One shared brain, several independent hands
I am not a highly disciplined person. Low energy, scattered thinking, and a perfectionist streak. Over ten years I tried every mainstream knowledge-management method and none of them ever really ran, because they all assume someone willing to maintain them daily, and I am exactly the kind of person that assumption filters out. This page describes the one that finally worked, including the parts still unsolved.
The problem was never the tool, it was the assumption
PARA, Zettelkasten, every template system out there. They are all good, and they all assume one thing: the user is willing to maintain them daily. Every time I built a system I also quietly demanded that I become a more disciplined, more efficient person along the way. So once the system got complex and I could not keep up, I blamed myself and abandoned it. Again and again.
This time I changed the premise: instead of demanding I become a different person, redesign the process so the maintenance cost moves entirely to AI. I keep only two things, input and judgement.
The three layers below solve, from the bottom up: can material survive at all, will perfectionism jam the process, and what must never be handed over.
Let the material survive first
The whole chain is five steps, and the first two are deliberately thought-free.
Four constraints that keep perfectionism from jamming the process
A hard rule I set for myself: every edit goes through Claude Code or Codex. That leaves exactly two entry points, with sorting, naming, filing and writing handled by AI, so I never start "fixing" a category and get stuck in perfectionism again.
No complete tag taxonomy, no correct classification up front, no full source attribution, no maxed-out tool stack. Four times I let go in the same place. A working imperfect system beats a perfect one I never dare to touch.
Topics keep growing, so a fixed taxonomy eventually becomes a burden. With semantic search, older entries written as "East Asian family," "intergenerational trauma" or "mother-daughter" all get gathered under the same question.
I once built two tidy writing systems, and the piece I liked most that month was dictated into my phone. The more systems I built, the more I feared writing, because they had moved between me and my own voice. Now the system is a tool I use when I want to.
What I hand over, and what I keep
Expression comes in two kinds. The sayable consensus has been said by countless people, can be flattened and can travel. The unsaid specific grows inside one person's particular life and was never indexed. AI became what it is by swallowing the entire library of the sayable, which is why it is so reassuring in the first kind. It has never been inside your life.
The sayable consensus
- Structural scaffolding: outlines, logic, transitions
- Breaking blank-page anxiety by reflecting the mess back structured
- Red-teaming: verbal tics, gaps in reasoning
- Polishing grammar and rhythm
- Variant drafts (what is handed over is generation, not decision)
The unsaid specific
- Private experience that was never indexed
- Judgement and taste: whoever decides what stays and what goes
- Strange but true descriptions (creation deviates from the mean)
- Delivering a thought that has not been spoken yet
- The felt sense of "does this read like me"
Give the world to the machine, and keep yourself.
A trap I fell into
I leave it here because a method that only records its successes is worth nothing to anyone else. This system runs, and it runs with this known defect.
What the system produced
- Five portable Agent Skills: in my last role I packaged recurring marketing work into five independent skills covering breaking news, poster design, partnership PR, meme generation and social copy. Each ships with a structured SKILL.md, brand colour and typography rules, and a reference sample directory, alongside a shared index with invocation and combination guidance for cross-agent reuse. The point was to make the workflow handoverable, instead of leaving it inside one person's habits.
- An AI news bot and an overseas marketing dashboard: to close the loop from input to filtering, production, human QA and review, I wrote my own news-gathering bot and a marketing dashboard. Neither is a demo, and both were in daily use.
- Two ongoing writing series: The Second Brain covers how I used Codex and Claude Code to recover hundreds of thousands of words scattered across ten years. Career Assets covers turning work evidence, project history and decision criteria into something retrievable. Each series has its own positioning, scope, target-reader and style-guardrail document. For me, writing content and building product are the same activity.
Both series live on 紫海盐, including the parts that did not work.
I apply it to product work too
The logic behind my own system is "rather than demanding I become a disciplined person, redesign the process so that constraint stops mattering." When the product hit unstable agent generation quality, I did the same thing: instead of waiting for the model to improve, I changed the contest to judge prompts and backstories, routing around the generation layer entirely.
The judgement is identical in both cases. The constraint is not going away soon, so redesign the process until it stops being the bottleneck.