b-open-io/prompts/remind

Remind

Recall and search previous Claude Code conversation sessions to find past discussions, decisions, solutions, and context.

How It Works

This skill searches conversation history using two backends:

1. Scribe DB (preferred) — SQLite FTS5 full-text index at ~/.scribe/scribe.db with 141K+ indexed messages across all AI coding tool sessions. BM25-ranked results, grouped by session.
2. JSONL fallback — Direct search through ~/.claude/projects/ JSONL conversation files when Scribe isn't available.

The search script at scripts/search.py handles both backends automatically.

When to Use

• User asks about something from a previous conversation
• User wants to continue or reference prior work
• User asks "when did we..." or "remember when..."
• User needs to find a specific past decision, approach, or code snippet
• User wants to see what they worked on for a project/timeframe

Workflow

Step 1: Search

Run the search script with the user's query:

python3 "${CLAUDE_SKILL_DIR}/scripts/search.py" "<query>" [options]

Options:

Results are ranked by a blend of BM25 relevance and recency. Recent conversations get a boost that decays exponentially (30-day half-life). Use --no-recency to disable.

Crafting good queries:

• Use 2-3 specific keywords, not full sentences
• Technical terms work best: "stripe webhook", "bap identity", "deploy railway"
• If too many results, add --project or --recent filters
• If too few results, try broader terms or --full for deeper search

Step 2: Present Results

Summarize what you found in a clear format:

• Which sessions matched and when they occurred
• What project each session was in
• Key snippets showing the relevant context
• Offer to dive deeper into specific sessions

Step 3: Deep Dive (if needed)

If the user wants more detail from a specific session:

python3 "${CLAUDE_SKILL_DIR}/scripts/search.py" --session <session-id> --json

This returns the full conversation transcript. Summarize the relevant portions — don't dump the whole thing.

Step 4: Read Raw JSONL (last resort)

If neither search backend returns results, you can read the JSONL files directly:

# List all project directories
ls ~/.claude/projects/
 
# List sessions in a specific project
ls ~/.claude/projects/-Users-satchmo-code-myapp/*.jsonl
 
# The JSONL format has one JSON object per line with these types:
#   user    — message.content is the user's input
#   assistant — message.content is Claude's response (array of text/tool_use blocks)
#   progress — tool execution progress
#   system   — system messages

Session index files (where available) provide quick metadata:

# Check for pre-built session index
cat ~/.claude/projects/<project-dir>/sessions-index.json
# Contains: sessionId, firstPrompt, summary, created, modified, gitBranch, messageCount

Tips

• Scribe reindex: If Scribe DB exists but results seem stale, the user can rescan with cd ~/code/scribe && bun run packages/cli/src/index.ts scan chat --provider claude
• Cross-project: Searches span all projects by default. Use --project to narrow.
• Session IDs: These are UUIDs like 69ad65a8-7499-4fd1-9bae-a3f0fcbb11ed. When presenting results, include session IDs so the user can ask for deep dives.
• Privacy: Conversation data stays local. Never output sensitive content (keys, tokens) from search results.