You open Claude, ask a question, get a great answer. The next day you open a new chat and start from scratch. It doesn't know what you're working on. It doesn't know your constraints. Every session is day one.
This is how most people use Claude. It works up to a point. You get useful outputs, but you spend half the session briefing the AI instead of working with it. The problem isn't Claude. The problem is the setup.
I've been running businesses, client accounts, and my own day through Claude for months now. What shifted everything wasn't a better prompt technique. It was building the context layer first — so every session starts already working.
Why reactive prompting has a ceiling
When you use Claude without context — a clean chat, no background files, no history — you're treating it like a search engine. Type a question, get an answer, close the tab.
That's fine for one-off lookups. It fails when you're trying to move real work forward.
Real work has context. Your projects have constraints, history, stakeholders, and decisions already made. When Claude doesn't know any of that, the outputs are generic — useful as a starting point, but needing heavy editing, or you spend ten minutes briefing it before the session actually starts.
The ceiling is real. Reactive prompting is useful; it's just not leveraged.
The two files that change how Claude works
The fix is straightforward, and it involves two plain text files.
The first is a CLAUDE.md file. It lives in your project folder and tells Claude who you are, how you work, what the project is, the decisions you've already made, and the tone you write in. You write it once, update it occasionally, and every session reads it automatically. It's the context layer.
The second is an actions.md file — a live task list that lives next to it. Every task is tagged with who owns it (you, Claude, or both) and roughly ranked by importance. Claude reads this at the start of a session, picks up from where you stopped, and logs what it finishes before it closes.
I wrote about setting this up in detail in how to turn Claude into your project manager with two files — worth reading if you want the full implementation. The short version: CLAUDE.md tells Claude who you are. Actions.md tells Claude what to do today. You need both.
The sessions change completely. You go from directing the AI every time you open it to opening a session that's already running.
The productivity system has to come first
Here's what I've found after running this for months: the context layer only delivers leverage if the work itself is organised first.
As I wrote in why AI is powerful but only works on top of a real system: the AI amplifies whatever structure you already have. A well-organised workload gets moved faster. A chaotic, unprioritised one gets moved faster too — you just move through the wrong things more efficiently.
AI won't fix your productivity problems if those problems are structural. The order matters.
Capture everything out of your head first. Process what lands using the 5 Ds — Do, Defer, Delegate, Drop, Diary — so nothing sits in inbox purgatory. Set your Rocks for the week: the three or four things that actually have to move. Then bring Claude in as the execution layer on top of that.
When Claude knows your Rocks, your constraints, and what's already been decided, the outputs are calibrated to your actual situation rather than a generic best-practice answer.
What this looks like in a real working day
A typical morning for me now: open a session. Claude reads the CLAUDE.md — it knows my businesses, my voice, my current priorities. It reads the actions.md — it knows what's in progress, what finished yesterday, what's waiting on me. I type a task or just say “continue.” The session is already working before I've thought about what to brief it on.
For a client account, it might pick up a half-written proposal, flag that the pricing section is outstanding, draft it to my format, and log the update. For my own content, it might draft a blog post against a brief I set last session, check the internal links, and queue the next one.
The sessions produce more output and require less mental overhead on my end. The context compounds — it gets richer with every session, so the outputs get better over time. That's the part that matters: it's not just faster, it's compounding.
As I covered in why context beats prompts — the way you structure what the AI knows is more valuable than the cleverness of what you ask it.
Frequently asked questions about using Claude AI for productivity
Do you need to be technical to set this up?
No. CLAUDE.md is a plain text file — no code, no configuration, just writing in plain English. Actions.md is a structured to-do list with simple tags. If you can use a notes app, you can run this system.
Does this work with the free version of Claude?
The free tier works for getting started and experimenting. For serious ongoing work, Claude Pro is worth it — the longer context window matters when you're reading project files and working through complex tasks in a single session.
What happens when Claude makes a mistake?
It does, occasionally. That's why the actions.md includes a log — you can see exactly what it did, verify it before it matters, and correct it. Every output is reviewable. Building in review checkpoints is part of the system, not an afterthought.
How is this different from just chatting with Claude normally?
Normal chat is stateless — each session starts from zero. The two-file setup creates persistent context, so you're not re-briefing. The difference compounds over weeks: the more history and context accumulates, the more calibrated the outputs become to your actual work.
Can this work for a team, not just an individual?
Yes. The same principles apply — a shared context file for the project, a shared task list. The main adjustment is making sure the task ownership tags are clear so Claude knows what's a human decision versus what it can move on. The context-first principle applies to teams as much as individuals.
Where to start
If you're using Claude and feel like you're constantly re-briefing it, you're not doing it wrong — you're just missing the context layer.
Build the CLAUDE.md file first. Add the actions.md. Let the context accumulate.
The productivity system provides the structure. Claude provides the execution layer on top. Between the two, the work moves faster and the decisions stay in your hands.
If you're not sure where to start with the underlying system, the NP System free course covers the full Capture/Process/Plan/Do framework — the structure that gives Claude something real to work with. Or take the productivity quiz to find out where your current system has the biggest gap.
