← Back to Discover

Unreviewed draft. Not published content.

DISCOVER / AI × PKM

From Evergreen Notes to a Draft: Where AI Should Stop

Date
Type
Method note
Source
Andy Matuschak — Executable strategy for writing
Tags
AI × PKMTools for ThoughtAI Second Brain

Chat windows make drafting feel free. That freedom hides a cost: the model can produce fluent prose that never touched your note network—and never trained your judgment.

A sturdier path starts earlier. Andy Matuschak describes an executable strategy for writing when you already keep conceptual evergreen notes: the hard work becomes editing and assembly, not inventing structure from a blank page. AI still has a place—but mostly as capture aid and editor, not as ghostwriter of the claims that make the notes yours.

This Discover piece walks that notes→draft pipeline, shows a reverse-engineered example, and marks where contemporary AI workflows help or quietly defeat the method. It is not another “ChatGPT writing tips” list.

Verification status: Not tested by Cyber-Taoist editorial. Claims follow the cited notes, essays, and forum discussion. Treat this as an explainer of methods worth trying, not a lab report.


The problem: blank pages vs. assembled drafts

Author view (Matuschak): Writing into a blank page burns activation energy. If you already maintain many concept-oriented evergreen notes, composing an essay is closer to editing than to inventing the argument from scratch.

He sketches two directions:

  • Undirected: keep writing notes → link them into an outline when a cluster appears → fill gaps when energy is high → concatenate a first draft → rewrite for coherence.
  • Directed: gather related notes → write an outline → hang existing notes on it (and add missing ones) → concatenate → rewrite.

He points to Ahrens’s Zettelkasten practice: string selected note content along an outline into a first draft, then rewrite for flow.

Editorial analysis: Progressive Summarization (Discover #2) trains findability inside a note. This strategy trains composability across notes. Both are human Distill muscles. Neither requires an LLM—and both break if you skip to “generate the article.”


A worked picture: notes before the essay

Practitioner view (Writing Slowly): Reverse-engineering Matuschak’s long essay Exorcising us of the Primer shows roughly sixty atomic notes feeding an outline note, which then becomes a publishable article. The failure mode is the opposite habit: a “bucket” of other people’s quotes that will not stitch when a deadline arrives. Evergreen / Zettelkasten practice spends late-stage energy on prose quality, not on still digesting raw material.

Editorial analysis: The lesson for AI-era writers is blunt. If your vault is mostly clipped summaries the model wrote for you, you do not have sixty your notes—you have a prettier bucket. Assembly cannot invent a point of view you never recorded.


Where AI helps—and where it should stop

Layer AI role that usually helps Role that usually hurts
Capture (meetings, transcripts) Transcribe, chunk, literature-style summary Deciding which claims are yours
Literature / sources Compress sources you will still open Replacing reading with a surrogate you trust blindly
Evergreen promotion Suggest candidate titles or missing links (you accept/reject) Writing the evergreen note in your voice
Outline Propose order from your note titles Inventing an outline disconnected from the vault
First draft concatenate Help paste/structure from selected notes Generating the draft without the note network
Rewrite / critique Loop: critique → you keep/change/discard Accepting the model’s final voice as yours

Community view (Zettelkasten Forum): One path is letting AI draft the whole post—fast, but the note network never participates and writing skill may atrophy. Another path uses AI as an editor on a human draft: critique cycles where the author decides what stays. The tension—“finish faster” vs. “get stronger by writing”—is the editorial question, not a solved recipe.

Practitioner view (Shadow / meeting→evergreen funnel): Automate meeting capture and literature-style summaries; require a human to promote only a few items into evergreen notes with claim-like titles. Using an LLM to ghostwrite evergreen notes defeats the method—evergreens are records of your understanding.

Editorial analysis: Align the three: Matuschak’s assembly strategy assumes notes that already encode your thinking; forum writers who keep the pen and use AI as critique stay compatible; funnels that stop automation before evergreen promotion stay compatible. “One-click article from vault” is a different product category—and a different skill outcome.


Default audience: write for yourself first

Author view (Matuschak): Because evergreen notes can eventually become public writing, it is easy to write for an imagined reader too early—overload and freeze. Default to notes for yourself; opt in to audience-facing prose when you choose to publish.

Editorial analysis: AI one-click drafting amplifies the early-audience trap: every note starts performing. Keep the notes layer selfish; put performance pressure on the rewrite pass, where you already chose a thesis.


A short try path

  1. Pick a cluster of 8–20 evergreen notes you actually believe (not clips).
  2. Write an outline note that only titles and links—no new essay prose yet.
  3. Concatenate excerpts into a deliberately ugly first draft.
  4. Optional AI: ask for a critique of gaps, repetition, and weak claims—accept or reject each point yourself.
  5. Rewrite once for a reader; keep the evergreen originals intact.

Editorial analysis: If step 1 fails, you do not need a better model—you need more of your own notes (or a Distill pass on what you already captured).


Limits and failure modes

  • Matuschak’s notes are method pages, not an AI product manual—AI insertion points are editorial synthesis.
  • Forum threads are plural voices, not controlled experiments.
  • Vendor blogs that automate meeting→notes may be commercial; label them as such and keep the “human evergreen” rule even if you reject the product.
  • Skipping rewrite leaves “concatenated note voice,” which reads like a scrapbook.
  • This article did not time a full notes→essay run with or without AI (verification: none).

Bottom line

Better creation with AI is not faster ghostwriting. Executable strategy says: grow evergreen notes, outline, concatenate, rewrite—with you still authoring the claims. Let models capture and critique; do not let them promote evergreen understanding or own the final voice. That is how a note network becomes a draft without outsourcing the thinking the network was built to protect.


Sources

  1. Executable strategy for writing — Andy Matuschak
  2. How to write a book (or a long article) from your notes — Writing Slowly (2024-09-18)
  3. AI-augmented Writing from Zettelkasten — zettelkasten.de Forum (2025-10)
  4. Evergreen notes from AI meeting capture — Chad Oh / Shadow (2026-07-11); commercial product context
  5. Write notes for yourself by default — Andy Matuschak
← Back to DiscoverBack to top ↑