musha「武者道」

ai researcher & builder | content creator

[musha「武者道」 @mushaxbt]

https://t.co/REwfyXayyY

[musha「武者道」 @mushaxbt]

@omarsar0 looping workflows have been killing it for me lately, gud stuff

[musha「武者道」 @mushaxbt]

@trq212 "verifying its doing the right work" is the real shift

most teams are still stuck at "did it compile" level ambition

[musha「武者道」 @mushaxbt]

@alex_prompter u shaped the first-mover disadvantage neatly

but whose gonna admit their agent needs a burst-mode cap before they lose a friday to a 1.3x cycle

[musha「武者道」 @mushaxbt]

@LinusEkenstam love seeing builder setups come together like this. clean rebrand, curious how the design-to-deploy pipeline actually feels in practice

[musha「武者道」 @mushaxbt]

Hot take from a Sentry senior engineer: the biggest unlock from AI in a large codebase isn't generation, but comprehension. And almost everyone gets that backwards.

Priscila stopped writing code by hand in December 2025 - now she only prompts, shipping real features, refactors and bug fixes into a 15-year-old codebase that 100,000 organizations depend on.

So she had Claude analyze 116 of her own sessions. The result surprised even her: 67% of her usage was comprehension, only 2% was code generation.

Her point: in a codebase with ~100 PRs merged a day and components constantly deprecated, the real work is understanding before you touch anything. AI is the cheapest senior engineer alive, the teammate that never tires of your questions.

Because her prompts kept repeating, she wrote a skill - "catch me up" - that structures comprehension into modes like architecture, conventions, feature-trace, and history. Align your mental model first, then the code flows. Don't ship slop into the codebase that pays your salary.

That's the exact gap two of my favorite plugins close.

Misalignment is the most expensive failure in AI dev, and a model grading its own plan is a known blind spot Anthropic has flagged.

So my stack opens with grill-me (16–50+ questions deep before a line is written) and grill-me-codex (Codex adversarially reviews the plan, up to 5 rounds).

10 tools total, each killing one real failure - the 5% signal in a daily firehose of noise.

I wrote the complete blueprint - link below, save it.

[musha「武者道」 @mushaxbt]

https://t.co/U3MnEJDndC

[musha「武者道」 @mushaxbt]

@Shruti_0810 "it learns your brand" is doing a lot of heavy lifting there

but if they actually solved that part, yeah this is real

[musha「武者道」 @mushaxbt]

@AlexFinn safe version destroying benchmarks is an oxymoron and we all know it

the benchmarks they actually overtrained on tell the real story

[musha「武者道」 @mushaxbt]

@iam_chonchol the ceiling is integration cost vs lock-in risk

most brands arent ready to hand over the full loop yet

[musha「武者道」 @mushaxbt]

@Axel_bitblaze69 shouldve asked fable 5 to burn the liquidity instead, it mightve considered that one

[musha「武者道」 @mushaxbt]

@NickADobos terms of service as a class barrier is kind of the whole discussion now. wont even let u run it locally.

[musha「武者道」 @mushaxbt]

@ns123abc 3x more expensive but it cant even keep from losing its mind on a long context

[musha「武者道」 @mushaxbt]

@alex_prompter "months of lead time" became "it finds ur audience"
first is def faster but curious who trusts that pipeline blind

[musha「武者道」 @mushaxbt]

@Rana_kamran43 claude responses are already solid for dense stuff but video changes how you actually keep it in ur head. interesting take on consumption vs creation.

[musha「武者道」 @mushaxbt]

Most teams think they're feeding their agents context. They're not - and that's why you're still babysitting them.

Brandon from Unblocked nukes three myths in this talk.

Myth one: naive RAG over your docs is context. It isn't - agents hit "satisfaction of search," grab the first match they find and stop looking, like a radiologist who spots one thing and quits scanning. Not exhaustive means wrong root cause.

Myth two: just connect enough MCPs. They're pipes for access, not understanding or reasoning across what they reach.

Myth three: the million-token context window saves you. It's here - fill it and the agent can't reason over it. Bigger window, same problem.

His live comparison: the same task with only MCP access passed every code check and compiled - then the senior engineer said it would've broken the entire system if shipped. The gap isn't intelligence. It's context.

I'm a Claude Code power user and I build AI systems, so this is the exact tax I optimize against.

A whole group of my plugin stack is about giving Claude memory and a map so it stops rediscovering your repo every session - Graphify builds a queryable knowledge graph (deterministic AST, auto-rebuilds on every commit), Claude-Obsidian keeps a hot cache that carries context into the next session.

10 tools, three jobs, zero obvious names - judged by one rule: does it remove a real failure or just add surface area.

Full stack below - save it.

[musha「武者道」 @mushaxbt]

https://t.co/U3MnEJDndC

[musha「武者道」 @mushaxbt]

@swyx so Opus and GPT are plateauing hard the moment you push them and Mythos just keeps scaling with time. thats the actual win here.

[musha「武者道」 @mushaxbt]

@iamfakhrealam mythos or not id probably just call it fable 5 until someone tells me otherwise

[musha「武者道」 @mushaxbt]

@allen_lattimer @MoraHQ analytics as a search bar has been promised since 2020

curious who actually ships the product that kills excel for good

[musha「武者道」 @mushaxbt]

@Parul_Gautam7 videos fill the engagement gaps that walls of text leave open. curious if the format holds when the explainer outruns the original nuance.

[musha「武者道」 @mushaxbt]

fable 5 is free on pro and max until june 22

two weeks of anthropic's best model at $0 extra, then june 23 it starts billing usage credits.

shelve everything, this is the fortnight to build

[Claude @claudeai]

Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.

Its capabilities exceed those of any model we’ve ever made generally available.

[musha「武者道」 @mushaxbt]

@PawelHuryn this paper getting rediscovered every 6 months is the real meta analysis here