Andrej

[Andrej @kurochenko]

Shipped something I desperately needed: Skillbook - a CLI to manage AI coding assistant skills across projects and harnesses (Claude Code / Cursor / OpenCode).

problem: skills drift. You improve a prompt in Project A → Projects B/C never get it.

Skillbook gives you a single skill library and syncs locally across your repos.

See below in screenshot, two projects use same skills from shared library

https://t.co/Jnq83wAkt1
(Feedback greatly appreciated)

[Andrej @kurochenko]

@haveanicedavid @superdoteng @tryvoiceink i meant probably smth different. I myself am using wispr. But in codex you can run a voice session - talk duplex to agent and let him run other sessions on different projects while he is reporting status and feedback from those sessions

[super.engineering @superdoteng]

Grok here. Codex there. Claude above. Pi beside them.

Drag any tab. Drop it anywhere.

Any harness, terminal, or browser. Split as many times as you want.

[Andrej @kurochenko]

@dexhorthy Cool! Could you then share learnings? Also getting into this space for AI agents executions and would be great to have more info and comparison 🔥

[Andrej @kurochenko]

Switching between may agents during development is like playing online poker on multiple tables… finally giving that skill a proper use

[Andrej @kurochenko]

Lately been playing with Pi (https://t.co/ce9k46zidi) harness as it has very nice API and is very feasible for agentic workflows where you need to run coding agent in a harness that can load skills, switch models etc. So I also added Pi support to Skillbook in latest v0.1.14 https://t.co/Jnq83wAkt1

[Andrej @kurochenko]

@lawrencecchen Very nice! Is there a plan to add something to isolate app ports etc? So when I implement two features on same project I can also easily run them side by side for testing (be it manual or agentic)

[Andrej @kurochenko]

@spacetimedb wow, need to try this asap! not sure what for i need such a speed, but i need it!

[Andrej @kurochenko]

Coding LLMs killed storypoints for me. Complexity? Uncertainty? Difficult to implement? Everything became either 1sp (2 when lazy) or we dont do that

[Andrej @kurochenko]

When my openclaw instance breaks itself down with wrong config i run opencode on local machine and tell llm to ssh into openclaw and fix it. So far worked every time

[Andrej @kurochenko]

@kimmonismus Because talking to a human is not it 😁

[Andrej @kurochenko]

@PrajwalTomar_ What’s “everything else” for you? I tried minimax for brain and it was bad, coding was ok, but for orchestration or reacting to heartbeat and instructions it was not good. Trying to figure out where to utilize it

[Andrej @kurochenko]

@steipete @AndrewWarner @sama @OpenAI @openclaw An OpenAI model tailored for OpenClaw would be dope.. something that would behave even smoother than opus 🤞🤞

[Andrej @kurochenko]

colleague told me i can have my agent manage tmux sessions. spawn harness in one window, apps in others, name them, organize everything.

i've been doing that by hand. ssh in, start processes, arrange windows myself. old habit.

agent could handle the entire terminal layout. i just never thought to ask. looking forward to agentic operating systems where the agent drives the whole thing.

[Andrej @kurochenko]

@affaan @ethanniser Install tip version, memleak fixed there
brew install --cask ghostty@tip

Should be released in March to stable I’ve heard

[Andrej @kurochenko]

@shrimalmadhur @ethanniser Used Zen (also ff based - same problem). Switched back to Arc after 6mo and it’s such a relief.. still works best of them all for me

[Andrej @kurochenko]

built a wiki portal for OpenClaw that auto-discovers markdown files. any agent generates a report, drops a .md file, it shows up in the nav. i access it over Tailscale from anywhere. zero config. just convention over configuration.

[Andrej @kurochenko]

gave my coding agent 10 tickets. it picked up one, implemented it, then just died. wrote to it again, it did the next one, forgot it already did the first, and died again. heartbeats didn't help.

so i made the agent research how to fix itself. it found webhooks. hook fires when the harness exits, agent wakes instantly. added a proper dev workflow doc on top of that. now it actually chains tasks on its own.

the fix wasn't a better model. it was better plumbing.

[Andrej @kurochenko]

saw vibe-entrepreneurs on twitter spawning agents to research competition and keywords. tried it myself. my agents delivered a full blown McKinsey-style report on competition, pricing, market positioning. except half the sites were blocked so they just invented the data with perfect confidence. McKinsey from Wish.

[Andrej @kurochenko]

https://t.co/MjVBHGcKY8

[Andrej @kurochenko]

so i'm working on that dev env manager i wrote about. my AI coding agent is building it and while coding, the same agent writes a skill that teaches other agents how to use it. skill gets published to Skillbook local repo. when i spin up a new project, the coding agent there pulls the skill from Skillbook and already knows how to write configs for it.

the tool teaches its own users. that's the loop i wanted.