Kevin Wolf

Currently building Dobby, my very own coding agent on top of @badlogicgames's Pi. After spending hours trying to get Claude Code working the way I wanted, I decided to try Pi, and its extensibility is huge. Highly recommend giving it a try. Will continue posting updates about Dobby in case anyone wants to try it 🧦✨

@brolag @fmontes En que te ha dado errores? yo llevo usándolo como una semana y no me ha fallado. Mas bien me encanta el contepto de poder agregar tabs dentro de un split.

@kvnwolf @fmontes Mae esta increíble esto es justo lo que necesito. Pero esta extremadamente buggy... Cuando lo mejoren esto definitivamente va a reemplazar Ghostty vanilla.

Necesito un TODO-List app donde cada item sea una terminal (ojo una terminal, no un chat...) ¿Alguien conoce o costruyó algo así? Pregunto antes de inventar el agua tibia.

Mae si pero eso es lo que se hace al analizar imágenes, analizar un conjunto de bytes para compararlo con la data de entrenamiento. Pero piense en convertir esos bytes a tokens, la cantidad de tokens que hay que analizar vs analizar un API que indica en dónde están los elementos interactivos. Al final eso es lo que hacen los render engines, parsear JSON o algún contrato similar y pintar el UI.

@kvnwolf Yo si creo que pronto lo vamos a ver. Al final lo que vemos en el monitor es solo un stream de bytes que una IA fácilmente podria analizar en tiempo real y enviar señales de teclado y mouse para interactuar como lo hace un humano. Ya deben estar trabajándolo.

Hay algo que siento que aún le hace bastante falta a las IAs para ser ya next level útiles y es poder ejecutar UIs (no solo web, por ej, también videojuegos) y probar cosas directamente sin tener uno que ser sus "ojos". Se que para web ya hay algunos tools, pero

@jamonholmgren Looks good, just a quick suggestion to be more token efficient; agents doesn't need a markdown link, you can just put the path and they'll follow when needed, that saves you tokens.

@kvnwolf I haven't yet found a place where skills work better than my current system, so I've kept it simpler with just markdown docs. Progressive disclosure is a good way to say it. AGENTS is the router, each doc may have references to deeper docs as needed. It's worked great so far.

My current agentic workflow is about 5x faster, better quality, I understand the system better, and I’m having fun again. My previous workflows have left me exhausted, overwhelmed, and feeling out of touch with the systems I was building. They also degraded quality too much. This is way better. I’m not ready to describe in detail. It’s still evolving a bit. But I’ll give you a high level here. I call this the Night Shift workflow.

@jamonholmgren Uhm.. but almost every agent supports skills as of now, why don't you like skills? Also, how are you instructing the agent to discover those markdown files? progressive disclosure?

@kvnwolf Sounds pretty good. One of my goals is to use as vanilla of setup as possible. I do NOT use agent skills, for example. Everything is a markdown file in the repo that is agent agnostic. AGENTS.md will work for any agent. Claude, Codex, Cursor, rotating, may even try others.

I highly recommend you taking a look at Pi, you can customize it the way YOU want to work, it has all the primitives you need. That's what I'm doing, moving from a set of Claude skills for planning and executing plans to building them and executing them more programmatically with Dobby. x.com/kvnwolf/status…

@kvnwolf Claude Code or Codex or Cursor, depending on what I feel like that evening. And a bunch of markdown files!

Pi is the Vim of agentic coding harnesses. Minimal core: just read/write/edit/bash and a tiny prompt. No bloat, no forced workflows, pure extensibility. I'm building my own harness exactly my way, skill by skill, extension by extension. It truly feels like an extension of how I think and code. Big thanks @badlogicgames for bringing back the joy of crafting tools that feel mine. pi.dev

@josealvaradoo @jeudyx Pongale un hook que rechace operaciones sobre el .env

@jeudyx @kvnwolf Una vez me borró el .env y como eso no se commitea, lo perdí 😭

A veces estoy breteando con Claude Code, mando el prompt y me voy 10 mins a hacer otra vara. Regreso para ver el resultado y me encuentro esto:

Claro. Otra manera es poner post tool use hooks que corran contra el código recién generado, corriendo linter, typecheck y tests contra el archivo recién editado, asi se agarran errores rápido. Combine eso con un pre commit hook y ya despreocúpese de revisar el código y preocúpese más por shippear.

@kvnwolf @jeudyx A mi si me gusta revisar. Creo que es un tema más personal. Por que he probado el código de claude, gemini y minmax. A veces suele hacer cosas con fallos de seguridad y así. Quizas también dependa de como definiste los skills y los agentes de tus proyectos.

@RADIKAL4 @jeudyx Ponga a otra IA a revisarlo. Yo antes me estresaba por el código y me costó aprender a delegar, pero si la IA va a tocar siempre el código, por qué no desestresarse de eso?

@jeudyx @kvnwolf pero a la final te toca revisar el código o bueno yo si lo reviso por mi mismo

@tan_stack I’ll need to build a realtime voice agent soon, but using a real phone number (Twilio) does this support it?

Building realtime voice AI is usually a pain: WebRTC, audio streams, tool execution, state 😵 TanStack AI now handles that w/ useRealtimeChat()🎙️ 🔊 Voice in / Voice out 🧠 Tools/State 🖼 Multimodal ⚡ Realtime transcripts Supports OpenAI Realtime & ElevenLabs 🔗⬇🧵

AI is now writing 100% of my code. Working in TDD mode, using RED-GREEN-REFACTOR, writing E2E tests when needed and formalizing their learnings at the end for future tasks by writing skills. What a time to be alive <3

Don't sleep on skills! They're the thin line separating current models from AGI.

I built a planning & execution framework for Claude Code that treats every feature like a mini-project: /plan interviews you exhaustively about the task — edge cases, auth states, entity flows — until there are zero assumptions left. Then it explores the codebase, proposes vertical-slice tasks, and generates structured YAML plan files. /execute reads those files and orchestrates an agent team. Each task gets its own task-executor agent running in an isolated git worktree. They work in parallel, commit independently, and merge back to the plan branch one at a time. The system discovers and installs missing skills automatically, recovers from context limits by spawning replacement agents, and walks you through a guided QA review before delivering a PR. It's basically a tiny project manager that plans like a staff engineer and delegates like a tech lead.

@official_taches A personal finance tracker

What are you building?

@cescastro_ Claro papi! Escríbame

@kvnwolf Enséñeme 🤓🤓🤓 #nojoke

La IA esta haciendo a la gente no solamente perezosa si no que peores personas. Que adulan y chupan las medias hasta del mal trato que el usuario puede hacer. Validacion de la antipatia.