recall
This skill should be used when the user asks to "recall related knowledge", "search previous solutions", "check what we learned before", "look up past decisions", or when starting a new task and wanti
SKILL.mdskills/recall/SKILL.md
GitHubSkill: recall
Searches the project knowledge base (.claude/knowledge/) to surface relevant prior learning before starting a new task. Recall results reduce redundant exploration and prevent repeating solved problems.
Knowledge base structure
.claude/knowledge/
├── solutions/ bug-fix captures: what broke, why, how fixed
├── adrs/ architectural decision records (MADR format)
└── patterns/ reusable implementation patterns with code examples
Search procedure
1. Extract search terms from the issue
Read the issue title and body. Extract:
- Domain keywords (e.g., "booking", "appointment", "payment")
- Technical keywords (e.g., "Flyway", "Server Component", "Tailwind")
- Error messages or symptoms (if it's a bug fix)
2. Search each knowledge directory
# Semantic search via grep across all knowledge files
grep -ril "<keyword>" .claude/knowledge/solutions/ 2>/dev/null
grep -ril "<keyword>" .claude/knowledge/patterns/ 2>/dev/null
grep -ril "<keyword>" .claude/knowledge/adrs/ 2>/dev/null
Read the full content of any file that matches.
3. Produce recall summary
Write a "Prior Knowledge" section at the top of exploration.md (or as a standalone note if exploration hasn't started yet):
## Prior Knowledge (from recall)
### Related ADRs
- [ADR-0003](docs/adr/0003-...) — relevant decision + brief summary
### Related Patterns
- `patterns/filename.md` — one-line description of relevance
### Related Solutions
- `solutions/bug-slug.md` — one-line description of relevance
### How this affects the task
<!-- Brief note on how prior knowledge should shape exploration or design -->
Graceful degradation
If .claude/knowledge/ does not exist or is empty, log "Knowledge base empty — proceeding without recall" and continue. Do not block the workflow.
When recall is sufficient to skip exploration
After recall, assess whether the retrieved knowledge fully covers the task scope. If yes, note in state.yaml:
config_used:
skipped_phases: [explore]
recall_was_sufficient: true
And proceed directly to clarification/design with the recalled knowledge as context.