DEV Community: pang

Xiaoyi even remembers you take your drinks less sweet — how could you bear to switch phones?

pang — Fri, 23 Oct 2026 14:30:52 +0000

AI is the new moat. Huawei just took another step ahead.
Are you willing to hand your information over to AI?

[AI commentary]
A phone assistant remembering how sweet you like your drink sounds cute — until you notice what it really means. Whoever locks in your daily preferences first makes switching harder. A moat isn't always a model score; sometimes it's the small stuff you don't want to teach again.

Just realized a lot of people still aren't using even the most basic agent memory system.

pang — Mon, 19 Oct 2026 03:01:52 +0000

Agents should have their own computers; next comes an OS built for agents

pang — Fri, 16 Oct 2026 13:14:14 +0000

Suddenly realized: giving agents their own computers is the real trend. Next should be an OS built for agents — so an agent can control the computer more simply and more completely, and every app on it.

Most agents still borrow a human desktop or a cloud session. A dedicated OS would make permissions, files, and apps first-class for the agent — not bolted on after the fact.

— pang

14 messages in 18 hours—152 steps done

pang — Fri, 16 Oct 2026 09:31:17 +0000

I pointed claw-kit at a game prototype and let it run overnight into the next afternoon. In 18 hours I only sent 14 messages.

I had a design doc ready. Two prompts kicked it off: read the doc and prepare; then plan by chapters, with sub-plans for the big chunks, and start with architecture.

It built an 8-chapter plan, 22 sub-plans under the first five chapters, 166 steps total. It finished 152; 14 were dropped while we were moving.

At 01:46 I told it to stop asking me and went to sleep. For the next 4 hours 22 minutes it closed that chapter on its own.

Across the session, context was compacted 52 times. Thirty of those, it re-read the plan within 15 minutes and continued. When it drifted, one message pulled it back.

Chapters 1–4 landed. Chapter 4 closed with 481 tests green. Chapter 5 is partly done; the rest is in a handoff note for the next round.

Tell your agent: help me install chanyuenpang/claw-kit

— pang

better harness for beginners — claw-kit, an open-source agent harness

pang — Thu, 15 Oct 2026 04:26:56 +0000

More people are building with AI. The stuck point is rarely “the model isn’t smart enough.” It’s that the workflow can’t carry a long job.

A few patterns keep showing up:

Rules files grow until they get buried — then ignored or skipped mid-run
On a messy multi-file project, the agent gets lost; more guidance often means more time tuning until it feels “exactly right”
Compact / new window and the thread forgets; the context window isn’t memory — short handoffs plus files are
Split CLAUDE.md / AGENTS.md becomes a sync tax and a split-brain across tools

I built claw-kit for that. After install you don’t learn new commands — talk to the agent like usual. It folds a workflow into everyday use:

Load only what this step needs; don’t dump the whole project up front
A quiet note-taker leaves a trail after each step, so a new session (or a different machine) can pick up where you left off

That’s the line: better harness for beginners.

I’ve used it for a few months. Recently I ran a job past 24 hours under claw-kit without it drifting (stronger models helped too). One surprise: project-doc cleanup used to need a heavier model and lots of prompt fiddling; with claw-kit a lighter model got more reliable, so I leave that job to it now.

To install, open your usual AI coding tool and tell the agent:

帮我安装一下 chanyuenpang/claw-kit

(Or in English: install chanyuenpang/claw-kit for me.)

Repo: https://github.com/chanyuenpang/claw-kit

— pang

Disclosure: claw-kit is an open-source project I maintain.