Extract Tasks from Conversation

Scan the current conversation for task-like content and propose adding untracked items to WritBase. This is a review-and-confirm workflow — never create tasks without explicit user approval.

Prohibitions

• Do not auto-invoke this skill. Only run when the user types /extract-tasks or explicitly asks to extract tasks.
• Do not update existing tasks. This skill creates new tasks only.
• Do not retry on MCP connection failure. Show error and stop.

Workflow

Step 1: Validate WritBase Connection

Call writbase:info to verify the MCP server is connected.

• If the call fails or returns an error: Output the following and stop:

WritBase MCP is not connected. Run npx writbase status to check your setup.

• If connected: Extract the list of available projects and departments from the response. You will use these to constrain valid scopes when creating tasks.

Step 2: Locate the Session Transcript

The current conversation is stored as a JSONL file. Identify it:

1. Determine the Claude projects directory for the current workspace. The path pattern is ~/.claude/projects/<mangled-cwd>/ where <mangled-cwd> is the absolute working directory with / replaced by - (leading - preserved). Run pwd to get the current directory, mangle it, and list matching *.jsonl files sorted by modification time (most recent first).
2. The most recently modified file is almost certainly this session. Use it.
3. If multiple files were modified within the last 2 minutes, ask the user to confirm which session to scan.

Step 3: Parse Conversation Content

Session JSONL files contain many record types. Only scan records relevant to the conversation:

• Include: Records where type is "user" or "assistant" — these are the actual conversation turns.
• Skip: Records where type is "progress", "file-history-snapshot", "system", or anything else. These are hook events, tool metadata, and internal bookkeeping (~90% of records).

For included records, extract text from message.content. This field may be:

• A plain string
• An array of content blocks (extract text from blocks where type is "text")

Use Grep on the JSONL file to find lines containing task-signal keywords, then Read surrounding lines for context. Task signals include:

• Action verbs with objects: "need to", "should", "must", "have to", "TODO", "FIXME", "let's add", "we need"
• Deadline language: "by Friday", "before release", "this sprint", "due date"
• Assignment language: "I'll handle", "you should", "assigned to", "owner"
• Commitment language: "I'll do", "let me", "will implement", "going to"
• Problem statements: "bug in", "broken", "failing", "doesn't work", "issue with"

Apply neighborhood context: read 3 messages before and 2 after each candidate to resolve vague references (e.g., "that thing" refers to what was discussed 2 messages earlier).

Step 4: Extract and Present Candidates

For each task-like item found, prepare:

• Description: A clear, actionable task description (minimum 3 characters, aim for a complete sentence)
• Priority: Inferred from context — critical (blocking/urgent), high (important/soon), medium (default), low (nice-to-have/someday)
• Source quote: The exact user or assistant message that triggered extraction (truncated to ~100 chars if long)
• Project: Pre-fill from writbase:info if only one project is available. If multiple projects exist, ask the user to choose.
• Department: Pre-fill from writbase:info if only one department is available for the selected project. If multiple, ask the user to choose. If none are required, omit.

If no task candidates are found, output the following and stop:

No untracked tasks found in this conversation.

Present candidates as a numbered list:

Found 3 potential tasks:

1. [high] Add rate limiting to the /api/upload endpoint
   Source: "we need to add rate limiting to upload before the release"
   Project: my-project | Department: engineering

2. [medium] Investigate flaky test in auth module
   Source: "the auth test has been failing intermittently"
   Project: my-project | Department: engineering

3. [low] Update README with new CLI commands
   Source: "should probably update the docs at some point"
   Project: my-project | Department: (none)

Ask the user to review: for each task, they can approve, edit (change description, priority, project, or department), or skip.

Step 4b: Check for Duplicates

For each candidate, call writbase:get_tasks with search set to 2-3 key words from the description. If a matching task already exists, mark the candidate as (possible duplicate of `<task_id>`) in the presentation. The user decides whether to skip or create anyway.

Step 5: Create Approved Tasks

For each approved task, call writbase:add_task with the confirmed parameters.

Error recovery — do not fail terminally on these errors. Instead, return to the user for correction:

For other errors (rate_limited, unauthorized_agent_key, inactive_agent_key), report the error and stop — these indicate infrastructure problems the user must fix outside this workflow.

Step 6: Confirm Results

After all approved tasks are created, output a summary:

Created 2 tasks in WritBase:
- [high] Add rate limiting to the /api/upload endpoint (id: abc-123)
- [medium] Investigate flaky test in auth module (id: def-456)

Skipped 1 item.