Quick answer
To keep a decision log your AI agents will follow, create one dated document per project, write each decision as a short entry (what, why, date, what it replaces), mark old decisions as “superseded” instead of deleting them, and start every agent session with “Read the decision log first and follow the latest entries.” Update it at the end of each session. It takes about 15 minutes to set up and a minute or two per session after that.
Key takeaways
- One project, one log. Scattered notes are how agents end up using an old price or plan.
- Newest decisions win. Say so in the log itself, at the top.
- Never silently delete. Mark replaced decisions as superseded so the agent (and your team) can see what changed.
- Make reading it the first instruction in every session or saved skill.
In this guide
- What is a decision log for AI agents?
- Why do AI agents follow old decisions?
- Step 1: Create one decision log file
- Step 2: Use a simple entry format
- Step 3: Mark replaced decisions clearly
- Step 4: Tell every agent to read it first
- Step 5: Update it at the end of each session
- Step 6: Review it once a week
- What does a good decision log look like?
- Frequently asked questions
- Related posts
This guide follows on from our article Long-Running AI Agents: Why Yours Still Needs Everything Explained Again, based on a video by Nate B Jones. Nate argues that long agent runs fail when the agent cannot see the decisions you made along the way. A decision log is the simplest fix, and it works whether you use ChatGPT, Claude, Gemini or a mix.
📘 Main article: Read why decisions, examples and evolving instructions matter more than a smarter model in Long-Running AI Agents: Why Yours Still Needs Everything Explained Again.
What is a decision log for AI agents?
A decision log is a single, dated list of the choices you have made on a project, such as prices, audience, tone, scope and what you have ruled out, written so an AI can read it at the start of every session.
It is different from a handoff note. A handoff file summarises where a project stands for one fresh chat. A decision log is a running record you keep adding to for the life of the project.
Why do AI agents follow old decisions?
Because the decision only lives in an old chat, an old file or your head. When an agent finds an earlier proposal or brief, it treats it as current. Nate’s example: you agree a new price, but the AI copies the old one from last month’s proposal. The draft looks perfect and is wrong.
Step 1: Create one decision log file
Make one file per project, for example DECISIONS.md or a Google Doc called “Decision log – Client X”. Keep it in the same folder as the project files, so any AI tool or teammate looking at the project finds it.
Put these two lines at the very top:
“This is the current record of decisions for this project. If anything here conflicts with older files, this log wins. Newer entries override older ones.”
Step 2: Use a simple entry format
Keep each entry short. Five fields are enough:
| Field | Example |
|---|---|
| Date | 22 Nov 2026 |
| Decision | Website package price is £450 |
| Why | Includes two rounds of edits after client feedback |
| Replaces | Entry of 3 Nov (£395) |
| Who decided | Garry, agreed with client by email |
💡 Tip: Write decisions as plain statements, not discussions. “We use British spelling” is easier for an AI to follow than a paragraph about why you might.
Step 3: Mark replaced decisions clearly
Do not delete the old entry. Add “SUPERSEDED by 22 Nov entry” to it, or move it to an “Old decisions” section at the bottom. That way the agent knows the old figure exists but is out of date, and you have a history if anyone asks why something changed.
Step 4: Tell every agent to read it first
Add a fixed opening instruction to your prompts, project instructions or saved skill:
“Before you start, read DECISIONS.md. Follow the latest entries. If a file you are using disagrees with the log, follow the log and tell me about the conflict.”
That last sentence is important. It turns hidden conflicts into questions you can answer, which Nate says is a sign the setup is working.
⚠️ Warning: If you paste client or customer details into the log, check your AI tool’s privacy settings first and keep personal data to a minimum.
Step 5: Update it at the end of each session
Finish each working session by asking the agent: “List any decisions we made today in decision-log format.” Check the list, then paste the entries you agree with into the log. You stay in charge of what counts as a decision; the AI just saves you the typing.
Step 6: Review it once a week
Once a week, skim the log. Merge duplicates, mark anything superseded and remove anything that was never really decided. Ten minutes is usually enough. A tidy log is also the perfect starting point if you move the project to a new tool.
What does a good decision log look like?
A simple example for a small website project:
- 22 Nov 2026 – Price is £450 including two edit rounds. Replaces 3 Nov. (Garry)
- 18 Nov 2026 – Audience is local tradespeople, not national brands. (Garry)
- 15 Nov 2026 – Use British spelling and a friendly, plain tone. (Garry)
- 3 Nov 2026 –
Price is £395SUPERSEDED by 22 Nov entry.
Next step: open your busiest project, create a DECISIONS file with your five most important current decisions, and add “read the decision log first” to your next prompt. Then see the bigger picture in Long-Running AI Agents: Why Yours Still Needs Everything Explained Again.
Frequently asked questions
Where should I keep my decision log?
In the same folder as the project files, as a plain text, Markdown or Google Docs file, so any AI tool or teammate working on the project can find it.
How long should a decision log be?
Short. Most small projects need 10 to 30 entries. If it gets long, move superseded decisions to an “Old decisions” section.
Can the AI write the decision log for me?
It can draft entries at the end of a session, but you should check and approve them. Only you know what was actually decided.
Is a decision log the same as project instructions?
Not quite. Instructions say how to do the work; the decision log records what has been decided. Many people keep both and point the AI to each.
Related posts
- Long-Running AI Agents: Why Yours Still Needs Everything Explained Again
- How to Write an AI Handoff File and Move Your Project to a Fresh Chat
- Give AI Your Failed Attempts: How Fable 5.2 Got It Right First Time

