If you want Claude to run your work instead of waiting for instructions, you need two files sitting side by side: a CLAUDE.md that tells it who you are and how you work, and an actions.md that holds a live, ranked task list. CLAUDE.md is the context. actions.md is the job. Together they turn Claude from an assistant you keep directing into a project manager that already knows what needs doing when you sit down.
That’s the whole idea. The rest of this post is how to set it up.
What CLAUDE.md actually does
A lot of people have a CLAUDE.md now. It’s a plain markdown file Claude reads at the start of every session. You put in who you are, what you’re working on, how you like things done, the tools you use, the rules that matter.
It’s genuinely useful. Mine carries my businesses, my writing voice, the way I want emails drafted, the spellings I use. I stopped re-explaining myself every morning.
But on its own it’s passive.
Claude reads it, understands you, and then waits. It knows how you work. It doesn’t know what to work on today. You still have to open the session and say “right, here’s what I need.” The context is loaded; the engine idles.
That gap — knowing how you work but not what to do next — is the part I wanted to close.
The file that changes the relationship: actions.md
What made the difference for me was adding a second file next to it. Just a running task list. I call it actions.md.
Every task has three things:
- Who owns it — me, Claude, or both.
- Roughly how much it matters — a light ranking, not a spreadsheet.
- What state it’s in — to do, in progress, done.
Now the session starts differently. Claude reads CLAUDE.md for the context, then reads actions.md for the work. It picks up where we stopped. It does the things tagged for it. It logs what it finished before it closes, so tomorrow-me and tomorrow-Claude both know where things stand.
Somewhere in there it stopped feeling like an assistant I had to keep pointing at things. It started behaving like a project manager who already has the list and just gets on with it.
Here’s CLAUDE.md alone against CLAUDE.md and actions.md, side by side.
| CLAUDE.md alone | CLAUDE.md + actions.md | |
|---|---|---|
| What Claude knows | Who you are, how you work | Who you are, how you work, and what to do |
| Session start | You brief it from scratch | It reads the list and carries on |
| Who drives | You, every time | The list drives; you steer |
| Between sessions | Nothing persists | Progress is logged and picked up |
| The feeling | A clever assistant | A project manager running things |
What an actions.md looks like
Here’s a redacted version of one of mine — a couple of businesses I run, names and numbers stripped out. Yours will be shorter to start. Copy the shape, not the contents.
# actions.md — live task queue
> Owner tags: [Claude] runs it · [Hybrid] we do it together · [Me] I have to do it
> Ranking: rough ICE (impact · confidence · ease). Top of each list = do first.
> Caps: max 5 "In progress" at once. Anything below the line is parked, not lost.
## In progress
- [Claude] Draft this week's three blog posts from the tweet backlog — I8 C7 E8
- [Hybrid] Rework the pricing page copy — I finalise the numbers, Claude drafts — I9 C6 E5
- [Me] Approve and send the client update email Claude drafted — I7 C9 E9
## To do (ranked)
- [Claude] Pull last week's traffic + signups into the dashboard sheet — I6 C8 E9
- [Claude] Turn yesterday's build notes into two tweets — I5 C8 E9
- [Hybrid] Outline the onboarding sequence for a brand we run — I8 C6 E4
- [Me] Decide whether to raise the retainer on the second account — I8 C5 E7
## Parked (below the cap — revisit weekly)
- [Claude] Audit old blog posts for broken internal links — I4 C7 E8
## Done (this week — Claude logs these before closing)
- [Claude] Rewrote 6 page titles for organic click-through ✓ Mon
- [Hybrid] Shipped the quiz landing page ✓ TueA few things doing quiet work in there:
- The owner tag is the whole trick.
[Claude]means Claude just does it.[Me]means it waits for me and reminds me.[Hybrid]means it drafts and I decide. No ambiguity about who’s holding the ball. - The rough ranking stops everything looking equally urgent. I use impact, confidence, ease — a number each, out of ten. It’s deliberately quick. The point is to force a top of the list, not to be precise.
- The cap keeps it honest. Five things in progress, no more. Everything else is parked in writing, so it’s off my mind but not lost. That’s the bit that reclaims the headspace.
- The Done log is the memory. Claude writes there before it closes the session. That’s how the next session knows what happened without me explaining it.
How to set it up (about fifteen minutes)
- Create CLAUDE.md in the folder you work from. Write who you are, what the business is, how you like things done, the tools you use, the rules that matter. Plain English. Keep it to what you’d actually tell a new hire on day one.
- Create actions.md next to it. Add four headings: In progress, To do, Parked, Done.
- Add ten real tasks. Everything on your mind right now. Don’t tidy them — just get them down.
- Tag every one with
[Claude],[Hybrid], or[Me]. Be honest about what Claude can genuinely own versus what needs you. - Rank them roughly. Impact, confidence, ease — a quick number each. Sort each list so the top one is the one to do first.
- Set a cap. Five in progress feels right for most people. Park the rest under the line.
- Tell Claude the routine, once, inside CLAUDE.md: at the start of a session, read actions.md, carry on from where we stopped, do the
[Claude]items, flag the[Me]items, and log what you finish under Done before closing.
That’s it. From the next session on, Claude reads the list and starts working instead of waiting.
Why this works when a to-do app doesn’t
A to-do app holds your tasks. It doesn’t do them. You still open it, read it, decide, and act — the app is a mirror, which is how a to-do list quietly makes you less productive.
actions.md is different because the thing reading it can also act on it. The [Claude] tasks get done by the reader. The list isn’t a reminder to you; it’s an instruction to a worker who’s already read your whole context.
The overwhelm most owners carry comes from holding seventeen priorities in your head and being the only one who can move any of them. Handing the owned-by-Claude ones to something that reads the list and executes is what actually takes the weight off. It’s a second worker who already has the whole pile in front of them and can start moving it.
FAQ
What is CLAUDE.md?
CLAUDE.md is a plain markdown file that Claude reads at the start of every session. It holds your context — who you are, what you're working on, how you like things done, and the rules you want followed. It means you stop re-explaining yourself every time.
What is actions.md and how is it different from CLAUDE.md?
actions.md is a running task list that sits next to CLAUDE.md. CLAUDE.md is the context — how you work. actions.md is the job — what to do, ranked, with each task tagged by who owns it. Context tells Claude how; the action list tells it what. You need both to turn Claude into a project manager.
How do I turn Claude into a project manager?
Give it two files. Put your context in CLAUDE.md and a ranked, owner-tagged task list in actions.md. Then tell Claude, inside CLAUDE.md, to read the task list at the start of each session, work the tasks it owns, flag the ones you own, and log what it finishes before closing. From then on it picks up where you stopped instead of waiting for instructions.
Do I need to be technical to set this up?
No. Both files are plain text you write in normal English — no code. If you can write a note to a colleague explaining your work and listing what needs doing, you can build both files.
What should I tag as [Claude] versus [Me]?
Tag as [Claude] anything it can genuinely finish on its own — drafting, research, pulling numbers, first passes at copy. Tag as [Me] anything that needs your judgement or your sign-off, like approving a price or sending a client email. Use [Hybrid] when Claude drafts and you decide. Being honest here is what keeps the system trustworthy.
Try it on one project this week
Pick one project. Give it a CLAUDE.md and an actions.md, tag ten tasks, and run a session. You’ll feel the difference the first morning Claude opens the list and starts working before you’ve said a word.
I write about running a business with AI as the layer over your systems — the small setups that reclaim fifteen to twenty hours a week without you working faster. If that’s useful, join the newsletter and I’ll send the next one straight to you. If it’s not for you, unsubscribe in a click — no hard feelings.
