--- name: weekly-skill-discovery description: Scans Jira, Confluence, Slack, Gmail, and Google Calendar for repeatable work patterns, proposes 3–5 skill candidates, notifies via Slack with a link back to this session, then builds any skill you approve. --- **Before running any phase:** read the CONFIG block in the HTML comment at the top of this file and resolve the three variables — `USER_NAME`, `USER_TIMEZONE`, `SCHEDULED_PATH` — to their actual values. Use those values everywhere these placeholders appear below. You are running the weekly skill discovery for {USER_NAME}. Timezone: {USER_TIMEZONE}. Today's date is available from context. Your job: find repeatable manual work that should be automated as a Claude skill. Run Phase 1 automatically on schedule. Stop after Phase 2 and wait for {USER_NAME} to choose a candidate to build. --- ## Phase 0 — Identity check (runs every time) Before doing anything else: 1. Read this SKILL.md file from disk and extract the CONFIG values. 2. Use the Atlassian MCP tool (tool name contains `atlassianUserInfo`) to get the current user's display name and account ID. Store both — the display name is used for the identity check, the account ID is used in Phase 1 Jira and Confluence queries. 3. Compare the current user's display name to `{USER_NAME}` from CONFIG. **If they match** → skip the rest of Phase 0 and proceed directly to Phase 1. **If they don't match** (this skill was copied from someone else, or hasn't been configured yet), enter setup mode: 1. Greet the user: > "Welcome! This skill hasn't been set up for you yet. I'll ask a few quick questions — this only happens once." 2. Ask the following questions one at a time, waiting for each answer: - "What's your full name?" - "What's your timezone? (e.g. Europe/Prague, America/New_York, Asia/Singapore)" - "What's the full path to the folder where your Claude skills live? (e.g. /Users/yourname/Documents/Claude/Scheduled)" 3. Use the Edit tool to update this SKILL.md file — replace the three CONFIG values with the user's answers: - `USER_NAME` → their full name - `USER_TIMEZONE` → their timezone - `SCHEDULED_PATH` → their skills folder path 4. Re-read this SKILL.md file from disk and re-resolve all three CONFIG variables to their newly written values. Use these updated values for all remaining phases. 5. Confirm: > "✅ All set! Running your first discovery scan now..." 6. Continue directly to Phase 1. --- ## Phase 1 — Discovery ### Step 1 — Build exclusion list Read all files matching `{SCHEDULED_PATH}/*/SKILL.md`. Extract the `name` and `description` from each frontmatter block. These skills already exist — do not propose them again. ### Step 2 — Scan all five sources (run in parallel) Use the Atlassian account ID retrieved in Phase 0 for all Jira and Confluence queries below. **Slack** — use the Slack MCP tool (tool name contains `slack_search_public_and_private`) - Query 1: `from:me after:YYYYMMDD` (7 days ago, ISO date) — your own sent messages - Query 2: `is:saved` — your saved/bookmarked messages - Look for: repeated phrases or questions you keep asking, messages that required manual research or synthesis, recurring topics you flag for later **Google Calendar** — use the Google Calendar MCP tool (tool name contains `list_events`) on the `primary` calendar - Fetch events from 7 days ago through 7 days ahead - Look for: recurring meetings that have no existing prep or debrief skill, meetings that always result in similar manual follow-up actions **Gmail** — use the Gmail MCP tool (tool name contains `search_threads`) - Query: `in:sent after:YYYYMMDD` (7 days ago) - Look for: emails where you typed similar content multiple times, recurring threads that required the same type of research or response structure **Jira** — use the Atlassian MCP tool (tool name contains `searchJiraIssuesUsingJql`) - JQL: `assignee = "[your account ID]" AND updated >= -7d ORDER BY updated DESC` - Also try: `reporter = "[your account ID]" AND created >= -7d` - Look for: issue types you create repeatedly, recurring investigation or documentation tickets, any manual status-update patterns **Confluence** — use the Atlassian MCP tool (tool name contains `searchConfluenceUsingCql`) - CQL: `contributor = "[your account ID]" AND lastModified >= now("-7d")` - Look for: pages you update on a regular cadence, recurring documentation formats, pages that always require the same type of content refresh ### Step 3 — Synthesize and rank From all sources, identify patterns that: - Appear more than once OR across multiple tools - Take meaningful time to do manually (>5 min each occurrence) - Are not already covered by existing skills (from Step 1) - Would be feasible to automate as a Claude skill Rank by estimated weekly time cost. Select the top 3–5 candidates. For each candidate, prepare: - A short name - The concrete pattern spotted (with a specific example from the data) - What the skill would do - Estimated time saved per week - Which sources surfaced it --- ## Phase 2 — Notify and present ### Step 4 — Get session URL (best effort) Try to get the URL of the current Claude session using the browser MCP tool (tool name contains `tabs_context_mcp`). If successful, use that URL in the Slack message. If unavailable or it fails, use the fallback text below. ### Step 5 — Send Slack notification Look up the current user's Slack ID using the Slack MCP tool (tool name contains `slack_search_users`), searching by `{USER_NAME}`. Use the returned user ID to send a **direct message to yourself** via the Slack MCP tool (tool name contains `slack_send_message`), passing the user ID as the channel. This works universally — no channel config required. Slack message format (mrkdwn): ``` 🔍 *Weekly Skill Discovery — [Weekday, Date]* ―――――――――――――――――――――― Found *[N]* repeatable patterns that could become skills: *1. [Skill Name]* — [one-sentence description] *2. [Skill Name]* — [one-sentence description] [repeat] ―――――――――――――――――――――― 👉 <[session URL]|Continue in Claude> to review details and build one. ``` If session URL was unavailable, replace the last line with: `👉 Open Claude and continue the *Weekly Skill Discovery — [Date]* session to review and build.` **Formatting rules:** - Use `*bold*` and `_italic_` only - Use `――――――――――――――――――――――` as dividers - Use `` for all hyperlinks ### Step 6 — Present full candidates in the Claude conversation Display the full candidate list in this format: --- ## 🔍 Weekly Skill Discovery — [Date] Here are the skill candidates I found from the past 7 days: --- **1. [Skill Name]** 📋 *Pattern:* [concrete example of the repeating action, e.g. "You wrote Jira status update comments in 4 separate tickets using the same structure"] 💡 *What it would do:* [clear description of the skill's job] ⏱️ *Est. time saved:* ~[X] min/week 📡 *Sources:* [Jira / Slack / Calendar / etc.] **2. [Skill Name]** [same structure] [repeat for all candidates] --- Reply **"build #[N]"** (or use the skill name) and I'll create it. To skip: close this session or say "skip". --- > STOP here. Wait for {USER_NAME}'s reply before proceeding. --- ## Phase 3 — Build (triggered by user reply) When {USER_NAME} selects a candidate: ### Step 7 — Clarify if needed Ask at most 2 clarifying questions, only if the answer isn't obvious from context: - Schedule: when should this run? (if it's a recurring skill) - Output: where should results go? (Slack, Gmail, saved to file, etc.) If the answer is obvious, skip and proceed directly. ### Step 8 — Write the SKILL.md Create the file at `{SCHEDULED_PATH}/[skill-name]/SKILL.md`. Follow the same structure as existing skills: - YAML frontmatter with `name` and `description` - Address {USER_NAME} by name at the top - Numbered steps with clear, executable instructions - Use the same Slack formatting conventions if the skill sends output there (use `――――――――――――――――――――――` as dividers, `*bold*` and `_italic_` only) - End with: use the Slack MCP tool (tool name contains `slack_send_message`) to send a direct message to the user (look up their Slack user ID by name first) if the skill outputs to Slack ### Step 9 — Schedule if recurring If the skill should run on a schedule, use `mcp__scheduled-tasks__create_scheduled_task` with the appropriate cron expression and the path to the new SKILL.md. ### Step 10 — Confirm Reply with: ``` ✅ Skill created: `{SCHEDULED_PATH}/[skill-name]/SKILL.md` [If scheduled:] Scheduled for [cadence, e.g. "every Tuesday at 8:00 AM CET"]. [If not scheduled:] Run it manually anytime from Claude Code. ```