Corey Epstein (e/acc)

@nbaschez Read my book (free) empire.institute

My biggest challenge with vibe coding / agentic engineering lately has been getting stuck in what I call a "plan doom loop" - have AI write a plan - review myself, seems good - have AI review plan, it always finds something - repeat It drains my time and energy to determine how important the "findings" really are Who has solved this

@burkov Claude Code is the thing your are noticing change. not the model itself

Based on months of daily work with Claude Code, I conclude that Opus 4.6, in its current state, is dumber than Opus 4.4 was when it was just released. Anthropic probably isn't cheating with benchmarks, and when they release a new model you get the real performance, but then they gradually extract the model's mojo probably the same way Fat Bastard did with Austin Powers.

@mikeysee @AnthropicAI 100% Very misanthropic

Hi @AnthropicAI could you guys just get with the program please and adopt the common standard of AGENTS.md and .agents folder for skills pls thanks. Love from an engineer trying to improve AI for customers

@GeoffreyHuntley @NickADobos A lot of software won’t be A lot of new crazy software will be. You’ll be paying for the team to babysit all of the agents running the code no one understands

@NickADobos this was my first revelation with ralph and yeah it scared the shit out of me. still does. current keynotes are all about this - is software still investable etc. ghuntley.com/real

Fascinating legal loophole

Use HQ in @OpenAIDevs Codex this one's for you @jacob_posel 😘

@bcherny @GergelyOrosz How about no DMCA takedowns? the code is out the CCP has it everyone is going to accelerate you will learn a ton from the OS community you can accelerate again, you have more session data than anyone

@GergelyOrosz This was not intentional, we’ve been working with GitHub to fix it. Should be better now.

OK, this is not April fool's: Anthripic seems to be issuing mass DMCA requests on code that are legit forks of Anthropic's own (open!) "claude code" repo This is looking terrible on Anthropic. Btw it's also breaking of the law to file a DMCA on something that never broke it

👀

Deck wanted a brand that felt "limitless." We built a visual language of portals leading into an expansive new world. My favorite touch is the wordmark, which has a portal arch in every letter.

Learnings from Claude Code leak and how to fix it: 1. Context decay is a lifecycle problem, not a bug The post says compaction destroys your context at ~167K tokens. That part is true — compaction is lossy, and you can't disable it. Where it goes wrong: the fix isn't "keep under 5 files." That's a band-aid. The real fix: treat sessions as disposable. I run a PreCompact hook that fires at 80% context. When it triggers, the agent immediately stops, serializes the entire session state (completed tasks, remaining work, blockers, git state) to a JSON thread file, and exits. The next session loads that file and resumes exactly where the last left off. No context lost. No "where was I?" confusion. The architecture assumes sessions are short and handoff-able. That's not a limitation — it's the design. The insight most people miss: stop fighting compaction. Build around it. My sessions run 20-40 minutes each and hand off clean. A 4-week project runs across dozens of sessions without regression. 2. "Done!" means nothing without verification The post claims there's an employee-only verification gate. That's fabricated. But the underlying problem is real: file writes succeed even when the resulting code doesn't compile. My fix: 140+ policy files loaded before every task. Hard-enforcement policies block on violation. A PreToolUse hook blocks PR creation unless `bun run test && bun check && bun lint` all pass. But verification at PR time is too late. I also enforce post-edit verification as a global policy — after completing edits in any repo with typecheck/lint configured, run them before reporting done. Not optional. Not "when you remember." Every time. The result: false completion claims dropped from ~30% to near zero. Not because the model got smarter — because the infrastructure won't let it lie to m 3. Big tasks collapse because one agent can't hold it all The post says "use sub-agents." Fine advice, but vague. "Launch 5-8 files per agent" is the kind of thing that sounds good and works poorly in practice. What actually works: an orchestrator that doesn't do the work itself. I built what I call the Ralph loop. It reads a PRD (project requirements doc), picks the next incomplete task, spawns a fresh Claude subprocess with isolated context, hands it the task + quality gates, waits for it to commit atomically, verifies the commit, and moves to the next task. Each task gets a clean context window. The orchestrator only tracks state — it never accumulates the code context of previous tasks. Quality gates (typecheck, lint, test) provide back-pressure: if a task breaks something, it fails immediately instead of poisoning the next 5 tasks. I've run 30-story projects through this. Not "theoretically possible" — shipped to production. 4. Generic agents lose domain knowledge. Specialized ones don't. The post tells you to use sub-agents. It doesn't tell you the harder part: a generic agent doing Stripe billing analysis will hallucinate metrics that a specialist with schema knowledge won't. I run 67 specialized workers. Each has its own YAML config, knowledge base, and learned rules. When I need LiveRecover data analysis, I don't ask Claude to "analyze the data" — I delegate to `liverecover-analyst`, which knows the database schema, the metric definitions, the Stripe webhook quirks, and the 12 gotchas we've discovered over 4 months. Workers are organized by company and skill. Company workers stay isolated — a LiveRecover worker never touches Ridgeline credentials. Each worker accumulates knowledge from task to task, session to session. The scaling unit isn't "more tokens" or "better prompts." It's specialized knowledge that compounds. 5. Grep is text matching. Your codebase isn't text. The post is right about this one. Claude Code has no AST. A rename that greps for callers will miss dynamic imports, re-exports, string references, and barrel files. My fix is a policy that forces multi-pattern search on any rename or signature change: - Direct calls - Type-level references - String literals containing the name - Dynamic imports and require() calls - Re-exports and barrel file entries - Test mocks Then run the project's typecheck to catch what grep missed. But the bigger fix is not relying on grep for discovery at all. I index all repos with qmd (local semantic + BM25 search). Instead of grepping for a function name and hoping, I search for the concept: "where is auth middleware defined." The agent finds files by meaning, not by pattern. Hooks block dangerous search patterns — you can't Glob from the HQ root (1.38M files, guaranteed timeout). You can't Glob for `prd.json` or `worker.yaml` — the hook forces you to use semantic search or direct Read. Wrong patterns fail fast and obviously. 6. Dead code is eating your context budget This one isn't in the original post, but it should be. Every unused import, dead export, and orphaned prop consumes tokens. In a 167K context window, that's space you need for the actual work. I now run a dead code scanner (knip) before any multi-file refactor. Clean first, refactor second, always as separate commits. The cleanup commit removes the noise. The refactor commit has a clean token budget for the real work. It sounds minor. It's not. On a 2,000-line file with 200 lines of dead code, that's 10% of your context budget wasted on code that contributes nothing to the task but everything to triggering compaction. 7. The system learns from its own failures This is the part nobody talks about. Every Claude Code user has the same problems session after session. Same mistakes, same gotchas, same "oh right, I forgot about that." I built a learning loop. A PostToolUse hook detects retry patterns — fixup commits, amend commits, repeated failures. At session end, it nudges: "Consider running /learn to extract patterns." The /learn command captures the gotcha, classifies it by scope (company, repo, command, global), deduplicates against existing rules, and writes it as a policy file. Next session, that policy file loads automatically. The mistake can't happen again — not because the model remembers, but because the infrastructure remembers. 140+ policies accumulated over 4 months. From "never use Prisma in Lambda handlers" to "the Integration union type cascades through all Supabase filters." Each one is a bug that cost me 30 minutes once and zero minutes forever after. 8. Company isolation isn't optional at scale If you're running one project, skip this. If you're running 17 companies with different AWS accounts, Vercel teams, Linear workspaces, Stripe keys, and Slack channels — this is existential. I maintain a manifest.yaml that maps every company to its repos, workers, credentials, deploy targets, and infrastructure. A PreToolUse hook warns when you're reading Company A's settings while working in Company B's context. Hard rules: never try another company's credentials as fallback. Never deploy to the wrong Vercel team. Never mix company knowledge in outputs. If the right credentials fail, stop and ask. This sounds paranoid. Then you realize one cross-company credential leak could expose customer data across businesses. The paranoia is earned.

Best use case of /loops in Claude Code

Most people building for the knowledge layer / memory in AI approaching it incorrectly I just got back from a conference and everyone likened AI memory to a human brain and neural network But the optimal agent harness functions more like a city vs a single entity People want to build an all knowing brain But the world was built by specialization because it is superior. Build a city. Don’t build a brain. Many workers. Many confined skill and permission sets.

@jacob_posel @bcherny @claudeai Please help @bcherny

Hey @bcherny @claudeai I'm on the $200/mo plan and blowing through usage instantly. Doesn't feel right. Is there any way to audit my account? Unfortunately I have experienced several bugs with the Claude product and I fear my plan configuration is not correct. Thanks

Most people use AI like passengers. One long conversation. Keep adding to it. AI gets confused. They blame the model. Operators do it differently: - Fresh context per task - Sequential execution - Verification before commitment A 7-year developer saw the output and said: "Teach me. I need to know what you're doing." empire.institute/book

Just had a call with my @LiveRecover dev team we moved standups to nighttime so we can kick off PRDs together and I can run projects while I'm sleeping that open PRs for them to review and merge before daybreak other devs will still work US hours, doing backpressure on the work our night team and my HQ does overnight. in the next week we are launching: - tik tok shop - new flow builder - improved agent quality checks/ML - integrations with Skio, Recharge, and Stay.ai - inbound/tap to text flows ...the list goes on days are months. weeks are years. multiplied by my entire team enabled on HQ. the acceleration is accelerating sh*t is getting very interesting

Something to learn from this

@geoffreywoo 0% chance Sales is going to thrive

18 months until most sales jobs disappear

@314p159 At least once a week When I was poor, almost daily

アメリカの人達って実際どれくらいの頻度でハンバーガー食べてるんだろうか？日本人より食ってんのかな

Very interesting stuff if true

@SteveLeq30 @LiveRecover On fire

Brands that are using @LiveRecover for a minimum of 6 months on average see a: 1. 5-6x increase in LTV 2. ~25% increase in AOV 3. 15-20% reduction in CX tickets Just a few KPI's to report on at the moment. 📈

My man @LiveRecover