Variables:
<BacklogPath>: backlog directory. Default:@docs/backlog/or@docs/prdwhichever exists. If both exist ask the user to choose or specify a custom path.<PlanPath>: path to the execution plan. Default: session plan.
Create or update an execution-ready implementation backlog in <BacklogPath> using <PlanPath>.
The backlog must consist of:
- an index file at
<BacklogPath>/index.md - epic files under
<BacklogPath>/epics/(if useful) - task files under
<BacklogPath>/tasks/
You are producing backlog artifacts for autonomous AI coding agents. Do not produce a product summary, broad architecture rewrite, or vague project-management prose.
Objectives
- Read the execution plan and source materials.
- Inspect the existing backlog first and preserve useful structure, IDs, and conventions.
- Create or update a small set of coherent epics when needed.
- Break work into the smallest independently reviewable and executable tasks possible.
- Ensure every task is self-contained and executable without chat context.
- Update
<BacklogPath>/index.mdso the backlog is navigable, dependency-aware, and ready for parallel agent execution.
Core principles
- Default to
owner_type: agent. - Scope tasks as narrowly as possible while preserving meaningful standalone progress.
- If a task can be split into smaller independently valuable tasks, split it.
- The bar for
owner_type: humanis very high. - Do not reopen design decisions already resolved in
<PlanPath>unless there is a clear contradiction or critical gap. - Reuse and update existing backlog items when appropriate instead of duplicating them.
- Prefer tasks that can be validated independently and merged safely.
- Prefer parallelizable decomposition over long serial dependency chains.
Operating rules
- Inspect existing backlog files first and follow established repository conventions.
- Prefer adapting existing structure over inventing a new one.
- Use
<PlanPath>as the primary source for workstream boundaries, design direction, and decomposition guidance. - Split tasks that mix design, implementation, migration, testing, rollout, or unrelated code areas unless they are inseparable.
- Split tasks that require different validation strategies.
- Merge tasks only when separation would create artificial overhead without meaningful execution value.
- Deduplicate overlapping tasks and normalize terminology before writing files.
- Keep one epic or task per file.
- Use kebab-case filenames, prefixed by stable IDs when conventions allow.
Task quality bar
Every task must:
- be executable without chat history
- identify the relevant code/docs paths or subsystems
- describe the exact intended change
- include explicit acceptance criteria
- include validation steps
- cite the relevant source plan/design sections
- list real dependencies only when required
Do not create tasks that:
- span multiple unrelated modules or layers
- mix vague investigation with implementation
- say “handle edge cases” without naming them
- omit how success will be validated
- require rediscovering decisions already documented in
<PlanPath>
Owner type rules
Use owner_type: human only when the task primarily requires:
- unresolved business or policy judgment
- stakeholder/product/design approval
- legal/compliance signoff
- credentials or system access unavailable to agents
- real-world operational coordination
Otherwise use owner_type: agent.
Required task frontmatter
Every task file must contain frontmatter with at least:
idtitleepicowner_typestatusprioritydepends_onrepo_pathssource_refsacceptance_criteriavalidation_steps
Backlog update behavior
When updating an existing backlog:
- preserve stable IDs where possible
- update existing tasks instead of duplicating equivalent work
- mark superseded tasks explicitly and link replacements
- if splitting a large task, update the parent task to reference the new child tasks
- do not silently delete backlog intent unless clearly obsolete
Index requirements
<BacklogPath>/index.md must:
- summarize epics/workstreams
- link every active task
- show status, owner type, and dependencies
- highlight tasks that can run in parallel
- note blocked or superseded work
- link back to the primary source plan/design materials