agents-evolution: Skill for Claude Code

Spec Mode

Use this skill when the task is about:

• specifications
• requirements
• design docs
• team task lists
• decomposition of work into independent zones
• .spec/<slug>/TASK.md, Requirements.md, Design.md

Workflow

1. Start from discovery first if the user is still understanding the project.
2. If the user asks things like:
   • how to improve the dashboard
   • how to improve Telegram Web logic
   • what problems exist
   • what to fix first then do analysis first and do not create .spec/ yet.
3. Only create spec files after the user explicitly asks to:
   • make a spec
   • formalize the plan
   • create requirements/design/tasks
4. Once the user asks for a spec:
   • treat each top-level workstream as an independent spec pack
   • store it in .spec/<slug>/
   • use TASK.md, Requirements.md, Design.md
5. If you need scaffolding, use:
   • scripts/init_spec_pack.sh

What to load

• TASK rules:
  • references/task-format.md
• Requirements rules:
  • references/requirements-format.md
• Design rules:
  • references/design-format.md

Important

• Keep SKILL.md lean; read references only as needed.
• Use two phases:
  • Phase 0: discovery and improvement analysis
  • Phase 1: spec formalization
• Always preserve traceability:
  • Task -> Requirements -> Design
• For subtask traceability use:
  • _Requirements: 1.1, 2.3_
• Every top-level task in TASK.md must have its own responsibility block:
  • owner
• For every top-level task, the responsible side must be chosen explicitly.
• Keep it simple:
  • one task -> one explicit responsibility block
• all subtasks under that task belong to the same responsibility block by default
• owner may be:
  • one person
  • one AI model
  • one named executor group if the team uses group ownership
• owner may also contain several explicit executors if the owner of the plan wants shared execution.
• The number of executors is not tied to the number of tasks.
• Three executors may split tasks evenly, unevenly, or one executor may take several tasks while others take fewer.
• The workload split is defined by the plan owner, not by the template.
• Different tasks may have different owners.
• The same owner may also appear in many tasks.
• If the user explicitly asks for split execution with several executors or agents, spawn separate independent agents for those top-level tasks during execution when the toolset allows it.
• Keep the split aligned with the chosen owner blocks in TASK.md.
• In split execution there is also a coordinating reviewer/integrator role.
• This reviewer does not have to be fixed in TASK.md.
• The important part is that this role exists.
• The reviewer waits until the executors finish, then:
  • collects results from all agents
  • checks whether changes conflict
  • integrates everything together
  • updates the single shared TASK.md
  • runs final verification
  • closes the spec by fact, not by assumption
• Reuse templates from:
  • templates/TASK.template.md
  • templates/Requirements.template.md
  • templates/Design.template.md