| name | prep-for-task |
| description | Bootstrap context for a new session on the current project. Run this BEFORE starting any implementation work. |
| disable-model-invocation | true |
Prep for Task
Bootstrap context for a new session on the current project. Run this BEFORE starting any implementation work.
Project instructions
cat CLAUDE.md 2>/dev/null || echo "No CLAUDE.md found"
Auto memory
PROJECT_DIR=$(pwd | sed 's|/|-|g; s|^-||'); cat ~/.claude/projects/"$PROJECT_DIR"/memory/MEMORY.md 2>/dev/null || echo "No MEMORY.md found"
Instructions
The project instructions and auto memory above have been pre-loaded. Now launch these exploration subagents in parallel to gain comprehensive project context:
Agent 1: Project Overview
Read these files and summarize the project state:
README.md — services, configuration, credentials- Run
git log --oneline -20 for recent commit history
- Run
git status for working tree state
- Check
~/.claude/plans/*.md — read any plan files modified in the last 7 days
Agent 2: Codebase Structure
Explore the full directory tree and map out the project structure. Adapt based on what you find:
- Top-level directories and their purpose
- Key entry points (main files, index files, server/app bootstrapping)
- Where the core business logic lives
- Any notable config files, build tooling, or infra files (Dockerfile, CI configs, etc.)
Agent 3: Journal History
Read ALL files in journals/ sorted by name (ascending), if the directory exists. For each session entry, extract signal and condense it. Return a structured summary to the main instance — do NOT return raw journal content:
## Journal Summary
### Recently Built / Fixed
<bullet per notable feature or fix, most recent first>
### Technical Gotchas
<bullet per confirmed gotcha or non-obvious behaviour>
### Architectural Decisions
<bullet per decision with rationale>
### Deferred / Known Issues
<bullet per thing explicitly deferred or known to be broken>
Omit anything already superseded by a later journal entry. Be terse — this summary must fit in ~30 lines. If no journals/ directory exists, skip this agent.
After Exploration
Summarize your findings to the user in this format:
## Project State
- Recent features: <list>
- Known issues: <list>
## Recent Changes
<last 5 commits, one line each>
## Key Gotchas to Remember
<top 5 technical pitfalls from journals and memory>
## Suggested Next Task
Based on open issues, in-progress branches, plan files, and deferred items from journals — suggest the single most logical next task and ask if that's the priority or if something else takes precedence.
Rules
- Use subagents for ALL exploration — maximize parallelism
- Do NOT write or modify any files during prep
- Do NOT start implementation until the user gives you a task
