matteotitta /discover — Weekly Skill & Automation Discovery
Proactively identifies repeatable work patterns across your tools and recommends what to automate.
For full process (subcommands, 6 phases, scoring, persistence) → the premium reference.
Claude Code Triggers
Invoke this skill when user says:
- "/discover"
- "What should I automate?"
- "Find repeatable patterns"
- "What skills should I build?"
- "Discovery scan"
- "What am I doing repeatedly?"
Do NOT invoke when:
- User wants to run a specific skill (run that skill directly)
- User wants to create a specific skill (use
/skill-creator) - User wants to audit workspace health (use
/audit)
Subcommands quick reference
| Command | Action | Detail |
|---|---|---|
/discover |
Run full discovery scan (default) | 6-phase scan |
/discover status |
Pattern DB stats + top 5 BUILD candidates | SQL summary |
/discover dismiss <name> |
Mark pattern dismissed | Excluded from future reports |
/discover built <name> <path> |
Mark pattern as built | Links to artifact |
/discover history |
Pattern evolution over time | Top 20 by recency |
Parse $ARGUMENTS to determine subcommand. Default (no args) = full scan.
For full SQL queries per subcommand → the premium reference.
Core decisions
Scoring framework (8 dimensions, 0 or 1 each)
| # | Dimension | Question |
|---|---|---|
| 1 | frequency | Will this run 5+ times per month? |
| 2 | time_savings | Saves 30+ minutes vs. doing it manually? |
| 3 | quality | Automation produces better/more consistent output? |
| 4 | context_dep | Inherits context from other skills? |
| 5 | distinct | Fills a gap no existing skill covers? |
| 6 | reuse | Works for multiple clients with minimal adaptation? |
| 7 | measurable | Clear criteria for good vs. bad output? |
| 8 | recurrence | Pattern appeared in 3+ separate weekly scans? |
Verdicts
- Score 6-8 = BUILD — create the automation
- Score 4-5 = DEFER — watch for another week
- Score 0-3 = KILL — not worth automating
Automation type selector
| Type | When to recommend |
|---|---|
| skill | Multi-step process with structured output, repeatable across clients |
| hook | Automatic trigger before/after a specific tool use |
| scheduled-agent | Time-based recurring task |
| prompt-chain | Sequence of skills that always run together |
| mcp-workflow | Cross-tool data movement or enrichment |
Auto-promotion rule
If a pattern's occurrence_count >= 3 in the database and it was previously DEFER, set recurrence: 1 which may push to BUILD.
For full classification guide and per-pattern JSON schema → the premium reference.
Data Sources (6)
- Session recall DB (last 7 days)
- Linear (last 7 days)
- Slack (last 7 days)
- Gmail (last 7 days)
- Granola (last 7 days)
- Google Calendar (last 14 days, catches biweekly patterns)
Full SQL queries + MCP calls per source live in the premium reference.
Anti-Hallucination Guardrails
- Verify against existing catalog. Don't mark "distinct" without checking
.claude/skills/meta/catalog/skill-catalog/SKILL.md. - Use real evidence. Each pattern needs source-cited snippets — no invented examples.
- Update existing patterns. Set
is_existing: trueif the pattern already exists in the DB; don't create duplicates. - MCP failures don't block scan. Log the failure, continue with available sources.
Notes
- The session recall DB is the richest source. Linear, Slack, Gmail, Granola, and Calendar add cross-tool visibility.
- Pattern hash uses first 16 chars of SHA-256 — sufficient for personal scale.
- Dismissed patterns stay in DB for history but are excluded from future reports.
- Scheduled-agent runs should complete in 5-10 min. Prioritize sessions + Slack if MCPs are slow.