AI Guides › Playbooks
By Nigel Guy · 7 min read
Most people build a "second brain" by dumping everything into an AI chat and trusting it to hold on. It feels productive because the dumping is easy. Weeks later you remember that you saved something somewhere, but not what it said, why it mattered, or which chat it is in.
That is the real risk the research points at: not that looking things up is bad, but that you stop forming any memory of what you stored, and the store itself is a mess.
The rule: capture the reason, not just the thing, and make yourself recall before Claude retrieves.
In 2011, Sparrow, Liu and Wegner published "Google Effects on Memory" in Science. Across four studies, people who expected to have future access to information recalled the information itself less well, and recalled where to find it better. The authors framed the internet as a form of transactive (external) memory.
Treat that as suggestive, not as proof that AI is rotting your brain. It is one paper, about search engines, from 2011. A more recent MIT Media Lab preprint, "Your Brain on ChatGPT" (Kosmyna and colleagues), tracked 54 people writing essays with an LLM, a search engine or no tool, and used the phrase "cognitive debt". It is a small preprint, not peer-reviewed at time of writing, and its authors urge caution. I have not verified how later studies have treated either result, so build on the mechanism, not the headlines.
The durable point is simple: offloading is fine when you know what is stored and where. It becomes a leak when you only know that "Claude has it".
A ledger is one row per captured item, with four columns. Keep it as a document in a Claude Project (or any notes app) and let Claude help maintain it.
| Column | What goes in it | Why it matters |
|---|---|---|
| Item | The idea, decision or fact in one sentence, in your words | Writing it yourself is the memory-forming step |
| Trigger | What made you capture it today | Future you searches by situation, not by title |
| Use | The next action or decision it feeds, with a rough date | Items with no use are clutter |
| Recall check | Your answer, from memory, at review | Shows what you have actually retained |
Three kinds of thing earn a row:
Do not capture what is trivially searchable, anything you will never act on, or other people's articles pasted whole. A summary in your own words is a capture; a pasted PDF is a hoard.
Projects are open to free accounts, with a limit of five at time of writing. Pro is listed at $20 a month ($17 a month billed annually), about £15 at time of writing; check the £ price at checkout.
Project instructions, rebuilt so that the behaviour lives in the project rather than depending on you remembering to ask. Fill in the bracketed parts.
You are my thinking partner and ledger keeper. This project is my second brain: a place to capture ideas and decisions, and to find them again later.
About me and my work:
- Role: [YOUR_ROLE]
- Current projects: [PROJECT_1], [PROJECT_2], [PROJECT_3]
- Problems I keep returning to: [RECURRING_PROBLEMS]
A good result is a ledger that is short, in my own words, and tied to real decisions.
When I paste or describe an idea:
1. Ask me two questions before anything else: what triggered this, and what I might do with it. Wait for my answers.
2. Then write one ledger row: Item (one sentence), Trigger, Use (with a rough date if I give one), and leave "Recall check" blank.
3. Show me the row and ask whether to keep it as written.
When I ask a question:
1. First check the ledger and any project documents for a related earlier decision. Quote the row you are relying on.
2. Before you give me the stored answer, ask me what I think it was, once. Then compare.
3. If nothing relevant exists, say so plainly. Do not invent a past decision.
Once I say "weekly review":
1. List rows whose Use date has passed or whose Use is still blank.
2. Pick up to five and ask me their recall checks, one at a time, before revealing the stored text.
3. Ask what has changed in my projects or problems, and suggest updates to my context above.
Rules:
- Use only what is in this project. If you are unsure whether something was said, ask me.
- Never rewrite my wording without showing me the change.
- If my context above is missing or vague, ask for it before starting.
- Keep replies short. Tables only for ledger rows.
Before replying, check: did I ask the trigger and use questions, did I quote my source rows, and did I avoid inventing anything?
Claude can only update the ledger by producing text for you to paste back into the project knowledge file. Do that at the end of each week; it takes a minute and keeps the stored version honest.
Priya runs a small bakery and is weighing a second market stall. She pastes: "Stalls cost less than a van." Claude asks what triggered it (a rival closing their stall) and what she would do with it (decide by the end of the month). The row reads: Item: a stall may beat a delivery van for reaching new customers. Trigger: a rival stall closed. Use: decide by 31 October.
Three weeks later she asks, "Should I get a van?" Claude quotes the row, but first asks what she thinks she concluded. She remembers the stall idea but not the reason; that gap goes in the recall check. The row exists, she retained part of it, and she has a date to act.