Codebase Understanding (read-first, low-token)
Goal: locate the relevant code in the fewest tokens. NEVER scan the whole repo to answer a question the context store already answers.
Order of reading (stop as soon as you have what you need)
.claude/context/SERVICE-MAP.md— which service/db/queue/API owns this area..claude/context/FINDINGS.md— is this already known/solved?.claude/context/DECISIONS.md— is the current behavior intentional?- The ONE relevant
<service>/CLAUDE.md— contracts, key files, gotchas. - The shared-models
CLAUDE.md— only when tracing a routing key / RPC across services. - Source — targeted, by
file:linefrom the above, via Grep/Read. Not whole-file dumps unless necessary.
Cheap search patterns
- Routing key / RPC:
grep -rn "<routing.key>" <service>/then open the dispatcher. - A symbol:
grep -rn "MethodName\|TypeName" <service>/ --exclude-dir={obj,bin}. - A flow: start at the SERVICE-MAP "In → Out" column, follow the routing key to the next consumer.
Output
A short "where it lives" answer: service, file:line, and the one-line role. Do not paste large code blocks — cite locations.
Don't
- Don't re-derive the architecture; it's in SERVICE-MAP/CLAUDE.md.
- Don't read every service. Read the one(s) the map points to.