# Ask The Codebase Questions Before Editing It ## Claim A new user of an agentic coding tool builds calibration faster by interrogating the codebase first than by jumping straight to edits. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for onboarding to an unfamiliar codebase with an agentic tool that can read files, search, and inspect Git history without side effects. - The mechanism is calibration, not correctness: Q&A teaches the user what the tool can answer confidently versus what needs hand-holding, before any edit is at stake. - Distinct from the feedback-loop claim (agent output quality scales with the check you give it) — this is about the human's ramp-up sequence, not the agent's self-correction loop. - Doesn't establish the onboarding-time reduction is generalizable; the 2–3-weeks-to-2–3-days figure comes from one well-documented internal codebase and is a single, unverified data point. ## Evidence Boris Cherny, describing how Anthropic teaches Claude Code to new hires: "The thing I recommend above everything else is starting with codebase Q&A." He reports the practice replaces manually spelunking code and interrupting other engineers, and that Claude answers at a depth beyond text search — reading Git history to explain who wrote a function, when, and which issues drove it, unprompted. ## Links - [[AI & the Future of Knowledge Work MOC]] — how calibration with an unfamiliar system is built faster by interrogation than by action - [[3. Resources/Permanent Notes/Agent Output Quality Scales With the Feedback Loop Not the Prompt]] — contrasts; that note is about the agent's self-correction loop during editing, this one is about the human's ramp-up sequence before editing starts - [[3. Resources/Permanent Notes/Time-to-Hello-World Is the Critical Developer Experience Metric]] — applies-in; Q&A-first is a concrete tactic for compressing the time between "get the tool" and "trust the tool" - [[3. Resources/MOCs/DevRel MOC]] — applies-in; a reusable onboarding pattern for any developer-adoption program built around an agentic tool ## Tests / falsifiers - Compare edit-acceptance rate and time-to-first-successful-PR for new users onboarded Q&A-first versus edit-first on the same codebase — if there's no meaningful gap, the claim overstates Q&A's calibration value. - Find a codebase where Q&A produces confidently wrong answers (stale docs, sparse Git history) — this would bound the claim to well-documented repos rather than all repos. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [[3. Resources/Readwise/Articles/How to Use Claude AI for Coding Master Claude Code in 30 Minutes]] — Boris Cherny (Anthropic), on Claude Code onboarding practice.