How to Give Your AI a Head Start

Welcome to AutomatED: the newsletter on how to teach better with tech.

In each edition, I share what I have learned — and am learning — about AI and tech in the university classroom. What works, what doesn't, and why.

Today, I close this three-part series by handing you the Work Ledger starter and walking you through installing it, using it, and judging it.

Hi—

My August 28 post argued that the software around an AI model — the harness — matters as much as the model (Agentic Harnesses Need a Proper Workspace). My August 31 post described my answer to this problem (the “Work Ledger”) and the lab-notebook discipline behind it (Teach Your AI to Keep a Lab Notebook). I also described some alternatives and their pros and cons. I ended that second post with a promise: a starter version of the Ledger you could put on a real project of your own, plus a guide to get it running.

Here they both are. I'm not going to re-argue either post today. What follows is the practical part: whether this is worth your time, and if it is, exactly what to do.

📦 What You Get Today

Work Ledger v1 is a free starter: a folder of plain text files and instructions that you open as a new project in an AI app you already use. You can download it from the release page, with the self-contained ZIP named work-ledger-v1.zip. The maintained step-by-step guide is START-HERE.md, but below I’ll walk you through the setup.

It’s been tested extensively in two agentic harnesses: Anthropic’s Claude Code and OpenAI’s Codex. You can use the former through Claude Desktop's Code tab in Local mode and the latter through the ChatGPT desktop app in Local mode. ("Local" just means the app works directly with files on your own computer.) It also works in the CLI versions of those apps, if that’s your style.

In terms of operating systems, it works on either Windows or Mac, and it is designed to be portable across those surfaces, with parallel functionality in any permutation of them.

I am not finished making it compatible with Claude Cowork. That comes later this year, along with confirmed compatibility with a range of open-source harnesses (I believe several are already compatible but further testing is needed).

It also isn’t compatible with the browser version of either assistant. If that is how you use AI day to day, this release is not for you.

I’d recommend keeping a backup of the ledger folder, the way you would for anything you would be annoyed to lose. (Git also keeps a history of changes inside that folder, which I’ll explain below. You can push your Git updates to a private online storage location as a form of backup.)

🧭 Is This Worth Trying?

This is for people with projects that keep going across many sessions: a course redesign that spans a semester, a grant revised over months, an accreditation cycle, a committee rewriting a policy, a research paper. Work where resuming from the wrong assumption costs you something — time, a bad draft, a meeting spent defending a decision that was already reversed.

If you mostly ask an assistant to tighten an email or suggest five discussion questions for tomorrow, you do not need any of this.

I’ve used it extensively for months, along with some beta users. It has performed extremely well for me across ~30 billion tokens and 5 heavy workstreams (including two with a lot of research and coding). My productivity is 5x what it was before I ended in roughly the mature Ledger state I’m sharing with you. Beta users report similarly transformative outcomes for heavy and long-running workloads.

But this is the first public release, and I expect to keep improving it as people put it to work on their own projects. So if something looks different from what I describe below, that is useful information for me, not proof that you did something wrong.

One more honest note. The first hour is more technical than every hour after it. You will open a terminal, confirm that a program called Git is installed, paste a few commands, and send one exact first message. After that, the assistant handles the mechanics and you go back to asking for things in ordinary language. Allow some uninterrupted time for setup, especially if you still need to install Git.

🧰 Before You Open Anything

Windows: install Git for Windows, which includes Git Bash. (Feel free to ask your AI assistant what Git is, or read this, if you aren’t familiar with it.) If you use Codex on Windows, you also need PowerShell 7.5 or later, which Microsoft distributes free; the PowerShell that ships with Windows is older and will not do.

Mac: macOS 14 or later. Open Terminal and type git --version. If Git is not there, macOS may offer to install Apple's command-line developer tools; let it, then try again.

After installing a prerequisite, restart your AI app before continuing.

🪜 The Walkthrough

Follow these with this newsletter open beside the app. Where Windows and Mac, or Claude and Codex, differ, the step says so; otherwise it is the same for everyone. If you get stumped, share this newsletter with your favorite AI assistant and ask for help about the relevant step, or just email me (respond to this newsletter in your inbox).

1. Download and extract. 

On the release page, under "Assets," download work-ledger-v1.zip. Skip the "Source code" links GitHub adds automatically; they are not the starter. Extract the ZIP into a new folder and keep everything inside it together. The folder you will open in your app is the one that directly contains a file called START-HERE.md — not the folder around it. After extracting you may see one folder sitting inside another; that is normal. Keep opening folders until you find the one with START-HERE.md in it.

2. Open a terminal in that folder. 

Windows: right-click an empty area inside the folder, choose "Open in Terminal," and make sure the tab says PowerShell.

Mac: open Terminal, type cd followed by a space, drag the folder in from Finder, and press Return.

On either system, then run:

git --version

Continue when it prints a version number. If it does not, stop there and email me rather than pushing forward.

3. Claude only: activate the folder's permission profile, once, with Claude closed. 

This copies one settings file into place. It tells Claude which routine file and Git commands are already allowed in this folder, so ordinary work does not stop to ask you every few minutes. Codex does not need this step; skip to step 4.

Windows (PowerShell), one line:

if (Test-Path -LiteralPath ".claude/settings.local.json") { throw "Stop: .claude/settings.local.json already exists; do not overwrite it." } else { Copy-Item -LiteralPath ".claude/policies/desktop-windows-standard.settings-local.example.json" -Destination ".claude/settings.local.json" }

Mac (Terminal), the whole block:

if [ -e "./.claude/settings.local.json" ] || [ -L "./.claude/settings.local.json" ]; then
  printf '%s\n' 'Stop: the local settings file already exists; leave it unchanged and ask for help.' >&2
else
  /bin/cp -n "./.claude/policies/desktop-macos-standard.settings-local.example.json" "./.claude/settings.local.json"
fi

Success normally prints no message. Never run the Windows version on a Mac or the Mac version on Windows.

4. Both apps: put your name on your local saves. 

In the same terminal, run these two lines, typing your real name and email inside the quotation marks in place of the examples:

git config --local user.name "Your Name"
git config --local user.email "[email protected]"

This applies only to this folder. It labels the saves the Ledger makes on your computer; it is not your AI account, and nothing is uploaded.

5. Open the folder in your app. 

Choose the exact folder from step 1 as a Local workspace — the folder itself, not its parent — and work in it directly, with worktree isolation off (a setting that keeps you working in this one folder rather than a hidden copy of it). Claude: select "Accept edits" (not Auto, not bypass). Codex: use normal editing inside this workspace. You may see one prompt asking whether you trust this new folder. That one is expected. What is not expected is a stream of permission dialogs after it. If a permission request appears that you do not understand, choose Deny and email me; do not click "Always allow" to make it go away.

6. Send the exact first message. 

Copy the message for your app, and where it says [Windows or macOS], type the one word that is true for you, without the brackets.

Claude:

Codex:

The first thing you see back should be a card with five choices. The "card" is just the assistant's reply in the chat, not a separate screen. It should arrive before the assistant does anything; if it starts acting before you have answered, stop and email me.

7. Answer the card. 

In plain terms, the card asks five questions, each with two or three choices shown right on it, and "Recommended" takes the first choice on each:

• What kinds of files will you add? (Recommended: work files already approved for this AI workspace. The alternative is a mix of work, personal, and client files, which it asks about group by group.)

• What should become reusable understanding? (Recommended: stable goals, decisions, constraints, and open questions.)

• How detailed should the source notes be? (Recommended: balanced.)

• How should it talk to you? (Recommended: brief, handling the mechanics for you.)

• How should it save? (Recommended: automatically, as ordinary local Git saves in this folder that upload nothing. The alternatives are to be asked each time, or to manage Git yourself.)

To take the recommended choices, reply:

using a short, non-secret label you recognize for your account or organization, and the same name and email from step 4 — never a password or key. Correct any line the card shows that is wrong, and fill in any line marked "not yet supplied." Once setup is complete, the assistant confirms its first local save.

8. Add your first files. 

Put them directly into the folder called add-files-here — for a first trial, something small and not confidential, with your originals kept elsewhere. Keep filenames simple — ordinary unaccented letters and numbers, spaces, periods, underscores, parentheses, and hyphens; start with a letter or number; no double periods and nothing ending in a period or a space; no longer than 120 characters. If the assistant flags a reserved or unsupported name, rename the file and try again. Then say:

The assistant stores a copy of each eligible file with a note of where it came from, shows you a short review of the understanding it extracted, and asks once what should become reusable. The review offers three choices: A saves the understanding as shown; B saves it with corrections you provide in the same reply; C keeps the files as sources only and saves no reusable understanding. It saves and ends with a readiness summary.

9. Work normally, one conversation at a time. 

Ask for real tasks in your own words. At the start of a conversation, list everything you already know the task needs; add related work to that same conversation while it is unfinished; start the next conversation after the work is completed and saved. Keep one active writing conversation per ledger folder. Two conversations changing the same folder at once can leave it half-written, and that is a bad way to learn what "half-written" means. You can always reopen an old conversation to read it.

Saving finished everyday tasks has its own one-time opt-in, separate from the card. When the assistant offers it, reply SAVE COMPLETED TASK FILES AUTOMATICALLY if you want that; otherwise tell it to leave the work unsaved as a checkpoint. The draft files remain in the folder, but they are not recorded in the Ledger's saved history.

10. Three phrases for later. 

"Update this workspace from the files I added."

"Show what this workspace currently understands."

And "Run the garden," which is designed to trigger a review of whether your Ledger contains anything stale, unresolved, or contradictory — it shows you what it has recorded first and asks before changing anything. I run it twice a week.

Once you are set up, though, I’d recommend asking your harness directly if you have any questions about what you can do. It will review the Ledger and answer accurately. For instance, if you want to setup a connector to your email account, just ask it how and it will explain how it could handle data in such a case, allowing you to specify your preferences.

🧪 The Test That Matters

Do not judge the Ledger by opening the folder and admiring the structure. The folders are arranged with care, but so is my spice rack, and neither fact tells you whether anything gets cooked.

Give it one real task on a real project. Ingest the relevant context in the form of old files. Say you are six weeks into a course redesign: the faculty group adopted a shared rubric, the oral-assessment question is still open, and you ask it to revise the assessment instructions to match the rubric without settling the open question. Then leave the project alone for several days. Come back in a fresh conversation and hand it a related second task — the syllabus-facing summary of the new plan, say — without re-pasting the background.

Then ask honestly. Did it resume from where things actually stood, or from a plan someone later dropped? Did it keep the open question open instead of quietly resolving it for you? Could it tell you why the rubric looks the way it does, and point to the record, rather than just asserting it? How much did you have to re-explain? And on balance, did it save more effort than the setup cost?

Let me know your answers!

🏁 Closing the Series

Three posts, in order. Friday: the environment around the model matters as much as the model. Monday: serious projects need something closer to a lab notebook than a chat transcript. Today: a notebook you can open and use.

This was a deliberately finite series, and it ends here. I'm not promising the return of a weekly cadence — I said at the start that I wouldn't, and I'm keeping to it. I may write other things later. What I want from you is simpler: try the Ledger once on something that matters to you, and tell me whether it earned its keep. I think it will, but you be the judge.

For installation or everyday-use help, respond to this email and it will come to my inbox. For a reproducible bug, open a GitHub issue and include the release filename, your operating system and app, the step, and what happened, with private material removed. The guide at https://github.com/automatedteach/work-ledger-starter will stay current.

A separate technical companion is planned, so stay tuned to my GitHub if you want more details on how the functionality you see in the folder relates to the architecture described in the prior posts, as well as the development process leading to this release.

P.S. I have created a new tool, the Schedule Workbench, for the unglamorous problem of assigning faculty to courses, which most departments still do through spreadsheets, email, and one person's memory. It organizes the agreed inputs, proposes assignments a human scheduler can edit, lock, use, or leave alone, and keeps a history of what changed. You can watch a video walkthrough in ~6 minutes here: scheduleworkbench.com. And if your department schedules some other way, I would like to hear how, as I am expanding its functionality every week.