Use this with ChatGPT, Claude, Codex, or any other assistant when you want help converting a plan into CAR-compatible tickets.
Prompt pattern:
- Share your implementation plan.
- Attach or paste this skill doc.
- Ask the assistant to produce
TICKET-###*.mdfiles that follow these rules.
- Filename:
TICKET-###*.mdExamples:TICKET-001.md,TICKET-120-api-parity.md - Tickets must be in ascending numeric order.
- Numbers do not need to be consecutive. Leave gaps if follow-up tickets are likely.
---
ticket_id: "tkt_0123456789abcdef"
agent: "codex"
done: false
---Required (linted):
ticket_id: stable logical ticket identity used by CAR state; normal CAR writer flows auto-generate this, and repair tools can backfill it for older ticketsagent: registered CAR agent id for this repo (for examplecodex,opencode) or the special valueuserdone: boolean
Do not use assistant product names as agent values (for example chatgpt or claude) unless those exact ids are configured in CAR for this repo.
Common optional fields:
title: short, outcome-focused summarygoal: one-sentence statement of intentmodel: pin a specific model when necessary
Keep tickets concise and independently verifiable.
Preferred sections:
## Tasks- concrete implementation steps## Acceptance criteria(or## Exit criteria) - observable outcomes## Tests- commands or explicit verification steps## Notes- only if necessary
Write criteria so another agent can prove completion without guesswork.
CAR's ticket_flow:
- Executes tickets in ascending order
- Picks the first ticket where
done != true
Implications:
- Put prerequisites in lower-numbered tickets.
- Never make a lower-numbered ticket depend on a higher-numbered one.
- If reverse dependencies appear, reorder or split tickets.
- Each ticket must be independently completable when its turn arrives.
- Implementation work -> repo agents (
codex,opencode, etc.). - Final human review/signoff -> user-assigned ticket near the end.
- In PMA mode, prefer delegation, not direct code edits.
A good ticket:
- Has a single, well-scoped outcome
- References specific files/modules when useful
- Includes explicit verification (tests, checks, or observable behavior)
- Avoids vague language (
"improve","clean up","fix issues") without criteria
---
title: "<Outcome-focused title>"
agent: "codex"
done: false
goal: "<What will be true when this ticket is complete>."
---
## Tasks
- <Concrete implementation step>
- <Concrete implementation step>
## Acceptance criteria
- <Observable behavior or artifact>
- <Observable behavior or artifact>
## Tests
- <Commands to run or explicit checks>- Missing or invalid frontmatter
- Cross-ticket dependency deadlocks
- Tickets that rely on unstated or hidden context
done: truewithout evidence against acceptance criteria