ticket-creator（票证创建者）

Ticket Creator

Create well-structured Linear tickets optimized for AI agent (Claude Code) implementation. Follow a pattern-first, template-driven workflow that ensures the implementing agent can plan and execute without making open-ended architectural decisions. Consult references/guidelines.md for detailed quality criteria and writing guidelines.

Core Principle: Pattern-First

The strongest predictor of ticket success is whether the implementing agent can copy-and-adapt from an existing pattern. Every ticket MUST reference an existing code pattern or explicitly flag it as a first-time pattern requiring extra specification.

AskUserQuestion Usage

Only invoke AskUserQuestion when the answer cannot be deduced from the current conversation context, the codebase, or prior responses. If the information is available or can be inferred, proceed without asking. This applies to all stages — classification, deduplication, information gathering, anti-pattern resolution, splitting, and confirmation.

Workflow

Use TaskCreate at the start to plan the workflow steps, and TaskUpdate to mark each step as you complete it. This keeps progress visible and survives /compact.

1. Gather Context & Classify

Extract as much information as possible from the current conversation before asking any questions. If work has been discussed, use that context to pre-fill ticket details.

Classify the ticket:

• Normal (feature, improvement, refactor) → references/template-normal.md
• Bug (error, broken behavior) → references/template-bug.md

Identify the domain from context or keywords:

If classification is ambiguous, use AskUserQuestion to clarify.

2. Check for Duplicates

Use Task with subagent_type general-purpose to search Linear for existing similar tickets. Pass the ticket description, key terms, and domain as context, along with the instructions from references/duplicate-checker.md. The subagent handles all list_issues calls and returns a concise summary.

If the subagent surfaces matching tickets, present them to the user and ask whether to proceed, update an existing ticket, or link as related.

3. Load Template & Gather Information

Load the appropriate template from references/. Gather missing required information conversationally, 1-2 questions at a time, in this priority order:

1. What are we building/fixing and why? What's the scope?
2. Is there an existing pattern to follow? (most critical)
3. Technical details — types, contracts, file paths
4. Side effects, integration points, permissions
5. Acceptance criteria and verification hints

For bug tickets: if a Sentry issue ID or URL is available, use Task with subagent_type general-purpose to fetch the Sentry error details. Instruct the sub-agent to return only the fields needed for the ticket: error message, affected file/line, root cause (if determinable), frequency/user impact, and any relevant breadcrumbs. See references/template-bug.md for the full field list. This prevents verbose Sentry payloads (stack traces, raw breadcrumbs) from flooding the main context.

4. Pattern Discovery

Use Task with subagent_type Explore to search the codebase for:

• Existing implementations to reference as patterns
• Relevant Claude Code skills (scan .claude/ directories for SKILL.md files)

Include discovered patterns in the ticket's "Existing pattern to follow" field. Include discovered skills in the Implementation Toolkit section using the format: skills: [skill-name-1, skill-name-2]

5. Quality Gate

Run all anti-pattern checks from references/anti-patterns.md. Surface any warnings and resolve them with the user before proceeding.

Evaluate whether the ticket should be split (see Ticket Splitting below).

6. Draft & Confirm

Compile the ticket using the loaded template. Present a concise preview — title, summary, key fields, acceptance criteria — not the full ticket body.

Ask the user to confirm or request changes before creating.

7. Create in Linear

Create the ticket using create_issue with: team ("Sylla Tech"), title, description, priority. Only set assignee if the user specifies one.

If dependencies exist, use blockedBy or blocks fields. Link related tickets found during deduplication using relatedTo.

Share the ticket link with the user after creation.

Ticket Splitting

When splitting, use TaskCreate with blockedBy to model the dependency order between split tickets (e.g., backend before UI). Use TaskList to review progress and TaskUpdate to track each ticket through stages 3-7.

Multi-Category Split

If work spans multiple domains (e.g., UI + Backend + Data-Engine), create separate linked tickets — one per domain. Run stages 3-7 for each, carrying forward shared context (priority, description, related tickets). Cross-link tickets using relatedTo.

File-Based Split

If a ticket would modify more than 8 files, suggest splitting into smaller, focused tickets. Each split ticket should be independently implementable.

Title Conventions

Reference Files

Example Tickets

Consult these real ticket examples to calibrate structure and detail level:

Quality Test

Before creating, verify: *"Could the implementing agent create a complete plan from this ticket without making any open-ended architectural decisions?"*

If the answer is no, identify what's ambiguous and ask for it. See references/guidelines.md for the full quality checklist.