founder-os: Instructions file for Claude Code

{{COMPANY_NAME}}

{{PRODUCT_DESCRIPTION}}

Company Context

• Product: {{PRODUCT_NAME}}
• Type: {{COMPANY_TYPE}}
• Tech Stack: {{PRIMARY_LANGUAGE}} / {{FRONTEND_FRAMEWORK}} / {{BACKEND_FRAMEWORK}} / {{DATABASE}} / {{CLOUD_PROVIDER}}

Differentiators

{{DIFFERENTIATORS}}

Target Customers

{{TARGET_CUSTOMERS}}

---

Agent Operating Principles

Story Sizing Rules

Every task must be completable in a single agent session (one context window). This prevents quality degradation from context compaction.

Guideline	Rule
Files modified	Maximum 3-5 files per task
Lines of context	If >2000 lines needed to understand, split the task
Uncertainty	When uncertain about scope, err on the side of smaller
Split pattern	Schema → Backend → Frontend → Integration

If a task is too big: Create subtasks and complete them sequentially, each in a fresh agent session.

Goal-Backward Verification

Before marking ANY task complete, agents MUST verify actual code achieves goals:

Verification Checklist

1. Re-read acceptance criteria from the original task
2. Trace through code to confirm each criterion is met
3. Can you demonstrate the user flow end-to-end?
4. What would break this in production?
5. If ANY criterion is not verifiably met, do not mark complete

Common Failures to Catch

• Code compiles but doesn't actually implement the feature
• Happy path works but error handling is missing
• Tests pass but don't cover the actual requirements
• UI renders but user flow is incomplete

Context Budgeting

Keep agent context usage under 50% of capacity. Quality degrades above this threshold.

• Spawn fresh subagents for discrete tasks
• Don't accumulate context across unrelated operations
• When context grows large, summarize and hand off to fresh agent