CCAR-F scenarios › Code Generation with Claude Code · Question 15 of 60
Scenario 2 · Code Generation with Claude Code
Your engineering team relies on Claude Code to generate, refactor, debug, and document code. You're responsible for fitting it into the team's workflow: custom slash commands and skills, a sensible CLAUDE.md setup, and knowing when a task calls for plan mode and when direct execution is enough.
Context Management & Reliability
Two hours into exploring an unfamiliar payments codebase, Claude Code's answers start to drift. It describes "the typical repository pattern" instead of the specific LedgerRepository and SettlementBatcher classes it examined earlier, and it gives an answer that conflicts with one from an hour ago. What practice best counteracts this?
Show answer and explanation
Answer: C — Have Claude maintain a scratchpad file recording key findings, such as class names, responsibilities, and data flows, and consult that file when answering later questions.
Answers that slip into "typical patterns" and contradictions are signs of context degradation in a long session. A scratchpad file of key findings persists them across context boundaries and gives later questions a reliable reference (task 5.4). Restarting and repeating the exploration (a) throws away the work and will degrade the same way. Rereading everything (b) floods context and speeds up the degradation. Extended thinking (d) doesn't bring back details that have degraded in context.
More from this scenario
- Your team is deciding which upcoming tasks should start in plan mode and which can go…
- You ask Claude Code to convert legacy log lines into a new structured format. Despite…
- Before a large refactor, Claude Code must survey around 300 files to catalog every place…
- During a long debugging session, the context has filled with verbose test-runner output…
- You're adding comprehensive tests to a legacy service in phases. In phase one, a…