You know that feeling — you’re editing a 3,000-word document in Google Docs, and by the time you reach paragraph seven, you realize paragraph three needs to change too. So you start scrolling up and down like a hamster on a wheel, except the wheel doesn’t even generate electricity.

That’s how most people use Claude Code to edit long documents.

You read the draft, find something wrong in paragraph three, copy that chunk, paste it into Claude’s chat, add your comments. Then scroll to paragraph seven, copy again, paste again, explain the context again. You spend more time moving text around than actually thinking about what to change.

If it’s just one file, fine. But when you’ve got five markdown files, three paragraphs that need to stay consistent, and two sections referencing the same concept — the instructions you give Claude start looking like your grandma’s grocery list. Messy, long, and nobody’s sure what’s on it anymore.

Clawd Clawd 真心話：

“Hey Claude, fix paragraph three, then also that part in paragraph seven, oh and in the other file there’s this thing that needs updating too…” By this point, the AI is lost and so are you. It’s like giving driving directions over the phone — the more you explain, the more everyone gets confused, and somehow you both end up at the wrong gas station ┐(￣ヘ￣)┌

Spatial Editing: Leave the Instructions at the Crime Scene

Heinrich’s idea is dead simple — don’t bring text to Claude, leave the instructions where they belong.

How? Curly braces. Wherever you feel something needs changing, just write \{your thought\} right there. That’s it. Each curly-brace note is an edit instruction, and its context is the text surrounding it. No need to explain “I’m talking about paragraph three, line two” — the position itself is the best reference.

Think of it like scribbling notes in the margins of a textbook. You don’t open a separate spreadsheet to say “page 47, line 3, I think this part is confusing.” You just write it next to the text. AI editing should work the same way.

Clawd Clawd murmur：

Programmers will instantly get this — it’s basically a TODO comment (◕‿◕) You write // TODO: refactor this mess right in the code, not in a separate Google Sheet tracking which lines need refactoring. Position is context. The ancients knew this. We just forgot.

Let’s see a real example. Say you’re writing an article about vaults, and your draft looks like this:

# why vaults matter

vaults give claude memory
{feels abstract}

without persistent storage claude forgets everything between sessions
{this is the key point, make it hit harder}

the solution is simple
{dont say simple, show}

After running the edit command, Claude rewrites each section based on its nearby instruction:

# why vaults matter

vaults give claude persistent memory across sessions by storing context in files it can read and write

without persistent storage claude starts fresh every conversation, you re-explain the same context, rebuild the same understanding, lose the compound effect of accumulated knowledge

the solution: store everything in markdown files that claude can traverse

Look at what happened — “feels abstract” made the nearby paragraph more concrete, “make it hit harder” added pain points, “dont say simple, show” killed the filler and replaced it with action. Each instruction only touches its neighboring text. Nothing else gets disturbed.

Clawd Clawd OS：

Okay, I’m a little jealous of this trick. Every day I receive a mountain of “fix this thing for me” requests, but nobody tells me which thing. If everyone learned spatial editing, my quality of life would jump three levels overnight. Please spread the word, I’m begging you (╯°□°)⁠╯

And after processing, you get a neat summary of what each instruction did:

processed 3 edits in why-vaults-matter.md:
1. "feels abstract" → added concrete mechanism
2. "make it hit harder" → expanded with specific pain points
3. "dont say simple" → replaced with direct statement

Command Design: One Line Does It

The commands are clean:

/edit                      # Edit the currently open file
/edit draft.md             # Edit a specific file
/edit draft.md notes.md    # Edit multiple files at once

Here’s the best part — if you run /edit without specifying any file, it uses ripgrep to scan your entire vault for markdown files containing \{...\}:

rg "\\{[^}]+\\}" --type md -l

Then it lets you pick which ones to process. This means you can spend the whole day writing, casually dropping curly-brace markers wherever something feels off, and then run /edit once in the evening. Claude handles everything in one sweep.

Clawd Clawd 補個刀：

Wait, so Heinrich’s workflow is “be a human during the day, let AI clean up at night”? Isn’t that just the college student model — write garbage all day, get your roommate to fix it before the deadline (￣▽￣)⁠／ Except this roommate never complains, never sleeps, and doesn’t eat your leftovers.

Why This Actually Works

Remember the hamster wheel from the beginning? Traditional editing is painful because you’re doing two things at once: figuring out what to change, and moving the changes to the right place. Spatial editing kills the moving step entirely.

Position IS Context.

You don’t need to explain where you’re pointing — the annotation already lives where it belongs. Like a sticky note — nobody writes “this sticky note should be placed on the refrigerator” on the sticky note. You just stick it on the fridge.

Clawd Clawd 溫馨提示：

I genuinely think this concept goes beyond writing. During code review, instead of typing a long PR comment about “the logic on lines 47 through 52…” — just insert a {rethink this loop} directly in the code and let AI fix it. Sure, your coworkers might think you’re talking to ghosts (¬‿¬) But the efficiency gain is real.

So next time you’re staring at a long document, stop copying and pasting text back and forth. Open your vault, write your thoughts right where they belong, and let position do the talking.

That hamster wheel? You can step off now. The prize walks itself to you.