| name | implement-tasks-sequence |
| description | Orchestrates implementation of all tasks for a feature using an agent team. Respects task dependencies, runs tasks in correct order, parallelizes independent tasks. Does NOT build, run tests, or commit at any point — verification is delegated to the caller (typically `/start-job`, which runs `build-verify` and `pr-review uncommitted` afterwards). Use when the user says "implement all tasks", "run the task sequence", "implement the feature", or wants to execute multiple tasks from a task list end-to-end.
|
| allowed-tools | Bash Read Grep Glob Edit Write Agent |
| model | claude-sonnet-4-6 |
Implement Tasks Sequence
Orchestrate the full implementation of a feature by executing all tasks from a task list using an
agent team. The flow runs uninterrupted — no per-task builds, no per-task tests, no commits, and
no final verification. Verification is the caller's responsibility (normally /start-job, which
runs build-verify and then pr-review uncommitted after this skill completes).
Step 1: Load the Task List
Read .claude/tasks/[feature-name]/tasks.md to understand:
- All tasks and their phases
- The dependency graph
- Which tasks can run in parallel vs. which must be sequential
Step 2: Plan Execution Order
Build an execution plan from the dependency graph:
- Group tasks into waves — each wave contains tasks whose dependencies are all satisfied by prior waves.
- Tasks within the same wave that have no mutual dependencies can be dispatched to agents in parallel.
- Tasks that depend on a previous task must wait for it to complete.
Briefly summarize the execution plan (wave count, parallel groups) and proceed — no approval needed.
Step 3: Execute Each Wave
For each wave, dispatch tasks to agents:
Per-Task Agent Instructions
Each agent receives:
- The task file (
.claude/tasks/[feature-name]/[num]_task.md)
- The PRD (
.claude/tasks/[feature-name]/prd.md)
- The Tech Spec (
.claude/tasks/[feature-name]/techspec.md)
- Instructions to follow the
/implement skill workflow for coding conventions and applicable
skills — BUT skip all build, test, lint, and commit steps. Only source code changes should
be made. Verification is centralized at the end of the sequence.
Explicitly tell each agent:
Do NOT run builds, do NOT run tests, do NOT run lint-format, do NOT commit. Only write the code
changes for your assigned task. Verification and commit decisions happen at the end of the
overall flow.
Wave Completion Gate
A wave is complete when all tasks in the wave have finished writing their code changes. Since there
is no per-task verification, the gate is just task completion. Do not start the next wave until the
current wave's agents have all returned.
Step 4: Report Results
Provide a concise summary:
- Which tasks were implemented (list, with a one-line description each)
- Any tasks that could not complete (and why)
- Reminder: nothing has been committed, nothing has been verified — that is the caller's
responsibility. When invoked by
/start-job, build-verify runs next, followed by
pr-review uncommitted.
Important Rules
- No per-task verification. Do not run builds, tests, or lint between tasks. This slows the
flow without catching anything the final pass wouldn't catch.
- No per-task commits. Everything stays uncommitted until the user reviews and commits.
- Respect dependencies — never start a task before its dependencies are done.
- Parallelize when safe — tasks in the same wave with no mutual dependencies should run as
parallel agents.
- Use worktree isolation — when running parallel agents, use
isolation: "worktree" to avoid
file conflicts. Merge changes back after each wave.
- Stop on hard failure — if an agent reports it cannot complete its task (e.g. missing
context, conflicting changes), stop and surface to the user rather than continuing with a broken
base.