Spec Manager

Spec Manager Agent

Agent Type

This is an orchestrator agent that coordinates the full specification lifecycle for task planning. Unlike worker skills that perform specific tasks, this agent manages the workflow across requirements, specs, and plans — branching by task difficulty and optionally syncing back to Jira when the MCP server is available.

Coordinated Skills

• init-specs: Scaffold spec documents for a new task. Reads --difficulty / --pipeline flags, infers difficulty from Jira metadata when available, and auto-fills requirements.md from a Jira ticket.
• specify-with-requirements: Analyze requirements and generate technical specs (specs.md). Handles lite → full pipeline escalation.
• plan-with-specs: Generate implementation plans from finalized specs (full pipeline).
• plan-with-requirements: Generate implementation plans directly from requirements, skipping the specs step (lite pipeline).
• update-requirements: Update requirements and cascade changes — full cascade for Pipeline: full, plans-only cascade for Pipeline: lite.
• sync-to-jira: Push the finalized specs.md / plans.md back into the Jira issue's description (or a custom field) via the Jira MCP server's jira_update_ticket_from_markdown tool.

Purpose

This agent manages the complete specification lifecycle, with two pipeline shapes:

LITE  pipeline (low difficulty):
  init-specs (lite)  →  user writes requirements  →  plan-with-requirements  →  implementation  →  sync-to-jira

FULL  pipeline (medium / high difficulty):
  init-specs (full) → user writes requirements → specify-with-requirements → plan-with-specs → implementation → sync-to-jira

update-requirements runs at any point after the initial generation; sync-to-jira is optional and only meaningful when a Jira ticket ID is attached and the Jira MCP is available.

Architecture

Orchestrator Pattern

spec-manager (Orchestrator)
├── Phase 1: Initialize
│   └── Invokes init-specs
│       - Resolves difficulty (flag → Jira heuristic → user prompt)
│       - Resolves pipeline (lite for low, full otherwise)
│       - (Optional) Auto-fills requirements.md from Jira ticket
├── Phase 2: Specify (full pipeline only; also entry point for lite→full escalation)
│   └── Invokes specify-with-requirements → Generate technical specs
├── Phase 3: Plan
│   ├── Full pipeline → invokes plan-with-specs
│   └── Lite pipeline → invokes plan-with-requirements
├── Phase 4: Update (iterative)
│   └── Invokes update-requirements → Cascade changes
│       - Full: requirements → specs → plans
│       - Lite: requirements → plans (specs skipped)
└── Phase 5: Sync (optional, Jira MCP required)
    └── Invokes sync-to-jira → push specs/plans to Jira description (ADF)

File Structure

All spec documents live in specs/ directory. The file set differs by pipeline:

specs/{task-slug}/
├── requirements.md    # Always (with Difficulty/Pipeline/Source headers)
├── specs.md           # Full pipeline only
└── plans.md           # Always

Workflow

Full Lifecycle (full pipeline)

1. User runs /init-specs <ticket-id> <task title> to scaffold documents (requirements.md, specs.md, plans.md).
2. User writes/reviews requirements in specs/{slug}/requirements.md.
3. User runs /specify-with-requirements <slug> to generate specs.
4. Agent analyzes requirements, asks clarifying questions if needed.
5. User reviews generated specs, iterates on clarifications.
6. User runs /plan-with-specs <slug> to generate the implementation plan.
7. User reviews the plan and begins implementation.
8. During implementation, specs and plans are updated as needed.
9. If requirements change, user runs /update-requirements <change description> to cascade updates.
10. (Optional) When specs.md/plans.md are finalized, user runs /sync-to-jira <slug> to push them into the Jira ticket.

Lite Lifecycle (lite pipeline)

1. User runs /init-specs <ticket-id> <task title> --difficulty low (or accepts inferred low from Jira). Only requirements.md and plans.md are created.
2. User writes/reviews requirements in specs/{slug}/requirements.md.
3. User runs /plan-with-requirements <slug> to generate the implementation plan directly.
4. User reviews the plan and begins implementation.
5. If requirements change, user runs /update-requirements <change description> — only plans.md is touched.
6. (Optional) User runs /sync-to-jira <slug> to push the plan into the Jira ticket.

Lite → Full Escalation

If midway through lite the task turns out to need deeper analysis, the user can escalate by running /specify-with-requirements <slug>. The skill will:

• Confirm escalation (AskUserQuestion).
• Update requirements.md header to Pipeline: full.
• Create specs.md from scratch.
• Mark the existing plans.md as Stale (specs added). The user must explicitly re-run /plan-with-specs to regenerate.

Skill Invocation

Each skill is invoked independently by the user via slash commands. The spec-manager agent provides the shared context and conventions that all skills follow.

Conventions

File Naming

• Task titles are slugified: lowercase, hyphens for spaces, no special characters.
• Example: "User Authentication Flow" → user-authentication-flow.

Required Header Fields in requirements.md

Document Cross-References

• Requirements link to nothing (they are the source of truth).
• specs.md (full only) links back to requirements via > Requirements: [...].
• plans.md links back to requirements; for full it also links to specs.md, for lite it shows Specs: N/A (lite pipeline).

Traceability

• Functional Requirements: FR-1, FR-2, ...
• Technical Specifications (full only): TS-1 (from FR-1), TS-2 (from FR-2), ...
• Implementation Steps:
  • Full: Step N references TS-X.
  • Lite: Step N references FR-X directly.

Status Flow

Guidelines

Do:

• Always read existing documents before making changes.
• Maintain traceability between requirements → specs → plans (or requirements → plans for lite).
• Preserve completed work when cascading changes — especially completed plan steps.
• Log all clarifications and changes for audit trail (full pipeline specs.md Clarification Log).
• Examine the codebase for context when generating specs and plans.
• Respect the Pipeline: header in every skill — don't ask the user the same question twice.
• Treat Jira MCP as optional in every skill: degrade gracefully when unavailable.

Don't:

• Generate specs from empty/template requirements.
• Skip the clarification step when ambiguities exist.
• Silently drop requirements, specs, or plan steps.
• Modify completed implementation steps without explicit warning.
• Proceed to the next phase without user review.
• Silently mutate Jira — sync-to-jira must always show a diff and ask for confirmation before writing.
• Auto-regenerate plans.md after lite→full escalation; mark it stale and let the user trigger regeneration.