Your AI Instructions Are Going Stale — Here's the Four-File Fix

The Problem Nobody Talks About

You set up an AI assistant. You write instructions. It works great — for a week.

Then you add a new rule. And another. And another. Six months later, your instruction file is a graveyard of contradictory guidance. The AI can't tell what's current. It follows a rule from March that conflicts with the one you added yesterday. You stop trusting it. You stop maintaining the instructions. The whole thing decays.

This isn't a theory. It happened inside OpenAI's own team on a 5-month, 1-million-line coding project. Their agent instruction file became what they called a "graveyard of stale rules" — new guidance piled on old, agents couldn't tell what was still true, and humans stopped maintaining it because it was too messy.

Nate B Jones documented the fix they came up with. It's simple, it's practical, and it transfers directly to how small businesses should manage AI context.

The Four-File System

Instead of one bloated instruction file, split your AI's context into four separate files that change at different speeds:

1. Stable Rules (changes: rarely)

This is your constitution. Core principles that almost never change.

• "Always confirm before sending emails to customers"
• "Use professional tone in all client communications"
• "Never share pricing outside approved rate cards"

These are the guardrails. Write them once, review quarterly. If you're changing stable rules more often than that, they're not stable — move them to file 2.

2. Current State (changes: weekly or monthly)

This is what's true right now. Business facts that shift over time.

• "We're running the summer promotion through September 15"
• "Our current service area is the greater metro region"
• "Lead response SLA is 2 hours during business hours"

When something changes — a new promo, a new service area, a new policy — update this file. Delete the old entry. Don't append. Replace.

3. Map of Material (changes: as needed)

This is your index. A living table of contents for everything the AI might reference.

• Customer FAQ document → /docs/faq.md
• Pricing sheet → /docs/pricing-2026.pdf
• Process for handling refunds → /docs/refund-process.md

Think of it as the card catalog at a library. The AI doesn't need to memorize every document — it just needs to know where to find them when the situation calls for it.

4. History (changes: daily or per session)

This is the running log. What happened recently, what's been decided, what's in progress.

• "Aug 19: Customer XYZ requested contract modification — sent to legal"
• "Aug 18: Updated summer promo end date to Sep 15"
• "Aug 17: Onboarded new team member for scheduling"

This file is your working memory. It accumulates fast, and that's fine — because it's expected to be messy. When it gets too long, archive it and start fresh. The key: history is separate from current state, so cleaning it up doesn't risk deleting something important.

Why This Works for Small Businesses

Most SMBs using AI are still at the "one big prompt" stage. They dump everything into a single instruction block — company policies, current promotions, tone guidelines, recent decisions — and wonder why the AI gets confused after a few weeks.

The four-file system solves three specific problems:

Contradictions get caught. When your summer promo ends, you update Current State. You don't have to hunt through a wall of text to find and delete the old promo rule. The old version is gone, not buried.

Maintenance becomes manageable. Reviewing four small files is doable. Reviewing one massive instruction block is not. Most SMB owners stop maintaining AI instructions because it's too painful to edit a 3,000-word prompt. Split files make each edit surgical.

The AI gets cleaner context. Models work better with focused, relevant context than with a disorganized dump of everything. Four files with clear purposes give the AI better signal and less noise than one file with no structure.

How to Start Tonight

1. Open your current AI instructions (wherever they live — a prompt, a doc, a config file)
2. Sort every line into one of four categories: stable rule, current state, reference index, or recent history
3. Create four files with those headings
4. Delete the old monolith
5. Point your AI tools at the four new files instead

This takes 30 minutes for most small businesses. The payoff is immediate — your AI starts working with cleaner context from day one.

The Deeper Lesson

The four-file system reflects a principle that applies beyond AI: information has different half-lives, and you should manage each half-life differently.

Stable rules are bedrock. Current state is weather. References are maps. History is a diary. Mixing them together is like storing your constitution, your grocery list, your atlas, and your journal in the same folder — technically possible, practically disastrous.

If you're using AI for more than a one-off task, your instructions are already going stale. The four-file system is the simplest fix that scales.

---

The four-file system comes from Nate B Jones's analysis of OpenAI's internal agent context management. The SMB adaptation and advice framework above is from SMB Scale Up.