3.8After the scaffold exists

Once Claude Code has created the initial repo state, the normal loop becomes:

Planner → Generator → Evaluator

At that point, the repo should already contain:

• CLAUDE.md (under 200 lines, with @imports)
• AGENTS.md
• .claude/rules/ path-scoped rules
• .claude/agents/ native subagent definitions
• .claude/skills/ repeatable workflows
• .claude/settings.json with hooks
• architecture docs
• standards docs
• plan templates
• starter page objects
• starter API clients
• starter fixtures
• starter tests
• CI skeletons
• artifact conventions

How to work after scaffolding

Start new sessions in acceptEdits mode for day-to-day work:

claude --permission-mode acceptEdits

Use skills to create new artifacts consistently:

/new-page-object ProductDetailPage
/new-test-spec "user adds item to cart"

Delegate investigation to subagents to keep your main context clean:

use a subagent to investigate how the checkout flow handles payment errors

Run the evaluator after completing a sprint:

use the evaluator subagent to verify the current sprint contract

Clear context between unrelated tasks with /clear. Long sessions with accumulated irrelevant context degrade Claude's performance.

Auto memory keeps learning. As you correct Claude's code style, naming, or patterns, it saves those corrections automatically. Future sessions start with that knowledge built in.

Then the agents can begin creating:

• new test cases
• new page objects
• new reusable components
• new API clients
• new fixtures
• deeper regression coverage
• refactors guided by the standards and plans