Skip to content

task.md

Latest commit

 

History

History
70 lines (51 loc) · 3.6 KB

task.md

File metadata and controls

70 lines (51 loc) · 3.6 KB
Raw

Skill /task

Defines a task for AI implementation from a ticket or free-form description. Explores the codebase via sub-agents, analyzes the architecture, asks clarifying questions, and produces a task file with context, requirements, and constraints. The output is a task file that /sp:plan turns into an implementation plan.

Input

$ARGUMENTS — ticket URL (GitHub Issues, YouTrack, Jira) and/or a free-form task description.

/sp:task https://github.com/owner/repo/issues/86
/sp:task add dark theme to settings

Phases

Phase Name What happens
1 Parse Fetch ticket contents, extract materials, form task-slug and TICKET_ID
2 Investigate Sub-agents explore the codebase (task-explorer) and architecture (task-architect). Strictly sequential
3 Synthesize Apply 5 dimensions to findings, determine type (frontend/general), form clarifying questions
4 Write Write the task file from examples (simple/complex), copyedit via sub-agent (Elements of Style)
5 Commit Auto-commit the artifact: TICKET docs(SLUG): add task definition
6 Complete Completion loop: run /sp:plan (recommended) / review via plannotator / finish

Output

File docs/ai/<slug>/<slug>-task.md with the following structure:

  • Header — Slug, Ticket, Complexity (trivial/simple/medium/complex), Type (frontend/general)
  • Task — one sentence: what exactly to do
  • Context — 4 subsections: area architecture, files to change, patterns to reuse, tests
  • Requirements — concrete, verifiable requirements
  • Constraints — what not to change, approaches to avoid, risks from findings
  • Verification — commands with expected results, edge cases
  • Materials — links and file paths from input

5 Synthesize dimensions

Dimension Check question
Intent Clarity Two different developers read the Task and arrive at the same implementation?
Scope Boundaries What's definitely in scope, and what's out?
Context Anchoring The implementer knows exactly what's needed — no more, no less?
Acceptance Criteria The implementer can verify each criterion without running the whole project?
Reuse Opportunities Which existing components can be reused?

Sub-agents

Agent Model Role
task-explorer sonnet Codebase exploration: files, patterns, tests, dependencies
task-architect sonnet Architecture analysis: patterns, integration points, risks, trade-offs

Example

/sp:task https://github.com/org/repo/issues/112

Result: docs/ai/112-password-reset-email/112-password-reset-email-task.md

Connections

/sp:task → /sp:plan → /sp:do → /sp:review

/task defines the task. /plan builds the implementation plan. /do executes the plan.