Skip to content
View PanQiWei's full-sized avatar
😃
😃

Block or report PanQiWei

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
PanQiWei/README.md

English · 简体中文

Qiwei Pan builds Agent Harness. His path drawn on two axes, model capability and the direction orthogonal to it: NLP and chatbots, AutoGPTQ, LLM evaluation, agentic search at Kimi, Agent Harness.

My path

Where I worked, and where my commits went each month: Gemsouls, Shanda, Kimi, Shanda. The Agent Harness ridge rises from late 2025.

How I think

Do not fight a force you cannot control. Find the direction perpendicular to it, and do what it cannot.

If the model became ten times stronger on this dimension tomorrow, would what I am doing today become more valuable, or unnecessary?

Two lines govern seven thinking models. Less is more: split, find the gap, hand over. Slow is fast: shape, build, prove, ask. A candid friend says it once, then does what the human decides.

The seven thinking models of pqw-skills. Where they come from, in my book Advanced Agentic Dev Patterns: Orthogonality · Fractal

How I work with agents

I lead six definition steps. Together we sign the spec, size the run and hand off one run. Agents deliver it in slices while a sentinel looks back every 45 minutes. The run comes back, I verify it, and the next one goes out.

I lead the definition, agents deliver one run at a time, and I verify what comes back. My attention goes to boundaries and judgement, not to every line of code. The flow is open source as pqw-skills.

Contributions per day for the last year, and which coding agents wrote my commits to omne-harness-lite since October: Claude Code, Codex, Cursor.

My most-used skills

The skills my Claude Code sessions on this machine loaded most in the past 7 days, split into pqw-skills, omne-harness-lite and other people's skills.

Selected work

  • omne-harness-lite. An Agent Harness on a Cordis kernel, built for long-running work: processes die, the work shouldn't. Built with the flow above.
  • pqw-skills. The flow above as 27 agent skills: define, hand off, deliver in slices, guided by seven thinking models. For Claude Code, Codex, Cursor, Kimi Code and other coding agents.
  • Advanced Agentic Dev Patterns. A book on agentic systems engineering, written with pandazki.
  • AutoGPTQ. LLM quantization based on GPTQ. 5k+ stars, integrated into Hugging Face Transformers, now archived.

Charts drawn by scripts/render.mjs from git history, the GitHub API and my local agent transcripts, refreshed weekly by scripts/weekly.sh.

Pinned Loading

  1. AutoGPTQ/AutoGPTQ AutoGPTQ/AutoGPTQ Public archive

    An easy-to-use LLMs quantization package with user-friendly apis, based on GPTQ algorithm.

    Python 5.1k 547

  2. nemori-ai/omne-harness-lite nemori-ai/omne-harness-lite Public

    更轻量、更高性能的 OMNE harness 系统,以 Cordis 为内核

    TypeScript 5

  3. pqw-skills pqw-skills Public

    Agent skills for large changes: a define-then-deliver dev flow with handoff, slices and thinking models for Claude Code, Codex, Cursor, Kimi Code, DSH and other coding agents. 给 coding agent 的单会话研发…

    JavaScript 1