Thread by @php100

Claude Code went from genius to garbage.

That’s what most devs say - and I felt it too.

So I logged 72 hours in the terminal. The results will surprise you.

Here’s what I found 🧵:

1/ I spent 3 straight days debugging SwarmStation using Claude Code.

Logged everything. Watched every regression.

Day 1: Claude solved complex issues.
Day 3: It started misfiring on basic prompts.

But that wasn’t the full picture.

2/ I tracked everything: inputs, outputs, session resets.

The degradation wasn’t random.

It followed patterns..

And those patterns pointed to one thing:
Claude is sensitive to how you prompt.

3/ The model didn’t "get dumber" -
it became more context-sensitive.

What looked like degradation was actually:
• Prompt fatigue
• Implicit memory gaps
• Command misalignment

Claude wasn’t broken.
It was reacting.

3/ Most devs use Claude like this:

[bash]
claude -p "fix the signup bug"

That worked fine.. at first.

But with repeated commands, ambiguous phrasing, or missing repo context...

Claude flailed.

4/ So I changed tactics.

- Added strict prompt structure.
- Injected issue context explicitly.
- Reset Claude sessions more often.

Suddenly, the model stabilized.
Success rate shot back up.

5/ I baked this into SwarmStation:

• One Claude agent per GitHub issue
• Clean environment per task
• Traceable PRs

The system works - at scale.

6/ Lesson?

Claude Code isn’t just a plug-and-play tool.
It’s a conversational interface with memory quirks.

To make it work, treat it like a senior dev:
• Clear task
• Full context
• Don’t overload it

7/ I embedded these insights into SwarmStation:

- Agents get clean prompts
- Each runs in isolated workspaces
- PRs created with proper traceability

The result?
Autonomous coding flows that actually ship.

swarmstation.com

8/ What started as a “garbage” story became a design lesson.

Claude doesn’t need babysitting -
It needs structured input and environmental isolation.

Get those right, and it works better than ever.

9/ Treat it like a high-context engineer:

- Clear prompt.
- No noise.
- Reset often.
- Guide with structure.

Claude Code still ships.
You just need to learn how to drive it.

Are you using Claude Code?

10x your workflow with my Swarm Station.

Join beta while it's free 👇
swarmstation.com