SuperAPI AI Directory

plan

Design & creation

View repository

Updated 10-15 21:49

About this skill

Plan

Turn ideas into complete Backlog specs in Linear.

Read the conventions first

Read references/conventions.md completely before any write. It defines preflight, Linear MCP tools, the spec body template, hierarchy, dependencies, and state rules.

When to use

  • Scoping, designing, or specifying work.
  • No Todo issue exists for the request.
  • Revising a Backlog spec.

A new build request starts here even when the user says "build". For tiny, clearly bounded edits such as a copy change, do not force the full Plan workflow and do not open an issue. State the assumption and ask whether the user wants the full process.

Plan workflow

Read references/plan.md completely and follow it.

Hard gates:

  • Do not draft a spec issue before the user has answered an alignment question or approved a recommended direction.
  • Every spec issue must include an exact ## Acceptance criteria section with observable pass/fail checkboxes.
  • Create and revise specs with Linear MCP. End at human approval into Todo. Do not write Todo yourself.
  • After Plan, the only workflow you invoke is Build. Do not implement product code here.

Requirements

  • Host Linear MCP with get_issue, save_issue, save_comment, list_comments, list_issues, and list_issue_statuses.
  • gh authenticated for the GitHub repo (read remote and existing PRs during context gathering).

Run the preflight in references/conventions.md before the first Linear write of a session.

Shared principles

  • Inspect the repo and the Linear backlog before making claims about project structure or existing work.
  • In Plan, ask focused alignment questions one at a time before drafting the issue.
  • Propose 2-3 approaches when more than one viable direction exists.
  • Prefer the smallest end-to-end slice a user can see, use, or evaluate.
  • Decompose epics by demonstrable behavior.
  • A spec issue is incomplete unless it has an exact ## Acceptance criteria section.
  • The Linear issue is the spec.
  • Surface uncertainty instead of filling gaps with guesses.