DEV Community

Ovidiu Drobotă
Ovidiu Drobotă

Posted on

How I Fixed Claude’s Biggest Limitation for Real Projects (And Stopped Wasting Tokens)

If you've used Claude for anything beyond quick prompts, you've likely hit the same wall I did.

At first, it feels almost magical.
It understands context, writes clean code, and explains things well.

And then… it forgets everything.

Not partially. Not gradually. Completely.

*The Real Problem Isn't Claude — It's How We Use It
*

Claude isn't designed to remember across sessions.
Every new conversation is a clean slate.

That becomes a serious issue when you're working on anything non-trivial:

multi-file projects
long-running features
iterative debugging
structured systems

You end up doing something absurd:

Re-explaining your entire project… every single time.

And it gets worse.

You forget to mention something → it breaks things
It “fixes" code that was already correct
You burn tokens rebuilding context instead of progressing

At some point, I realized:

This isn’t a prompt problem. It’s a system problem.

The Shift: Treat Claude Like an Agent, Not a Chatbot

The breakthrough came when I stopped treating Claude like something you “talk to”
and started treating it like something you operate.

Claude doesn't need better prompts.
It needs structure, memory, and rules.

So I built a simple framework around that idea.

The System That Fixed It

Instead of relying on memory (which doesn’t exist), I externalized it.

*1. Persistent Memory Files
*

I created a few simple files that act as Claude's “brain”:

project.md → what this project is, architecture, goals
rules.md → constraints Claude must follow
state.md → current progress, what’s done, what’s next

Every session starts by feeding this back in.

No guessing. No drift.

*2. Clear Behavioral Rules
*

Claude behaves much better when given strict boundaries.

For example:

don’t rewrite working code unnecessarily
verify assumptions before changing structure
respect existing architecture

Without this, it tends to “over-help” and break things.

*3. Session Handoff Protocol
*

Before ending a session, I force a structured summary:

what was done
what changed
what remains
known risks

That gets written into state.md.

Next session → instant continuity.

The Result

This changed everything.

Instead of:

repeating myself
fixing regressions
fighting context loss

I got:

continuity between sessions
stable progress
far less token waste

Claude went from "impressive but frustrating"
to something I can actually build long-term.

*Why This Matters
*

Most people are trying to solve this with better prompts.

That won't work.

You're trying to patch a stateless system with stateless inputs.

The only real solution is to introduce a state outside the model.

*Final Thought
*

Claude is already powerful enough.

The bottleneck isn't the model —
it's the lack of a system around it.

Once you fix that, everything changes.

*If You're Curious
*

I documented the full setup (with exact structure, examples, and workflow) in a short manual.

If you're building anything serious with Claude, it'll likely save you time and frustration.

Happy to share more details or answer questions here.

Top comments (0)