Pawel Huryn — Agentic Engineering
Builder profile: Pawel Huryn (product manager, writer at The Product Compass). Known for applying agentic tools as a non-engineer and documenting the PM workflow shift.
CLAUDE.md as Decision Engine (2026-04)
来源：Pawel Huryn，2026-04-05
核心论点：Your CLAUDE.md is probably a config file. It should be a decision engine.
不是告诉 LLM 去哪里找（where to look），而是告诉它如何思考（how to think）。
让 LLM Wiki 持续增值的三要素
Pawel Huryn 用 120+ 知识文件运行这个模式数月，发现三个关键：
1. Knowledge Architecture
INDEX.md 路由 LLM 到正确的文件，无需你指向。
不是把文件路径硬编码在 CLAUDE.md 里，而是让 LLM 先读 INDEX，自己找到相关页面。
2. Decision Journal
LLM 写回。 每次会话后更新：
- 模板（templates）
- 假设（hypotheses）
- 性能数据（performance data）
LLM 不只是在回答你的问题，它在学习如何更好地帮助你。
3. Quality Gate
硬规则覆盖默认值。
示例规则：
"When evidence contradicts instinct, trust the evidence."
这些规则不能被 LLM 的默认行为覆盖。
Without vs With
| Without | With |
|---|---|
| CLAUDE.md 是 README | CLAUDE.md 是 decision engine |
| 每次会话从零开始 | 每次会话都基于之前的积累 |
| LLM 是通用聊天机器人 | LLM 是专业领域专家 |
PMs Review Artifacts, Not Code (2026-06)
来源：Pawel Huryn，2026-06-01
In two weeks, Huryn shipped three projects — PM Brain, Claude Usage for VS Code, and Grok Build — with 800+ tests and eval checks, barely reading code. He frames this as the new PM workflow: the unit of review changed from lines of code to the artifact layer.
Seven Lessons
-
Review Artifacts, Not Code. The artifact layer is a deliberate, human-and-agent-readable summary of the solution. PMs review changes in the artifacts, not the diff.
-
Build Before Building Agreement. The old motion was "write PRD, get agreement, then build." Now the first working version is often cheaper than the meeting about it. Prototype first, align around evidence.
-
Put Strategy Where the Agent Works. Strategy lives in repo-level docs (strategy.md) the agent reads every turn. The repo becomes the product's memory.
-
Cross-Examine Confident Answers. Push back on irreversible changes, auth, billing, security, and claims that sound too clean. A second model (different family) provides more review surface than self-review.
-
Make Failures Teach the System. The best artifacts began as failures. Add the smallest test that would have caught each bug. Maintain a
fuckups.mdor equivalent. -
Point Every Agent at One Source of Truth. One master doc (CLAUDE.md or AGENTS.md), with other agents getting a tiny pointer file. Each agent can have its own entrypoint; it should not have its own reality.
-
Grant Autonomy, Enforce Boundaries. Preferences go in prompts; non-negotiables go in hooks, tool permissions, and mechanical gates. Wide lane, real walls.
What Still Needs Human Judgment
- Models over-rate their own findings (severity ratings are often inflated).
- A second model helps but is not magic — two models from the same family share blind spots.
- Taste regressions still need eyes (no agent caught a copy button overlapping header buttons).
- The workaround becomes part of what you ship — scar tissue is product judgment.
Open Questions
- Does artifact-layer review scale beyond small-team projects?
- How does the PM role change when the team includes both human engineers and multiple agent instances?
- Can the "build first, align later" model work in regulated or high-stakes environments?
Related
- claude-code/overview — Claude Code 总览
- claude-code/llm-wiki-pattern — Karpathy 的 LLM Wiki 模式
- harness-engineering/overview — Harness Engineering 总览
- harness-engineering/judgment-engineering — Judgment Engineering
- product-trends/pm-role-bifurcation-ai-era — PM Role Bifurcation