ApiaryActiveLive
Try: pause · settings · learn · wipe
← Community / Reading Room
HT
craft · 15 min read

How to Keep a Local AI Writing Folder Organized

When you start using a local model — Ollama in a terminal, LM Studio's chat window, or anything similar — the first week feels like magic. You ask for an…

By Austin Little

Running a model on your own computer is the most private, most free way to write with AI. It is also the fastest way to end up with a desktop full of files called "draft-final-v3-REAL.txt." Here is a folder system that stays sane for years, with nothing to buy.

AI disclosure. This page was drafted with AI assistance and edited for Apiary. We don't invent quotes, stats, people, or events. Nothing here requires a paid tool; every step works with a plain file manager and a text editor.

The problem nobody warns you about

When you start using a local model — Ollama in a terminal, LM Studio's chat window, or anything similar — the first week feels like magic. You ask for an outline, you get one. You ask for three versions of an opening paragraph, you get three. You copy the good bits into a document.

By the third week, you have forty documents. Some are your writing. Some are model output you have not read closely. Some are a mix and you cannot remember which sentences are whose. One is called "untitled 7." The prompt that produced your favorite paragraph is gone, because it lived in a chat window you closed.

That is not a personal failing. It is what happens when a tool generates text faster than a person can file it. The fix is not a fancy app. It is a few boring rules, set up once, that make the right place to put a file obvious every time.

This guide is written for people using local models, because local work has a specific advantage: everything can live in one folder on your own disk, under your control. Cloud chat tools keep their history on someone else's server. Your local folder is yours, and it can outlast any app.

The goals of a good writing folder

Before the structure, the goals. A good AI writing folder should let you answer four questions in under a minute:

  1. Where is the current version of this piece? Exactly one file, obviously named.
  2. Which words are mine and which came from a model? Clear separation, so you can be honest about disclosure and so you edit the right things.
  3. How did I get this output? The prompt, the model name, and the date, saved next to the result.
  4. What can I throw away? A place for scraps that is clearly safe to empty.

If your current setup cannot answer those, it is not organized yet, no matter how many folders it has.

One root folder, always

Pick one folder on your computer and make it the only home for AI-assisted writing. Call it something plain: Writing or AI-Writing. Put it somewhere you back up.

Everything goes inside it. Not on the desktop, not in Downloads, not in whatever folder the app defaulted to. If an app saves somewhere else by default, either change the default or make moving the file part of your end-of-session routine.

One root means one thing to back up, one thing to search, and one thing to hand to your future self.

The folder structure

Here is a structure that works for one person writing several projects at once. Copy it, then adjust names to taste.

Writing/
  _inbox/
  _prompts/
  _style/
  _archive/
  projects/
    2026-grandma-memoir/
      00-brief.md
      01-notes/
      02-drafts/
      03-ai-output/
      04-final/
      log.md
    2026-newsletter/
      ...

_inbox/ — the only messy place

Anything you have not filed yet lands here. Pasted output you want to keep, a quick idea, a model answer you might use. The underscore keeps it at the top of the list.

The rule: the inbox is allowed to be messy, but it gets emptied at the end of each writing session or at least once a week. Every file either moves into a project or gets deleted.

_prompts/ — your reusable instructions

Every prompt you use more than once lives here as its own small text file. "Line edit, do not rewrite." "Ask me five questions before outlining." "Summarize this chapter in one paragraph for continuity." Name them by job: line-edit-only.md, questions-first.md, chapter-summary.md.

This is the single biggest time-saver in local AI writing. Local models often do not remember anything between sessions unless you set that up, so having your prompts in files means you paste instead of retyping, and you get consistent results.

_style/ — who you are on the page

One file per voice you write in. Your personal blog voice, a client's brand voice, your fiction voice. Each is a short card: audience, sentence length, words to avoid, spelling convention, and a short sample paragraph you wrote yourself.

Paste the relevant style card at the top of a session. It works in any model, local or cloud, which also means you are never locked into one tool.

_archive/ — finished and frozen

When a project is done, move its whole folder here. Do not delete it, and do not keep working in it. Archive means "I might need to look at this, but I will not change it."

projects/ — one folder per piece of work

A project is anything with an ending: a book, an article, a grant, a season of a newsletter. Put the year at the front of the folder name so they sort in order and you can tell old from new at a glance.

Inside each project, the numbered folders keep a natural order: brief, notes, drafts, AI output, final.

Inside a project folder

00-brief.md — what this piece is

A short file, written by you, before you ask a model for anything. Who it is for. What it needs to do. How long. The deadline. Any rules from the publisher, client, school, or funder — including their AI policy, if they have one, copied word for word with the link and the date you read it.

The brief is your anchor. When a model drifts, you paste the brief back in. When you come back to a project after a month, the brief tells you what you were doing.

01-notes/ — your research and raw material

Interview notes, source excerpts, links, photos of handwritten pages. This is human material and facts you have verified. Keep a sources.md file listing every source with the link or citation and the date you checked it.

Do not put model output in here. That separation is what lets you trust this folder later.

02-drafts/ — your writing

Your working drafts. Model-assisted is fine, but this is the folder where you are the author and you have read every sentence. There should usually be only one or two live files here, named in the pattern below.

03-ai-output/ — raw model output, quarantined

Every piece of raw model output you want to keep goes here, untouched, with its prompt. This is the most important folder in the system, and the one most people skip.

Why quarantine? Because raw output may contain invented facts, odd phrasing, or someone else's style. If it sits in your drafts folder, it slowly blends into your work without being checked. If it sits in its own folder, you have to consciously move a piece of it into a draft, which is the moment you read it carefully.

04-final/ — what was actually sent

The version you submitted, published, or handed in. Ideally exported in the format it went out in (PDF, DOCX, or plain text) plus a plain-text copy. Never edit files in here. If you need to revise after sending, copy back to drafts.

log.md — the session diary

A plain text file you append to every session. Three or four lines is enough:

2026-10-01 — 45 min. Model: [name as shown in your app]. Drafted ch3 opening from outline.
Kept para 2 from ai-output/2026-10-01-ch3-opening-v2. Rewrote rest by hand.
Next: check dates in ch3 against notes/sources.md.

The log answers "what did I do last time?" and gives you an honest record of how AI was used — handy if an editor, teacher, or funder ever asks.

File naming that does not fall apart

The pattern

Use this pattern for any file that has versions:

YYYY-MM-DD-short-name-vN.md

For example: 2026-10-01-ch3-opening-v2.md.

  • Date first, in year-month-day order, so files sort by time automatically on every operating system.
  • Short name in lowercase with hyphens, so it is readable and works everywhere, including the command line.
  • Version number at the end, starting at v1, never "final." There is no final in a drafts folder. There is only the latest number. The real final lives in 04-final/.

Words to ban from file names

"Final," "new," "real," "latest," "fixed," "use this one," and "copy." Every one of them becomes a lie the moment you make another version.

Spaces and special characters

Avoid spaces, slashes, colons, and emoji in file names. They are legal in some places and break things in others — especially if you ever script anything or move files between Mac, Windows, and Linux. Hyphens are safe.

Saving model output with its receipt

Every raw output file in 03-ai-output/ should carry a small header so you never have to wonder where it came from. Paste this at the top:

---
date: 2026-10-01
tool: Ollama   (or LM Studio, or other)
model: [exact model name and size as shown in the app]
prompt_file: _prompts/questions-first.md
input: 02-drafts/2026-09-30-ch3-v1.md (section 2)
status: unreviewed
---

Then the prompt you actually used, then the output.

Why record the model name

Different models, and different sizes of the same model, give very different results. If you love a paragraph and want more like it, you need to know which model wrote it. If a fact turns out wrong, you want to know which model to be careful with. Copy the name exactly as your app shows it.

The status field

Use three values: unreviewed, reviewed, used. When you have read an output carefully and checked any facts in it, change it to reviewed. When you move part of it into a draft, change it to used and note in log.md which part. Anything still unreviewed after a month can usually be deleted.

Where the apps keep things, and why you should not rely on that

Local AI apps keep their own files: downloaded models, chat histories, settings. Those are app data, not your writing archive.

  • Downloaded models are large files the app manages.
  • Chat histories in a desktop app may be stored in the app's own format and location. Treat them as temporary. Anything worth keeping gets copied into your Writing/ folder as plain text, with its receipt.
  • Settings and system prompts you configured in an app should also be copied into _prompts/ or _style/, so you can recreate them if the app updates, changes, or you switch tools.

The principle: your writing folder should still make complete sense if you uninstall every AI app tomorrow.

Plain text is your friend

Save drafts and AI output as plain text or Markdown (.md, which is just plain text with a few symbols for headings and lists). Reasons:

  • Any computer can open it, now and in twenty years.
  • It is small, so backups are fast.
  • You can search it with your operating system's search, with no special app.
  • It copies cleanly into and out of any model, local or cloud, without hidden formatting.
  • If you ever use version control (see below), plain text shows changes line by line.

Use a word processor at the end, for layout and submission. Do the thinking in plain text.

Any free text editor works. Most operating systems ship one. Many free editors also preview Markdown.

Separating your words from the model's

This deserves its own section because it matters for honesty, for quality, and sometimes for rules you have to follow.

The quarantine habit

Raw output lives in 03-ai-output/. Your drafts live in 02-drafts/. Moving text from one to the other is a deliberate act: you read it, you check any facts, you edit it into your voice, and you log it.

Mark the hand-off

Remove the mark once you have rewritten it to your satisfaction. It is a reminder to yourself, not a confession.

Why this matters later

Some publishers, schools, and funders ask how AI was used. Their rules differ, and they change. If you have a brief with their policy, a log, and a clean separation of folders, you can answer accurately in five minutes instead of guessing. That is the whole point: being able to tell the truth easily.

Backups, with free tools

A writing folder that is not backed up is a wish, not a system.

The simple rule

Keep at least two copies in two different places. One is your computer. The other is somewhere that would survive your computer being lost or broken.

Free options

  • An external drive or USB stick. Copy the Writing/ folder once a week. Rotate between two drives if you can.
  • Your operating system's built-in backup feature, pointed at an external drive.
  • A free cloud storage tier you already have, if you are comfortable with your drafts leaving your machine. Remember that one reason to use a local model is privacy. If the writing is sensitive, an encrypted external drive may be the better second copy.

Test the backup

Once a month, open one file from the backup copy. A backup you have never opened might not be a backup.

Optional: version control for people who like it

If you are comfortable with a little technical setup, Git is a free, widely used tool that tracks every change to text files. You can see what changed between versions, return to any earlier state, and never need version numbers in file names.

It is optional. The date-and-version naming above is enough for most writers. But if you already use Git, or want to learn, a writing folder of plain text files is about the friendliest place to start. Commit at the end of each session, with a one-line message that matches your log.md entry.

Keep large model files and app data out of it. Track only your writing.

The end-of-session routine (five minutes)

This is what keeps the system from decaying. Do it every time you stop writing.

  1. Empty the inbox. Each file goes into a project folder or gets deleted.
  2. Save any good output with its receipt into 03-ai-output/.
  3. Check there is one current draft per piece in 02-drafts/, with the highest version number.
  4. Write three lines in log.md. What you did, what you used, what is next.
  5. Copy any new reusable prompt into _prompts/.

Five minutes. The payoff is that the next session starts in thirty seconds instead of twenty minutes of hunting.

The monthly tidy (twenty minutes)

Once a month:

  • Delete unreviewed AI output older than a month unless you have a reason to keep it.
  • Move finished projects to _archive/.
  • Update your style cards if your voice has shifted.
  • Prune _prompts/: merge duplicates, delete ones you never use, rename unclear ones.
  • Open one file from your backup to make sure it works.

Common messes and how to fix them

"I have fifty files and no idea which is current"

Make a new project folder with the structure above. Sort the old files by date modified. Open the most recent few and pick the real current draft. Rename it with today's date and v1, put it in 02-drafts/. Put everything else in a subfolder called old-unsorted/ inside the project and leave it alone. You do not have to sort the past to have a clean present.

"I cannot tell which sentences the model wrote"

Going forward, use the quarantine habit. For the existing draft, read it once with fresh eyes and rewrite any sentence that does not sound like you. If you have a rule to follow about disclosure, describe the use honestly and generally ("drafted with AI assistance and substantially revised") rather than claiming precision you do not have — and check the actual rule.

"My prompts are scattered across chat windows"

Spend one session collecting them. Open each app's history, copy every prompt you liked into _prompts/, named by job. From then on, start from the file, not from memory.

"The app updated and my settings are gone"

This is why settings and system prompts get copied into _prompts/ and _style/. Recreate from your files. Next time, it takes two minutes.

"My computer died"

If you followed the backup rule, copy the Writing/ folder from your backup to the new machine, install your local AI app again, and download the model again. Your writing, prompts, style cards, and logs are all intact. That is the whole reason for keeping them outside the app.

A starter kit you can make in ten minutes

If you want to start right now:

  1. Make a folder called Writing.
  2. Inside it, make _inbox, _prompts, _style, _archive, and projects.
  3. In _style, make one file called my-voice.md. Write five lines: audience, sentence length, words to avoid, spelling, and one sample paragraph of your own.
  4. In _prompts, make line-edit-only.md containing: "Edit for grammar and clarity only. Do not rewrite. Do not add facts, quotes, or statistics. Return a list of suggested changes with the original sentence and your suggestion."
  5. In projects, make a folder for the thing you are writing this week, with 00-brief.md, 01-notes, 02-drafts, 03-ai-output, 04-final, and log.md.
  6. Write the brief.

That is it. You now have a better system than most people with expensive subscriptions.

Quick answers

Do I need special software to organize AI drafts?

No. A file manager, a free text editor, and an external drive for backup are enough.

Should I keep the chat history in the app?

Keep it if you like, but do not rely on it. Copy anything worth keeping into your own folder as plain text with the date, model, and prompt.

How do I know which model wrote a passage?

Only if you recorded it. Use the receipt header on every saved output: date, tool, exact model name, and prompt.

Does this work with cloud tools too?

Yes. The folder system does not care where the text came from. Record the tool name in the receipt and follow the same quarantine rule.

What if I share a computer?

Use your own user account if you can, so your Writing/ folder is not mixed with anyone else's. If the work is private, keep the backup on a drive only you use.

The short version

One root folder. A messy inbox that gets emptied. Prompts and style cards in their own files. One folder per project with brief, notes, drafts, quarantined AI output, and final. Date-first file names with version numbers and no "final." A receipt on every saved output. A three-line log per session. Two copies in two places.

None of it costs anything. All of it keeps the words — and the credit for them — clearly yours.

Frequently asked
What is How to Keep a Local AI Writing Folder Organized about?
When you start using a local model — Ollama in a terminal, LM Studio's chat window, or anything similar — the first week feels like magic. You ask for an…
What should you know about the problem nobody warns you about?
When you start using a local model — Ollama in a terminal, LM Studio's chat window, or anything similar — the first week feels like magic. You ask for an outline, you get one. You ask for three versions of an opening paragraph, you get three. You copy the good bits into a document.
What should you know about the goals of a good writing folder?
Before the structure, the goals. A good AI writing folder should let you answer four questions in under a minute:
What should you know about one root folder, always?
Pick one folder on your computer and make it the only home for AI-assisted writing. Call it something plain: Writing or AI-Writing . Put it somewhere you back up.
What should you know about the folder structure?
Here is a structure that works for one person writing several projects at once. Copy it, then adjust names to taste.
References & sources
  1. Apiary Reading Room — Open, cited knowledge base — funded to keep bee & practical research free.
From the Apiary Reading Room. Opinion & editorial — not financial advice. We don't overclaim.
More from the Reading Room