A raw request becomes a backlog only after someone names the discrete work and writes acceptance criteria, then cuts the oversized pieces into shippable slices. This skill does that conversion. The output is drafts: a tracker write happens only at the final step, and only with explicit approval.

A well-formed issue states an outcome and carries the context to act on cold, ending on criteria that prove it is done. See the anatomy of a great issue for the full standard, the splitting heuristic, the labeling taxonomy, and a worked example.

Climb the determinism ladder: the spec validation and the exact creation command live in scripts/to-issues.py, never in prose. The judgment (what work, what criteria, what split) is the last rung, and the agent keeps only that.

Steps

Gather the source. Collect the conversation, plan, spec, or notes the issues come from into one place. Name the tracker (GitHub or Linear) and read its existing labels and priorities so drafts match the house taxonomy. Done when the source text and the target tracker are both fixed.
Identify the discrete work. List each separable unit of work the source implies as a one-line candidate. Merge duplicates; drop anything already shipped or out of scope. Done when every candidate names one outcome and no two candidates overlap.
Draft each issue into the spec. Per candidate, write title-as-outcome, context (the why), acceptance criteria, scope and non-goals, dependencies, and a size estimate into a JSON spec entry (title, body, acceptance, labels, assignee), following issue-anatomy.md. Done when no drafted issue is missing acceptance criteria.
Split the oversized. A draft that cannot ship in one focused pass gets cut into vertical slices, each independently shippable per the splitting heuristic in issue-anatomy.md. Done when no remaining draft fails the "finishable in one sitting" bar.
Label and link. Apply type, priority, and area labels from the taxonomy read at step 1, then record dependency and parent-child links between the drafts. Done when every draft carries a type label, a priority, and its stated links.
Validate the spec deterministically. Run scripts/to-issues.py check <spec.json>. The script rejects any issue missing a title or acceptance criteria, so a malformed draft never reaches the tracker. Done when the check exits zero.