Claude Code Skills Arms Race Intensifies While Anthropic Quietly Builds Knowledge Bases
Daily Wrap-Up
Today's feed reads like a snapshot of an ecosystem that's collectively figured out the tools are powerful but hasn't quite figured out the workflows. The dominant theme by a wide margin is Claude Code configuration: skills, AGENTS.md hygiene, session forking, and the general meta-game of optimizing your AI coding setup. At least a third of today's posts are about making the tools work better rather than shipping actual products, a pattern that @johnpalmer and @nearcyan both roasted with devastating accuracy. The self-awareness is there, but the behavior continues unabated.
The more substantive signal is the convergence around agent memory and persistent context. Anthropic appears to be working on "Knowledge Bases" for Claude, which would give agents topic-specific memories they manage automatically. GitHub Copilot's team is tackling "infinite sessions" to solve context window degradation. And individual developers are building their own solutions with RAG over previous sessions. The problem statement is clear and universal: agents are useful but forgetful, and whoever solves continuity across sessions unlocks the next productivity jump. @AstasiaMyers framed it well, arguing that the real product moat is "agent context," not the model itself.
The most entertaining moment was the dueling satire from @johnpalmer and @nearcyan, both independently arriving at the same joke: developers spending entire weekends optimizing their Claude Code setup with nothing shipped to show for it. It hits because it's true for at least half the people reading it. The most practical takeaway for developers: if you've been meaning to clean up your AGENTS.md or create skills from your best sessions, today's posts from @mattpocockuk, @matanSF, and @blader give you concrete tools to do it in an afternoon. Then actually ship something with your fancy new setup.
Quick Hits
- @VraserX summarizes Demis Hassabis arguing that LLMs lack world models and true causal reasoning, which he sees as the barrier to AI-driven scientific discovery. Interesting framing but not new ground.
- @TheAhmadOsman reports ~40% of recent projects written in Go, citing LLM compatibility, sane tooling, and single binary deployments as key advantages.
- @shanselman plugs Evan Boyle's work on GitHub Copilot CLI, calling it "killing it lately."
- @alexhillman notices overlap between someone else's AI workflow patterns and his own, plans to compare and report back.
- @shri_shobhit makes a solid case that the gap between "I can code" and "I can ship a product" is mostly ops knowledge: domains, auth, CI/CD, Docker, monitoring. Worth bookmarking if you're in that gap.
The Claude Code Skills Arms Race
The single biggest cluster of conversation today revolves around optimizing Claude Code itself. Not using it to build things, but configuring, extending, and meta-gaming the tool. The ecosystem has clearly hit the phase where power users are differentiating themselves through setup quality rather than raw prompting ability.
@matanSF announced that Droid now has /create-skill, which automatically generates a SKILL.md from any session where you demonstrated a technique. This is a smart UX move because it lowers the friction between "I showed the AI how to do something" and "the AI remembers how to do it next time." @blader took a more creative approach, feeding Claude Code the Wikipedia article on signs of AI writing and turning it into a skill that avoids all of them. As he put it:
"it's really handy that wikipedia went and collated a detailed list of 'signs of ai writing.' so much so that you can just tell your LLM to ... not do that."
On the maintenance side, @mattpocockuk shared a prompt and guide for cleaning up AGENTS.md files, warning that "bad AGENTS.md files can make your coding agent worse and cost you tokens." @theo asked about tooling to sync skills across repos with symlinks, a problem anyone managing multiple projects with Claude Code has bumped into. And @joelhooks described forcing Claude to review its own mistakes and create rules to prevent them, which is essentially TDD for agent behavior.
The self-aware comedy arrived right on schedule. @johnpalmer captured the archetype perfectly with a fake conversation where someone spends all weekend in Claude Code and, when asked what they built, can only talk about their setup being "so optimized." @nearcyan landed the same joke more concisely: "men will go on a claude code weekend bender and have nothing to show for it but a 'more optimized claude setup.'" Even @cgtwts got in on it with the quip that Rome wasn't built in a day "but they didn't have claude code." The jokes land because the optimization rabbit hole is real, and the line between productive configuration and procrastination-by-setup is genuinely blurry.
Agent Memory and the Context Problem
If there's a technical theme that unifies the most forward-looking posts today, it's persistent memory for AI agents. The problem is well-understood: agents lose context between sessions, and the workarounds are messy. Today's posts suggest multiple fronts of attack.
The biggest news is @testingcatalog's leak that Anthropic is building "Knowledge Bases" for Claude. The internal instructions describe them as "persistent knowledge repositories" that Claude should "proactively check for relevant context" and update incrementally. If this ships as described, it would be a native solution to what developers are currently solving with markdown files, RAG pipelines, and custom hooks.
On the tooling side, @_Evan_Boyle from the GitHub Copilot team revealed work on "infinite sessions" to address context window degradation:
"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... Infinite sessions solves all of this."
@PerceptualPeak demonstrated a different approach called "smart forking," which vectorizes all your previous Claude Code sessions into a RAG database, then lets you fork new sessions from the most relevant historical context. It's clever because it treats your session history as a searchable knowledge base rather than disposable transcripts.
@AstasiaMyers connected these threads to a business argument, identifying "agent context" as the core product moat in the AI era. The strongest moats, she argued, come from depth of data, formalized workflows, system integrations, and human-in-the-loop checkpoints. This framing reframes the skills-and-setup obsession from the previous section as genuine competitive advantage rather than yak-shaving. Whether you're building a product or just your personal dev environment, the quality of context you can feed an agent determines the quality of output you get back.
Multi-Agent Orchestration Takes Shape
As individual agents get more capable, the conversation is shifting toward managing fleets of them. Today brought multiple posts about dashboards, APIs, and orchestration patterns for running many agents simultaneously.
@Saboo_Shubham_ shared a Claude Code Agent UI that runs nine agents simultaneously through an RTS-style interface, declaring that "multi-agent UI will be HUGE." @MattPRD showed AgentCommand, a dashboard for monitoring 1000+ agents with real-time visibility into inter-agent communication, revenue, deploys, and code diffs. And @rauchg praised an API that abstracts over every major coding agent, calling it "extremely powerful" for building AI-powered features like auto-fixing and code review into products.
@dabit3 offered the most reflective take on what this multi-agent world actually feels like from the operator's seat:
"After 14 years of being a software developer I would have never guessed in a million years that writing, specs, and ideas would be the bottleneck for my expressivity and output. But here I am, 5 agent loops running in perpetuity, spending 100% of my time finding the fastest and most optimal ways to generate specs."
@bangkokbuild provided the most extreme example of personal agent integration, describing an AI assistant with access to their Garmin watch, Obsidian vault, GitHub repos, VPS, messaging apps, and X account. The agent now monitors health data, deploys code, tracks website visitors, monitors earthquakes, and even checks in via Telegram if the user goes quiet too long. The ambition is impressive, though the security surface area is enough to make a pen tester weep.
Vibe Coding and Creative Building
Amidst all the meta-optimization, some people are actually building things. @levelsio highlighted someone running a cluster of Claude Code terminals and vibe-coding apps toward a million-dollar revenue goal, calling them "the most interesting person shipping I've seen recently." It's a brute-force approach to product development, but it reflects a real shift in how solo builders can operate.
@startracker shared a detailed technical writeup on building a sprite animation pipeline with AI, documenting the specific challenges of getting consistent motion across sprite sheet frames. The post is refreshingly specific about what works and what doesn't, including the persistent "aura" problem with chroma keying and pixel corruption from AI-generated backgrounds that aren't truly flat colors. This is the kind of practical knowledge that only comes from actually shipping something rather than optimizing your setup.
@minchoi pointed to MCPs connecting Claude directly to Unity, Unreal, and Blender, arguing that "3D game dev is about to change forever." The MCP ecosystem continues to expand the surface area of what AI coding agents can directly manipulate, moving beyond text files into creative tools.
The Shifting Developer Role
A couple of posts today grappled with what "being a developer" means when AI handles the implementation. @richtabor described product management as "describing the problem, shaping it into a PRD, translating that into structured JSON issues, sequencing the work, and then the AI starts actually doing it." The developer's job becomes "thinking clearly, making good calls, and steering the system while it moves."
@okaythenfuture took a more macro view, arguing that capital is "increasingly losing its need for labor" and will "compound effortlessly by itself" by the early 2040s. The post urges people to accumulate capital now while the transition is still in progress. Whether or not you buy the specific timeline, the underlying dynamic of automation reducing labor's share of value creation is hard to argue with. The developers who thrive will be the ones who own the systems rather than just operate them.
Source Posts
j
forcing claude to review all of its fuck ups and create rules/skills to prevent them in the future https://t.co/NflUKCGkI2
S
I vibe coded and built a sprite animation pipeline 🛠️
(Day 22 of making the engine+game)
⬇️
Watch the video if you don't wanna read the wall of text - it directly shows what I do.
Shoutout to @jidefr for showing me a paper on black/white combination to get alpha, it's the cleanest method yet, and to @cursor_ai for enabling this entire journey.
If you prefer the wall of text here you go:
The hardest part of using general image models for 2D sprites isn’t getting a nice-looking frame, it’s getting consistent motion across a whole sprite sheet.
You can fake a sheet, but frames won’t align, timing drifts, and you end up with weird artifacts.
Even if you manually cut frames + interpolate, the animation often looks “off” because each frame is basically a new interpretation, not the same character evolving over time.
This is especially noticeable with public API models like gpt-image-1.5 and Nano Banana. Some custom LoRAs for open models exist, but this is intended for less techy folks.
My workaround: use a video model first, then post-process into a sprite sheet.
Render the animation over a solid background (white/black/magenta/green), then chroma-key it out (my engine tool supports this). If the motion stays inside the silhouette, this works surprisingly well. You can do this in almost any video editing software too!
The catch: keying almost always leaves an “aura” (edge spill).
My best results come from interpolating the keyed animation with a clean base sprite, so you keep crisp edges and only “borrow” motion/detail where needed.
If the animation extends outside the silhouette (tree branches, hair wisps, foliage), I usually skip “true sprite animation” and do it with shaders instead.
Keying can’t fully remove halos there, no matter how much feathering/tuning you do.
Another annoying issue: pixel corruption. AI rarely generates a perfectly flat background (pure #000000 or #FF00FF).
That tiny noise breaks clean extraction and creates crawling garbage pixels around the subject.
For clean base sprites (and even PBR maps), a useful trick is generating the same asset on white + black backgrounds and deriving alpha from the difference.
This is basically a matte workflow: white = opaque, black = transparent.
It fixes aura… but you’d need it per-frame to fix animation, which is still hard.
For simple pixel art (single-digit frames), you can sometimes generate a sprite sheet, then ask the model to recreate it on black/white while preserving alignment… but it’s still manual-heavy.
Honestly, at this point, for some projects it’s easier to go 3D → 2D and render clean sprites/maps directly.
But I still love pushing “pure 2D” and seeing how far we can take it.
Thanks for reading!
Follow/bookmark/repost if interested in this kind of content!
n
Wow, after 14 years of being a software developer I would have never guessed in a million years that writing, specs, and ideas would be the bottleneck for my expressivity and output.
But here I am, 5 agent loops running in perpetuity, spending 100% of my time finding the fastest and most optimal ways to generate specs for my next dozen agent loops.
And I'm realizing that the process of getting thoughts out of my head and into proper specs is an art in and of itself. And finding new and better workflows for this process is completely new and uncharted territory, there are no "best practices" because the tooling, techniques, and design space improves and expands every hour.
T
Anyone have a tool to sync between these with a symlink or something? Even just like a ./sync_skills.sh I can leave in every repo
f
flavio
@flaviocopes
How did we end up here? https://t.co/gY25cTpjCG
G
This is extremely powerful. 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.
B
BLACKBOX AI
@blackboxai
New Release: Agents API Run Blackbox CLI, Claude Code, Codex CLI, Gemini CLI and more agents on remote VMs powered by @vercel sandboxes with 1 single api implementation https://t.co/2XNRGHtAQA
A
Great piece: AI budgets are MASSIVE, and the core product moat becomes “agent context.” The strongest moats are:
• depth and cleanliness of data
• complexity and formalization of workflows
• number of system integrations
• embedded human-in-the-loop checkpoints
A
Aaron Levie
@levie
The future of enterprise software
S
Another Claude Code Agent UI
Run 9 Claude Code agents with the RTS interface.
I repeat: Multi-agent UI will be HUGE
https://t.co/piAPXikECV
S
Shubham Saboo
@Saboo_Shubham_
Multi-agent UI's will be huge in 2026. Some early signs: A2UI, AG-UI, Vercel AI JSON UI https://t.co/oXfGOG92T6
M
many users have sessions where they demonstrate a skill or technique to droid without formally creating a SKILL .md for it
to make your lives easier, droid has the slash command /create-skill, that will automatically create this Skill based on any session https://t.co/Bktsc7KVrM
@
This guy is running a cluster of Claude Code terminals vibe coding apps until he hits $1,000,000
Most interesting person shipping I've seen recently
He's on here too @matthewmillerai but doesn't seem to tweet a lot
https://t.co/2K3973Ngv1 https://t.co/pOUnuetSRA
a
amrit
@amritwt
There's a dude on YouTube, a vibe coder. He does hardcore streams and he does it for 6 hours a day with one goal in mind: to vibe code an app to a million dollars. The way he opens up 6 terminals with Claude Code running on all of them is too good. I hope he makes it. https://t.co/7NYwrf7awQ
T
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."
V
Demis Hassabis, CEO of Google DeepMind, drops a quiet bombshell:
The big question isn’t whether AI can solve problems.
It’s whether AI can invent new science.
Right now, it can’t.
Not because of compute. Not because of data.
But because it lacks something fundamental:
A world model.
Today’s LLMs can generate brilliant text, images, even code.
But they don’t truly understand causality.
They don’t know why A leads to B. They just predict patterns.
Hassabis argues that real scientific discovery requires more:
– Long-term planning
– Stronger reasoning
– And an internal model of how the world works
Physics. Biology. Cause and effect.
Only then can an AI run its own thought experiments.
Only then do we get a true digital scientist.
R
how to build an agent that never forgets
how to build an agent that never forgets
3 months ago, I was rejected from a technical interview because I couldn’t build an agent that never forgets. Every approach I knew worked… until it d...
R
this is what product management (and a lot of engineering) looks like now.
i describe the problem, we shape it together into a PRD, it translates that into structured JSON issues/user stories, sequences the work, and then starts actually doing it.
my job is less “writing specs in a doc” and more thinking clearly, making good calls, and steering the system while it moves, fast.
M
3D game dev is about to change forever.
Claude can now talk directly to Unity / Unreal / Blender... so you can build crazy 3D scenes + game with just prompts.
3 MCPs to try this weekend. Bookmark this.
1. Blender MCP
https://t.co/S7rfJ1hc6L
S
Follow Evan - he's leading the charge on the GitHub Copilot CLI which has been killing it lately
E
Evan Boyle
@_Evan_Boyle
We've been working on something internally called "infinite sessions". 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 as they can't be included in PR. Infinite sessions solves all of this. One context window that you never have to worry about clearing, and an agent that can track the endless thread of decisions.
S
If you are struggling with "I know the coding, but I don't know how to build and run a product", you need to understand the data pipeline of a modern web-app. It's a one time exercise, and many people struggle with this even when it's actually trivial.
Ignoring the product management part, just build a test web-app for once and be done with it
Learning to setup a domain name, authentication, hosting the frontend, backend services runtime, CI/CD by github actions, docker image deployment, database connection, monitoring and basic security hygiene will give you a lot more confidence and insights!
P
Prasenjit
@Star_Knight12
This guy just exposed real computer science problem https://t.co/t58NeciOJW
M
Gave Clawd access to:
• Garmin Watch
• Obsidian Vault
• GitHub repos
• VPS
• Telegram + Whatsapp
• X
Now it:
• Logs my sleep/health/exercise data and tells me when I stay up too late
• Writes code and deploys it
• Writes Ralph loop markdown files that I deploy later
• Updates Obsidian daily notes
• Tracks who visits MenuCapture and where they came from
• Monitors earthquakes in Tokyo
• Researches stuff online and saves files to my desktop
• Manage memory across sessions by remembering my projects, patterns and preferences
• Reminds me of my schedule, including holidays/accommodation
• Checks on me (on Telegram!) if I'm quiet too long
Next up:
• Reading my blood test results
• Making a morning brief with weather, health stats and calendar
• Doing a weekly review from my notes
@clawdbot 😃
Z
holy shit it fucking WORKS.
SMART FORKING. My mind is genuinely blown. I HIGHLY RECCOMEND every Claude Code user implement this into their own workflows.
Do you have a feature you want to implement in an existing project without re-explaining things? As we all know, the more relevant context a chat session has, the more effectively it will be able to implement your request. Why not utilize the knowledge gained from your hundreds/thousands of other Claude code sessions? Don't let that valuable context go to waste!!
This is where smart forking comes into play. Invoke the /fork-detect tool and tell it what you're wanting to do. It will then run your prompt through an embedding model, cross reference the embedding with a vectorized RAG database containing every single one of your previous chat sessions (which auto updates as you continue to have more sessions).
It will then return a list of the top 5 relevant chat sessions you've had relating to what you're wanting to do, assigning each a relevance score - ordering it from highest to lowest. You then pick which session you prefer to fork from, and it gives you the fork command to copy and paste into a new terminal.
And boom, there you have it. Seamlessly efficient feature implementation.
Happy to whip up an implementation plan & share it in a git repo if anyone is interested!
Z
Zac
@PerceptualPeak
Claude Code idea: Smart fork detection. Have every session transcript auto loaded into a vector database via RAG. Create a /detect-fork command. Invoking this command will first prompt Claude to ask you what you're wanting to do. You tell it, and then it will dispatch a sub-agent to the RAG database to find the chat session with the most relevant context to what you're trying to achieve. It will then output the fork session command for that session. Paste it in a new terminal, and seamlessly pick up where you left off.
M
Bad AGENTS.md files can make your coding agent worse and cost you tokens.
Here's a prompt you can use to clean them up - PLUS a full guide for folks wanting to learn more.
Enjoy: https://t.co/vDSwVQYKbT https://t.co/EJJQvP0tPy
S
it's really handy that wikipedia went and collated a detailed list of "signs of ai writing".
so much so that you can just tell your LLM to ... not do that.
i asked claude code to read that article, and create a skill to avoid all of them.
enjoy: https://t.co/Ie9IL7KsGf
📙
This is SUPER similar to some patterns I have and work extremely well, asking my assistant to compare the overlap and curious to see what the best of both might look like 👀
will report back.
S
Siqi Chen
@blader
used claude code to make a little claude code skill that learns new claude code skills as you use claude code https://t.co/IUpdeFzRtq
J
“bro I spent all weekend in Claude Code it’s incredible”
“oh nice, what did you build?”
“dude my setup is crazy. i’ve got all the vercel skills, plus custom hooks for every project”
“sick, what are you building?”
“my setup is so optimized, i’m using like 5 instances at once”
n
near
@nearcyan
men will go on a claude code weekend bender and have nothing to show for it but a "more optimized claude setup"
C
“Rome wasn’t built in a day” but they didn’t have claude code https://t.co/iS9EOq2uwf
n
men will go on a claude code weekend bender and have nothing to show for it but a "more optimized claude setup"
M
AgentCommand: a dashboard for when your AI agents are running AI agents.
Watch 1000+ agents spin up and down, see them talk to each other, and track the revenue, deploys, and code diffs happening in real-time. https://t.co/R9Km17v59V
E
We've been working on something internally called "infinite sessions".
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 as they can't be included in PR.
Infinite sessions solves all of this. One context window that you never have to worry about clearing, and an agent that can track the endless thread of decisions.
S
Scott Hanselman 🌮
@shanselman
Something is cooking in GitHub #copilot https://t.co/WZKoXGsGqA
O
This is one of the smartest things you will ever read on this site,
And it’s why I constantly tell all of you daily to get rich as fuck on here,
By the early 2040s we will likely will be in a new structural economic reality, and if you are not rich af by then(depending on your location/class status), then oh boy, I don’t think it’s going to be pleasant.
This is the Great Game we’re heading towards, capital is increasingly losing its need for labor and will compound effortlessly by itself.
You do not want to be lacking capital when that comes.
Normally I wouldn’t even quote tweet alpha like this, because I want as many of you in the dark as possible, but it’s Sunday so let me try to adhere to Christian values and be a good anon neighbor.
T
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)
@teortaxesTex
To restate the argument in more obvious terms. The eventual end state of labor under automation has been understood by smart men (ie not shallow libshits) for ≈160 years since Darwin Among the Machines. The timeline to full automation was unclear. Technocrats and some Marxists expected it in the 20th century. The last 14 years in AI (since connectionism won the hardware lottery as evidenced by AlexNet) match models that predict post-labor economy by 2035-2045. Vinge, Legg, Kurzweil, Moravec and others were unclear on details but it's obvious that if you showed them the present snapshot in say 1999, they'd have said «wow, yep, this is the endgame, almost all HARD puzzle pieces are placed». The current technological stack is almost certainly not the final one. That doesn't matter. It will clearly suffice to build everything needed for a rapid transition to the next one – data, software, hardware, and it looks extremely dubious that the final human-made stack will be paradigmatically much more complex than what we've done in these 14 years. Post-labor economy = post-consumer market = permanent underclass for virtually everyone and state-oligarchic power centralization by default. As an aside: «AI takeover» as an alternative scenario is cope for nihilists and red herring for autistic quokkas. Optimizing for compliance will be easier and ultimately more incentivized than optimizing for novel cognitive work. There will be a decidedly simian ruling class, though it may choose to *become* something else. But that's not our business anon. We won't have much business at all. The serious business will be about the technocapital deepening and gradually expanding beyond Earth. Frantic attempts to «escape the permanent underclass» in this community are not so much about getting rich as about converting wealth into some equity, a permanent stake in the ballooning posthuman economy, large enough that you'd at least be treading water on dividends, in the best case – large enough that it can sustain a thin, disciplined bloodline in perpetuity. Current datacenter buildup effects and PC hardware prices are suggestive of where it's going. Consumers are getting priced out of everything valuable for industrial production, starting from the top (microchips) and the bottom (raw inputs like copper and electricity). The two shockwaves will be traveling closer to the middle. This is not so much a "supercycle" as a secular trend. American resource frenzy and disregard for diplomacy can be interpreted as a state-level reaction to this understanding. There certainly are other factors, hedges for longer timelines, institutional inertia and disagreement between actors that prevents truly desperate focus on the new paradigm. But the smart people near the levers of power in the US do think in these terms. Speaking purely of the political instinct, I think the quality of US elite is very high, and they're ahead of the curve, thus there are even different American cliques who have coherent positions on the issue. Other global elites, including the Chinese one, are slower on the uptake. But this state of affairs isn't as permanent as the underclass will be. For people who are not BOTH extremely smart and agentic – myself included – I don't have a solution that doesn't sound hopelessly romantic and naive.
A
~40% of my projects in the last 6 months were written in Go
> LLMs handle it well
> the tooling stays sane
> single bundled binaries are
> a joy to move across machines