A Smaller AI Brief Works Better

Our coding assistant read the same product rule eight times before answering one question.

The duplicate text wasted context. It also made the brief harder to maintain.

Context Is a Budget

Every AI tool has a context window. It can hold only so much text at once.

Instructions, code, conversation history, and reasoning share that space.

Repeated instructions leave less room for the current task.

Think of context like a desk. Eight copies of one memo leave little room to work.

What We Found

We reviewed 31 instruction files. They contained about 4,850 lines of setup guidance.

The same color system appeared in three files.

The same product rule appeared in eight places.

A 55-line changelog repeated information already available through git log.

Several files repeated the same test instructions.

We had added guidance over several months. We did not check for overlap.

Five Fixes

Find duplicates. Search for each important rule across the instruction files.

If a rule appears several times, choose one home for it.

Layer instructions. Put broad rules at the top. Put local rules near the code they govern.

Point instead of repeating. Link to the source of truth from other files.

A short pointer keeps the rule easy to find. It also prevents drift.

Delete tracked information. Git already records changes. We removed the duplicate changelog.

Measure the result. We removed about 800 lines. That freed roughly 7,200 tokens per session.

Why This Matters

Context cleanup is not a one-time task. Every new instruction has a cost.

Good editors keep the brief focused. They leave room for the work that matters now.

The same model can work better with a clearer brief.

Check Your Setup

Try this five-minute review:

1. Count your instruction files and their total lines.

2. Search for three key terms across those files.

3. Remove information your tools already track.

The assistant did not get smarter. We stopped making it read the same thing eight times.