# To Tickets > Draft, publish, or repair delivery tickets from settled work, with acceptance and real dependencies. Use only for explicit ticketing requests. - Skill: `stevennitesh/to-tickets` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add stevennitesh/to-tickets` - Raw SKILL.md: https://api.skillmd.com/api/skills/stevennitesh/to-tickets/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: stevennitesh (https://skillmd.com/u/stevennitesh) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/stevennitesh/to-tickets --- # To tickets Turn a settled outcome into useful delivery boundaries, preserving the meaning that each implementer needs. Drafting does not require a tracker. Publication is conditional, and implementation follows the user's original authorization. ## 1. Establish the source and need for decomposition Read the supplied spec, conversation, or audit direction and its decision-bearing references. Identify accepted behavior, constraints, acceptance, and remaining uncertainty. A parent spec is useful but not required. Inspect relevant current owners, callers, and proof to find practical boundaries; do not survey unrelated code or re-open settled requirements. For one coherent outcome without a useful ownership or dependency handoff, return one bounded work item instead of manufacturing a graph. If the user explicitly requests one tracker ticket, retain that publication request. Resolve ordinary technical details within scope. For consequential missing product decisions, identify the owner and affected work; for a real architecture question, use `$codebase-design` when available. Draft independent settled work while marking dependent items blocked. Do not label uncertain work ready or invent decisions to complete a plan. ## 2. Choose cohesive delivery boundaries Create the fewest tickets that each deliver a meaningful, independently checkable outcome. Prefer a narrow path through the affected behavior over separate schema, API, and UI tickets, but do not force every layer into every slice. Fold setup, documentation, and tests into the outcome they support. A technical prerequisite deserves its own ticket when migration, compatibility, ownership, or an independently verifiable enabling result makes that useful. Name what it unlocks and how to prove it. An early experiment or thin real path earns a learning role only when its result changes later work. Do not add a prefactoring stage to every feature. For wide changes, cross-ticket integration, or learning that changes later work, read [Delivery boundaries](references/delivery-boundaries.md). Keep coherent changes together when splitting would manufacture temporary compatibility or conceal unfinished integration. Size work so a fresh agent can recover the needed context; context-window estimates do not replace meaningful boundaries. ## 3. Preserve acceptance and real dependencies Each ticket needs its outcome, source pointer, settled decisions and scope, observable acceptance, consequential constraints, and actual blockers. A source link does not replace the acceptance that this ticket must satisfy. Preserve the source's distinguishing input, state, result, and evidence class rather than replacing them with generic instructions to handle edge cases or add tests. When a plausible wrong implementation could satisfy the ordinary happy case, retain the smallest source-supported example that distinguishes the intended behavior. Preserve settled meaning rather than inventing edge cases or new tests. Give each delivery-changing commitment an owning ticket. Repeat acceptance only at a consumer that can independently lose that meaning. When work produces an input another ticket needs, make that dependency explicit and prove the consuming path with the actual produced result. Use stable pointers or a compact accepted schema when they prevent guessing; leave routine implementation mechanics open. An edge means a required predecessor outcome, not preferred order or possible file overlap. Live concurrency and integration coordination belong to execution. Check for missing commitments, duplicate work, unknown blockers, cycles, and a truthful starting set. A required human decision or permission is an explicit prerequisite, not an executable agent ticket. Do not claim a starting set exists when every item is blocked. ## 4. Deliver the draft or publish Read each ticket as an implementer receiving it alone. Ensure it identifies the governing shared constraints, required predecessor result, and acceptance needed to avoid guessing. Include compact local details or accessible canonical pointers; do not copy the entire specification. For measured acceptance, preserve the source's workload, metric, threshold or baseline, and operating conditions rather than redefining them during decomposition. Present the proposed outcomes, acceptance, dependencies, and why the boundaries are useful. Use one local document or the conversation for a draft unless the user requests separate files. Revise when feedback changes the breakdown without silently changing accepted behavior. When publishing or repairing tracker items is requested, read [Tracker publication](references/tracker-publication.md). Prepare the concrete effects, follow existing authorization and repository policy, and verify the result. Missing tracker setup blocks publication, not the useful draft. Do not auto-publish merely because the skill is named to-tickets. Finish with the draft or verified ticket pointers, dependencies, actionable work, and material limits. Identify whether delivery covers a complete parent or a named subset; preserve the exact published identities and unresolved decision or permission gates. Readiness means prerequisites are resolved, not that an item is safe to run beside every other ready item. Parallel-implement owns that live independence check and turns acceptance into worker assignments. Already-authorized implementation can continue; a planning or ticket-publication request alone does not authorize implementing the graph, spawning workers, or closing a parent.