AI CRM
The guiding question for this context file
"What is true of every feature, component, and workflow within this product?"
Every piece of information you include should be answerable with "yes, this applies to everything in this branch." If it only applies to some children, it belongs lower down the tree. If it repeats something already in a parent file, leave it out.
What to include
- ✓What the product does and who it is for — a clear, jargon-free description an AI agent can use to answer questions
- ✓The product's current status and maturity — is it live, in beta, being deprecated?
- ✓The primary user journey — what does a user actually do with this product from start to finish?
- ✓The core technical architecture at a high level — what kind of system is it?
- ✓Integrations and dependencies — what does this product consume and what does it expose?
- ✓Key constraints every contributor must know — performance budgets, data access rules, accessibility requirements
- ✓Agent-specific guidance — how should an AI agent behave when helping users of or contributors to this product?
- ✓What this product explicitly does not do — its boundaries
What to avoid
- ✕Granular implementation details that belong in code comments or technical docs
- ✕Sprint-level plans or near-term tasks — context files describe stable truth, not current work
- ✕Duplicate of information already in parent context files — agents receive the full chain, so don't repeat what's above
Examples of good context sentences
"AI CRM uses a buyer intent model that scores each buyer's likelihood to purchase within 30 days. Agents should not reveal model scores directly to buyers — they are for dealer and internal use only."
"This product does not handle payment or contract signing. It hands off to the Dealertrack integration once a buyer expresses purchase intent."
General principles
🌲
Tree-aware: Agents receive all ancestor files too. Don't repeat what's already above you in the tree.
⏱
Stable, not live: Write things that stay true for months. Roadmap items, sprint tasks, and current incidents belong elsewhere.
🤖
Agent-first: Write as if briefing a capable colleague who has never worked here. Be explicit; don't assume implied knowledge.