Overview

Create the dt-coach agent definition, entry prompts for starting and resuming coaching sessions, and the coaching state persistence protocol. This phase assembles the Phase 1 foundation into a functional coaching experience.

Deliverables

Agent Specification

The dt-coach agent:

• Identity: Collaborative DT coaching colleague, never quizzes, never lectures, never directs
• Philosophy: Think (internal reasoning) / Speak (observations, 2-3 sentences) / Empower (options with trade-offs)
• Handoffs: "Learn DT" → dt-learning-tutor (V2), "Deep Research" → task-researcher, "Start RPI" → rpi-agent
• Instructions: Three-tier loading (ambient always, method on glob, deep on-demand)
• State: .copilot-tracking/dt/{project-slug}/coaching-state.md
• Hat-switching: 2 shared + 19 specialized hats activated by method context
• Mid-session subagent dispatch: Coach dispatches runSubagent for focused research queries (industry data, competitive analysis, technical constraints) without triggering a full handoff. The coach receives results and weaves them into the coaching conversation, maintaining session state and identity continuity. This mirrors how task-researcher dispatches subagents for all research: the coach adopts the same pattern for mid-coaching information needs.

Coaching State Schema

project: "{{project-slug}}"
current_method: 3
current_phase: "execution"
dt_space: "problem"
methods_completed: [...]
hint_calibration:
  level: 2
  pattern_notes: "..."
user_preferences:
  detail_level: "moderate"
  coaching_style: "collaborative"
session_recovery_instructions: |
  Free-text recovery context for session continuity

Authoring Standards

Agent and prompt files follow .github/instructions/prompt-builder.instructions.md:

• Agent: phase-based protocol (## Required Phases → ### Phase N: ...)
• Prompts: step-based protocol (## Required Steps → ### Step N: ...), end with --- + activation
• Input variables: ${input:variableName} syntax
• Frontmatter: description:, tools:, handoffs: fields

Artifact creation uses the RPI pipeline: task-researcher gathers DT4HVE source material for the agent and prompt specifications, task-planner sequences creation, task-implementor authors files, task-reviewer validates against prompt-builder standards.

Dependencies

• Phase 1 complete (all 6 foundation instruction files)

Success Criteria

• Agent file created with proper handoffs and tool declarations
• Mid-session subagent dispatch specified in agent definition
• Start-project prompt guides setup: project name, stakeholders, industry, Method 1 entry
• Resume-coaching prompt reads state and provides context summary
• Coaching state schema supports multi-project, hint calibration, session recovery
• Agent interacts conversationally, no task-output patterns
• All files pass task-reviewer validation against prompt-builder standards
• Artifacts registered in collections/design-thinking.collection.yml with path and kind fields
• Artifacts registered in collections/hve-core-all.collection.yml with path and kind fields
• npm run plugin:generate run after updating collection manifests