Damon-Stewart-1 clear-prep
Generate a self-contained handover prompt for a cleared Claude Code instance. No logging, no Productive, no sync-context, no git diff. Pure relay.
Step 0: Check for Viable Session
If the conversation has fewer than ~3 substantive turns AND no files were mentioned/edited AND no commands were run, output: "No active session to relay. Describe the project or task first, then run /clear-prep." Stop. A planning-only or research-only session with substantive decisions IS viable.
Step 1: Read MEMORY.md
Read ~/.claude/projects/$(basename $HOME)/memory/MEMORY.md (adjust path to match your setup). Store its contents for the delta check in Step 6. If unreadable: skip the delta check and add a high-priority entry to ## Warnings.
Step 2: Session Analysis
Scan conversation history to populate: Decision Log (explicit decisions and rejections with reasons), Session State (goal, completed vs. in progress, working directory), Blocked On (unresolved blockers/dependencies), Warnings (gotchas, tool failures, session-specific instructions), Environment State (export, source, alias, conda activate, nvm use, or similar ephemeral shell commands).
Step 3: Identify Active Scope
Check for /scope invocations, agent file references, or the working directory of most recently referenced files. List all projects/directories touched. If undeterminable, note "No active scope detected."
Step 4: File Snapshot Fingerprinting
Read references/fingerprint-patterns.md for the exact commands, output format, caps, and edge case handling. Fingerprint all relevant files.
Step 5: Dispatch Job Ingestion
ls -t ~/.claude/jobs/*.md 2>/dev/null
Filter to files modified within the last 2 hours (stat -f %m vs date -v-2H +%s). For complete jobs: extract results (up to 10 lines). For running jobs: note ID, task, status, and add a warning not to modify affected files.
Step 6: Memory Delta Check
Compare session context against MEMORY.md from Step 1. Include anything that meets ALL criteria: falls into a concrete category (user preference, project constraint, architectural decision, deployment target, API credential location, team/people context, or non-obvious workflow), is NOT already in MEMORY.md, and would be useful weeks from now. If nothing qualifies, omit the section.
Step 6b: Plugin Debrief
Scan this session for gaps in the plugin system. Look for three things only:
- A hook that should exist or a current hook that didn't catch something it should have
- A workflow that repeated enough to warrant a new skill or an addition to an existing one
- Something learned about a client that isn't in their agent
.mdfile
For each gap found, propose a specific diff: which plugin file (hooks/, skills/, agents/), what to add or change, and why. If nothing qualifies, say "No plugin gaps found." Do not propose CLAUDE.md edits or memory file writes here -- that's Step 6.
Step 7: Sanitization Pass
Read references/section-rules.md for the complete list of secret patterns to scan for. Replace any match with [REDACTED -- see Keychain or passwords app]. Never output a real secret.
Step 8: Assemble Output
Read references/handover-template.md for the exact output structure. Do not ask for approval. Output a single fenced code block. Read references/section-rules.md for which sections are required vs. optional.
Step 9: Copy to Clipboard
After outputting the fenced handover block, immediately copy its contents (the text inside the fences, not the fences themselves) to the macOS clipboard:
cat <<'CLEAR_PREP_EOF' | pbcopy
<the full handover content here, exactly as it appeared in the fenced block>
CLEAR_PREP_EOF
Use a heredoc with a unique sentinel (CLEAR_PREP_EOF) so embedded backticks, dollar signs, and quotes pass through untouched. Do not use echo or printf (they mangle special chars). After running, output one final line confirming: Handover copied to clipboard. Paste with Cmd+V into the cleared instance.
If pbcopy fails (e.g., no display, headless), say so and skip silently rather than blocking.