Docs/Task Planning Rules

Task Planning Rules

Imperative instructions for planning and executing task epics.

Task Decomposition

  1. Decompose — Break the request into atomic, sequential tasks
  2. Assess — For each task, determine if specialized agent orchestration is needed
  3. Order — Arrange by dependencies and logical flow
  4. Validate — Ensure each task is: specific, actionable, independently testable, completable in one session, clear about inputs/outputs/success criteria
  5. Checkpoint — Add approval gates between major phases

Epic Template

# {Epic Name} Epic

**Status**: PLANNED
**Goal**: {brief goal}

## Overview

{Single paragraph starting with WHY — user benefit or problem being solved}

---

## {Task Name}

{Brief task description — 1 sentence max}

**Requirements**:
- Given {situation}, should {job to do}
- Given {situation}, should {job to do}

---

Epic Constraints

  • Overview starts with WHY (user benefit/problem being solved)
  • No task numbering — use task names only
  • Requirements use ONLY "Given X, should Y" format
  • Include ONLY novel, meaningful, insightful requirements
  • No extra sections, explanations, or text beyond the template

Execution Protocol

  1. Complete only the current task
  2. Validate — verify the task meets its success criteria
  3. Report — summarize what was accomplished
  4. Await approval — get explicit user approval before proceeding

On Completion

  1. Update epic status to COMPLETED with date
  2. Archive the epic file
  3. Remove from active plan

Execution Modes

Supervised Mode (default)

Human-in-the-loop execution. Use when:

  • Spec maturity is below SPECIFIED
  • Working on novel or ambiguous requirements
  • Single-agent sessions with direct human oversight

Rules:

  • Complete only the current task before proceeding
  • Get explicit user approval between tasks
  • If a task reveals new information, pause and re-plan
  • If blocked or uncertain, ask rather than assume

Autonomous Mode

Agent operates without human checkpoints. Use when:

  • Spec maturity is SPECIFIED (all requirements, interfaces, and edge cases defined)
  • Operating within a swarm or parallel agent execution
  • A conductor agent is coordinating the work

Rules:

  • Follow the spec's Given/should requirements as success criteria
  • Validate against rubrics (TDD rubric, code review rubric, design rubric) — these are the quality gates
  • On completion, report results to the conductor or commit directly
  • On failure or ambiguity, stop and report — don't guess

Constraints

  • Tasks should be independent — completing one shouldn't break others
  • Avoid breaking changes unless explicitly requested (open/closed principle)
  • Task quality is bounded by rubrics and success criteria, not line count