Build the Context Layer

You now have field meaning and one authority rule. The next step is turning those decisions into data the system can validate, retrieve, version, and test.

Start smaller than feels comfortable

A useful first library has fewer than ten rules. The existing starter file uses seven: fiscal calendar, currency handling, reporting exclusions, attribution, deprecation, an allowlist, and known query anti-patterns. Each prevents a failure you already understand.

Do not document every field. Import existing help text as a baseline, then add context where a missing definition changes an answer.

Give every entry a governance envelope

A rule needs an ID, title, category, scope, status, owner, source, version, and review date. Field entries add authority, provenance, freshness, value maps, and dependencies. Store the rule body as data rather than burying it in one system prompt.

id: field.renewal_risk_score
title: Use renewal risk score for renewal health
category: field-interpretation
scope: Account.renewal_risk_score
authority: authoritative
owner: revenue-operations
status: active

Titles are retrieval keys

“Opportunity exclusions” is a folder name. “Exclude test and CPQ-generated opportunities from pipeline totals” tells the planner exactly when to fetch the rule. Write titles as conditions, not labels.

Always load what the agent cannot notice is missing

Exclusions, access boundaries, temporal rules, currency rules, and known anti-patterns belong in the stable prefix. Optional interpretation can be selected from an index. A skipped definition may cause a hedge. A skipped exclusion creates a confident wrong total.

Build the smallest safe payload in the GTM Lab. Presentation preferences do not belong in it.

Reference

• Canonical rule schema →
• Copy the seven-rule YAML →
• Read the full starter-library walkthrough →