"Ban AI entirely" is too coarse a rule. I use AI at full throttle building knowledge-gardener / supernemawashi / taskdog / ttymap, and none of it gives off the "I'm not actually thinking" feeling. But ask AI to "write 1500 words about X" for a blog post and it instantly reads as AI-written.
What's the difference? Lay it out and you see that build and write lose different layers.
So the core of "feels AI-written" is the loss of thinking-ownership, not the use of AI itself. This post raises the resolution on that, and shifts the boundary from the prompt to the workflow. It's also itself one of the test pieces written with the workflow described below.
Rewrite the rule and it becomes:
This isn't "relaxing the rule" — it's raising the rule's resolution so you protect exactly what you meant to protect.
But re-deriving this per-prompt leaves a judgement cost behind. Asking yourself "am I letting AI do the thinking right now?" on every write is tiring.
[before] [shifted to the workflow]
draw the boundary in prompt design thinking is already pooled in the vault
→ control AI at the moment of → by the time it reaches AI, thinking is done
writing → AI is a summarizer, not a thinker
Instead of guarding the boundary with a temporal shift, guard it with vault provenance.
| Layer | Content | Treatment in the workflow |
|---|---|---|
| Felt sense | Did I type it with my own fingers | Discard it (AI summary is fine) |
| Structural | Whose thinking is it recorded as | The vault commit log secures provenance |
Discard the felt sense and lean on the structural. As long as the git history of my thinking lives in the vault, the fact that "I thought this" doesn't change even if AI writes the prose at the blog stage.
[1] Accumulate thinking through normal vault operation (fleeting → permanent)
[2] When I want to blog it, I pick the source notes myself ← thinking layer
[3] I write the outline myself (angle, claims, order) ← thinking layer
[4] Ask AI to "compress this outline + these notes into an N-word post"
[5] I check the AI output (is it on-angle / are any claims missing) ← thinking layer
[6] Rewrite with my own fingers (intervene at the articulation stage) ← thinking layer
[7] publish
Steps [1]-[3] and [5]-[6] are the thinking layer. AI only touches [4] and partial reference-material generation in [6].
As a side effect, it dogfoods my own knowledge-gardener plugin, and the vault → AI summary → blog workflow itself becomes distribution material (a post like this one).
The thinking layer of this draft (angle, claims, examples, conclusion) comes from permanent notes in my Obsidian vault — how I run that vault is here. AI was involved only in dressing up the surface layer — making this post one of the first test pieces of the workflow it describes.
This post was written with AI assistance and reviewed by the author.