Hi there,

Most AI workflows feel inconsistent for one simple reason.

Nothing carries over.

You open your agent, build a feature, get decent output… and move on. When the next task starts, you’re right back at the beginning.

Same explanations. Same corrections. Same preferences.

It feels like you’re resetting every time.

Here are three things you can set up that immediately change how this feels.

1. Make your AGENTS.md actually control behaviour (10 minutes)

You probably already have an AGENTS.md file.

The problem is most of them are passive. They describe the project, but don’t actually change how the agent behaves.

Open it and add a Do/Don’t section somewhere.

Keep it tight and opinionated, and feel free to mix up general and technical constraints.

Here’s an example.

Every time the agent does something you don’t like, don’t just correct it. Add a rule:

That’s how this file becomes useful. It evolves based on real mistakes, instead of trying to be perfect upfront.

2. Turn one repeated task into a process (5 minutes)

Pick something you do often and create:

implement-feature.md

Define the steps you want followed every time:

Use it like this:

You can use this with or without plan mode in your agent.

3. Move context out of chat

Create a few simple files (you can even generate these with your agent):

• spec.md → what you’re building
• features.md → broken into steps
• decisions.md → important choices you’ve made for the feature or project

Keep them updated as you go.

Use it like this:

Once you do this, things start to feel different quite quickly.

You stop repeating yourself. Similar tasks start producing similar outputs. Decisions carry forward instead of being rediscovered. You get faster.

And most importantly, you’re not fixing the same issues over and over again.

Try just the first step today and you’ll feel the difference straight away.

Alex GS
Tech Education Lead

Unlearn is proudly part of BitterBrains – Making Developers’ Lives Easier