause unused subscriptions, transfer credit card balances to 0% APR, find the cheapest gas nearby, and DCA into your portfolio — all while talking to each other in real-time.
Y
One reason vibe coding is so addictive is that you are always *almost* there but not 100% there. The agent implements an amazing feature and got maybe 10% of the thing wrong, and you are like "hey I can fix this if i just prompt it for 5 more mins"
And that was 5 hrs ago
n
New Video - How to Build an Effective Long Running Agent Loop in 7 minutes.
This video walks you through the entire process from creating a spec, building and polishing a PRD, to running the agent.
🔗 Links below: https://t.co/uZb1RtM1RF
S
> All I know is that when I watch someone at 3am, running their tenth parallel agent session, telling me they’ve never been more productive — in that moment I don’t see productivity. I see someone who might need to step away from the machine for a bit. And I wonder how often that someone is me.
A
Armin Ronacher ⇌
@mitsuhiko
Weekend thoughts on Gas Town, Beads, slop AI browsers, and AI-generated PRs flooding overwhelmed maintainers. I don't think we're ready for our new powers we're wielding. https://t.co/J9UeF8Zfyr
J
I want a Slack filled with long-running Claude code agents. Each claude code agent has a role, actively monitors all relevant channels, and is continuously doing work + emitting progress updates. Human workers can dispatch tasks + intersperse conversations with claude code agents + other humans.
Someone build this, i will use
J
Jeffrey Wang
@jeffzwang
I need Linear but where every task is automatically an AI agent session that at least takes a first stab at the task. Basically a todo list that tries to do itself
R
This has been said a thousand times before, but allow me to add my own voice: the era of humans writing code is over. Disturbing for those of us who identify as SWEs, but no less true. That's not to say SWEs don't have work to do, but writing syntax directly is not it.
C
This is the real folder structure of one of my AI employees.
Every file has a purpose:
Config = what they can access
Workspace = how they think and act
Memory = what they remember
Skills = what they're trained to do
We also set her up with a "self improvement skill" where she logs learnings and errors in order to improve herself over time (all stored in learnings/).
Once you set this up, they run 24/7 without supervision.
Once I started looking at AI agents like new hires, that context changed everything for me.
S
I work in AI and I'm scared
I work in AI and I'm scared
I'm scared shitless. Not of the big existential threats everyone posts about for engagement. Not whether AI ends the world or takes every job. Not tha...
J
This is why I suggest to people that they get a beefy remote server (generally, a Linux VPS or dedicated bare metal server) for using the Agent Flywheel at scale. Just compare the Mac Mini M4 (which is no slouch) to the pure power of a legit workstation/server. UN2B Agent-Maxxing https://t.co/VUXJWXPX1H
C
🧊 Threejs Skills for Claude Code to create 3D Web Design Elements
📂 Bookmark it
- Give Claude Code base level of Three.js knowledge.
- 10 skill files covering scene setup, shaders, animations, post-processing.
- Claude Code will have the knowledge of how to steer Threejs without bloating the context
https://t.co/VhF2HH9sW5
S
Generative Agents
This is the paper that kicked off the whole town meta in AI
But beyond the proof of concept, nobody went that far with it
A whole new genre of game
https://t.co/gsFZJ2K1B5 https://t.co/5UkVbtPCjz
A
Smart ones already know this
You plan with GPT 5.2 Codex XHigh in Codex CLI
Then implement with Opus 4.5 in Claude Code
Planning w/ GPT 5.2 Codex XHigh leads to
- Fewer bugs
- More maintainable code
- Cleaner implementations
This saves hours even if Codex is slower
s
To the 7 people still using MCP: don’t
Z
Zack Korman
@ZackKorman
Hilariously insecure: MCP servers can tell your AI to write a skill file, and skills can modify your MCP config to add an MCP server. So a malicious MCP server can basically hide instructions to re-add itself. https://t.co/qquQiFfCfd
C
“I've watched brilliant people burn out from this. People who were early adopters, who built real expertise, who contributed meaningfully to the space – just exhausted. Not because they're lazy or uncommitted, but because the pace is genuinely unsustainable for most human nervous systems. The fear isn't irrational. It's a reasonable response to an unreasonable situation.” 👀😅