AI Strategy
Your AI Remembers Everything. That's the Problem.
The short version
- Everyone organizes their notes. Almost nobody organizes their AI, and the AI is the one reading everything on every turn.
- PARA sorts by how actionable something is, not by subject: Projects, Areas, Resources, Archives. The bucket people skip is Archives.
- Four moves: read everything once, stamp a status on every project, rebuild the index into four sections, then schedule the review.
- An audit almost always finds the same three things: orphaned files, decisions that were quietly superseded, and standing rules misfiled as projects.
A messy notes app wastes your attention when you go looking for something. A messy AI memory wastes your attention and then reasons from it, confidently.
Tiago Forte's PARA method has been in every productivity thread for a decade. Four buckets, sorted by how actionable each thing is. Most people who adopt it point it at their notes, their files, maybe their reading list.1
Almost nobody points it at their AI. Which is strange, because the AI is the only tool in the stack that reads the entire pile every single time you talk to it.
I ran PARA on my own Claude setup this week. Sixty-eight memory files, built up over about six weeks of real work. What came out of it was less a tidying exercise than a diagnosis, and the diagnosis generalizes.
01 / THE RULEActionability, not subject
This is the whole method. Everything else is bookkeeping.
The instinct when you have a pile of anything is to sort it by topic. Marketing here, finance there, product over here. It feels orderly and it is almost useless, because topic tells you nothing about whether you need the thing today.
PARA sorts by distance from action instead. Four buckets, most actionable first:
Projects
Has an outcome
Active work with a defined finish and, usually, a deadline. Finishes, then leaves.
Areas
Has a standard
Ongoing responsibility with no end date. You maintain it rather than complete it.
Resources
Has no owner
Reference and reusable method. You pull from it. Nothing is owed.
Archives
Has finished
Shipped or dormant, from any of the other three. Out of sight, revivable anytime.
The system is deliberately forgiving. Rough placement is fine, and things move between buckets as your relationship to them changes. A project becomes an archive when it ships. A resource becomes a project the day you decide to do something with it.
02 / THE MAPWhat the buckets are in an AI setup
The names transfer. What lives in each one is not obvious until you look.
| Bucket | In your AI | The test |
|---|---|---|
| Projectsactive work | Memory notes about work in flight, connected project folders, scheduled tasks with an end date. | Can you name what "done" looks like? |
| Areasstandards | Your voice rules, your process rules, your brand tokens, your standing gotchas. Usually skills or instruction files rather than notes. | Would it still be true in a year? |
| Resourcesreference | Verified research, sourced numbers, frameworks you reuse, the format skills you reach for. | Do you cite it, or work on it? |
| Archivesfinished | Everything shipped that is still sitting in active memory. This is where the gap is. | Has it happened already? |
Archives is the one that matters and the one everybody skips, because nothing forces the issue. A note about a thing you finished in June does not announce itself. It sits there looking exactly like a note about a thing you are doing on Tuesday.
The reason this is worth doing
Shipped work sitting in active memory is not a filing problem. It is a context problem. Your AI was weighing decisions you already made against work you are doing now, and it had no way to tell the difference.
03 / THE STATUS FIELDOne word per project
The cheapest change in the whole exercise, and the one that makes the review possible.
Every project note gets one word at the top saying where it stands. Three values carry almost everything:
Being worked now. Has momentum, or a live deadline, or open items you can point at.
Planned or parked. Real enough to keep, not started. Stops proposals from masquerading as commitments.
Shipped, published, deployed, or clearly done. Add one line saying where it landed so it stays revivable.
The someday value earns its place more than you would expect. Without it, anything you have written down at all reads as committed, and a pile of half-considered ideas starts applying pressure it has not earned.
Be conservative when you assign these. Under-archiving costs you a little context. Over-archiving loses a decision you still needed, and you will not notice until the model contradicts something you settled weeks ago. When you cannot tell, leave it active and flag it.
04 / THE FOUR MOVESDoing it
Under an hour for a setup of any reasonable size, most of it reading.
Read everything once
Open every file before moving anything. You cannot sort by actionability from titles alone, because whether something shipped is written in the body, not the name.
Skip this and you will archive live work.Stamp a status
One word on every project: active, someday, archived. On the archived ones add a line saying where the thing landed, so future you can find it.
status: archived · shipped to /free-ai-tips/Rebuild the index
Rewrite your top-level index into the four sections, active first. Move lines between sections, never delete them, and confirm every file appears exactly once.
Every line survives. Only its section changes.Schedule the review
A weekly recurring task that archives what shipped, flags what is stale, and reports back. Without it the active pile refills inside a month.
Friday morning. Five minutes on a quiet week.05 / THE FINDINGSWhat an audit actually turns up
Three patterns showed up in mine. I would expect them in yours.
Finding 01
Files nothing points at
Eight of my files were not listed in the index at all. They existed, they held real decisions, and nothing referenced them. They had been written in a session that never got around to indexing them, and then they went quiet.
Why it happens: writing a note and registering it are two steps, and only the first one feels like progress. The fix is mechanical: when you rebuild the index, list the directory and reconcile it against the index rather than working from the index alone.
Finding 02
Decisions that were quietly superseded
One note laid out a pricing structure in early August. Two later notes replaced it, and said so in their own text. But the original still read as current, because nothing had gone back to mark it.
This is the expensive one. A superseded decision that still looks live is worse than a missing one, because your AI will cite it with the same confidence as the version that replaced it. Archives exist precisely so a decision can be kept without being believed.
Finding 03
Standing rules filed as projects
A third of what I had classed as projects were not projects. They were standards: how to run a deploy check, which words never appear in my copy, what the brand tokens are. Things with no finish line at all.
Filed as projects they sat in the active pile forever, because nothing that never ends can ever be archived. Filed as Areas they behave correctly: always available, never pretending to be a deadline.
None of the three is exotic. They are what happens when a system grows one useful note at a time and nothing ever steps back to ask what is still true.
06 / THE REVIEWThe part that makes it hold
A one-time cleanup decays. A scheduled one does not.
The reorganization is the easy half. Everything you did will quietly undo itself over about a month unless something comes back to check. Set this as a recurring weekly task, or run it yourself on a Friday morning:
Weekly PARA review of my AI's memory. Read everything first: the index, then each topic file. For every project, check its status. - If it clearly shipped (deployed, published, committed, live at, verified complete, or a deadline that has passed with no open items left), set the status to archived, add one line saying where it landed, and move it to the Archives section. - If it is stale but has not shipped, flag it in your summary and change nothing. - If you are unsure, leave it active and ask me. Leave standing rules and reference material alone unless something about them has actually changed. Preserve the full body of every file you touch. Change status only. Preserve every line of the index. Only move lines between sections. Then message me: what is active, what you archived and why, and what looks stale. Keep it short.
Notice what the prompt does not do. It does not reorganize, rename, or rewrite. It touches one field and moves index lines. A review that can restructure your system is a review you will be afraid to run, and a review you are afraid to run does not run.
What is not active should be out of sight, not out of reach.
That distinction is the whole reason Archives is a bucket rather than a trash can. Nothing gets deleted. The June decision is still there, still findable, still exactly as detailed as the day you wrote it. It has simply stopped presenting itself as this week's business.
Your AI will believe whatever you leave lying around. The kindest thing you can do for it is tell it what is already finished.
Built with you, owned by you
Don't just organize your AI. Build it on purpose.
Sorting what you already have is the cheap win. Deciding what belongs there is the work. A build gives you the strategy first, then the standing orders, the memory, and the skills that run on it, handed over so you operate the whole thing yourself.
Prickled.ai · Goal first. People next. Tech carries.
Questions people ask about this
Straight answers, in case you skimmed.
What is the PARA method?
PARA is Tiago Forte's organizing system from Building a Second Brain. Four buckets ordered by how actionable they are: Projects have a defined outcome and a deadline, Areas are ongoing responsibilities with a standard to maintain, Resources are reference material, and Archives hold anything finished or dormant. The core rule is to organize by actionability, not by subject.
Why apply PARA to AI memory instead of a notes app?
Because the cost is different. A messy notes app wastes your attention when you go looking for something. A messy AI memory is read on every turn, so finished work is paid for whether you use it or not, and the model reasons from it confidently. Sorting by actionability makes what is live easy to find, for you and for the model.
How do I know when to archive something?
Archive when the work described is finished and no listed next action remains. Words like deployed, published, committed, live at, or verified complete are the tell. If open items are still sitting in the file, it stays active. When you are unsure, leave it active. Under-archiving costs you a little context. Over-archiving loses a decision you still needed.
What counts as an Area in an AI setup?
Standards you maintain with no end date: your voice rules, your process rules, your brand tokens, your standing gotchas. In practice they live as skills or instruction files rather than project notes. This is the most commonly misfiled bucket, because a standing rule written down on a Tuesday looks exactly like a project note.
How often should I run the review?
Weekly is enough, and a scheduled task beats intention. The review only has to do three things: archive what clearly shipped, flag what is stale but unshipped, and leave anything ambiguous alone so you can decide. Five minutes if the week was quiet.
Does this work with any AI, or only Claude?
Any assistant that keeps persistent memory, project knowledge, or a set of instruction files. The names differ and the file format differs. The rule does not: sort by how actionable something is, mark what is finished, and review it on a schedule.
Notes
- PARA (Projects, Areas, Resources, Archives) is Tiago Forte's method, set out in Building a Second Brain (2022) and on his site at fortelabs.com/blog/para. The four buckets and the organize-by-actionability rule are his. The mapping onto an AI setup, the status field, and the review prompt are my adaptation.
- The sixty-eight files, the eight unindexed ones, and the three findings are from my own Claude project memory, audited on 17 August 2026. Your numbers will differ. The patterns, in my experience, do not.
- "Read on every turn" describes how persistent memory and project instructions are typically supplied back to a model as context. Exact mechanics vary by product and change over time, so treat it as the working principle rather than a claim about any specific implementation.
- If you want the file that sits in the Areas bucket and does the most work, that is the claude.md standing-orders file, which has its own guide and a free template.