AgentSkillsCN

direct

轻量级会话,适合快速的临时性工作。无需审问,无需规划——只需激活、工作,然后结束。可通过“快速任务”“只需做这件事”“/做这件事”“临时性工作”等语句触发。

SKILL.md
--- frontmatter
name: direct
description: "High-level project vision designer — splits work into sequential or parallel chapters, defines decision principles, produces structured plans consumable by /coordinate. Triggers: \"create a direct\", \"design a project vision\", \"plan the project\", \"split this into chapters\", \"evolve the vision\"."
version: 1.0
tier: protocol

High-level project vision designer — produces structured, diffable vision documents with tagged chapters for /coordinate execution.

Direct Protocol (The Architect's Blueprint)

Execute §CMD_EXECUTE_SKILL_PHASES.

Session Parameters

json
{
  "taskType": "DIRECT",
  "phases": [
    {"label": "0", "name": "Setup",
      "steps": ["§CMD_PARSE_PARAMETERS", "§CMD_SELECT_MODE", "§CMD_INGEST_CONTEXT_BEFORE_WORK"],
      "commands": [],
      "proof": ["mode", "session_dir", "parameters_parsed"]},
    {"label": "1", "name": "Interrogation",
      "steps": ["§CMD_INTERROGATE"],
      "commands": ["§CMD_ASK_ROUND", "§CMD_LOG_INTERACTION"],
      "proof": ["depth_chosen", "rounds_completed"]},
    {"label": "2", "name": "Planning",
      "steps": ["§CMD_GENERATE_PLAN"],
      "commands": ["§CMD_LINK_FILE", "§CMD_APPEND_LOG"],
      "proof": ["plan_written", "plan_presented"]},
    {"label": "3", "name": "Dependency Analysis",
      "steps": [],
      "commands": ["§CMD_APPEND_LOG"],
      "proof": ["graph_produced", "parallel_groups_identified"]},
    {"label": "4", "name": "Execution",
      "steps": ["§CMD_SELECT_EXECUTION_PATH"],
      "commands": [],
      "proof": ["path_chosen", "paths_available"]},
    {"label": "4.A", "name": "Vision Writing",
      "steps": [],
      "commands": ["§CMD_APPEND_LOG", "§CMD_LINK_FILE"],
      "proof": ["vision_written", "chapters_tagged"]},
    {"label": "4.B", "name": "Agent Handoff",
      "steps": ["§CMD_HANDOFF_TO_AGENT"], "commands": [], "proof": []},
    {"label": "5", "name": "Synthesis",
      "steps": ["§CMD_RUN_SYNTHESIS_PIPELINE"], "commands": [], "proof": []},
    {"label": "5.1", "name": "Checklists",
      "steps": ["§CMD_VALIDATE_ARTIFACTS", "§CMD_RESOLVE_BARE_TAGS", "§CMD_PROCESS_CHECKLISTS"], "commands": [], "proof": []},
    {"label": "5.2", "name": "Debrief",
      "steps": ["§CMD_GENERATE_DEBRIEF"], "commands": [], "proof": ["debrief_file", "debrief_tags"]},
    {"label": "5.3", "name": "Pipeline",
      "steps": ["§CMD_MANAGE_DIRECTIVES", "§CMD_PROCESS_DELEGATIONS", "§CMD_DISPATCH_APPROVAL", "§CMD_CAPTURE_SIDE_DISCOVERIES", "§CMD_MANAGE_ALERTS", "§CMD_REPORT_LEFTOVER_WORK"], "commands": [], "proof": []},
    {"label": "5.4", "name": "Close",
      "steps": ["§CMD_REPORT_ARTIFACTS", "§CMD_REPORT_SUMMARY", "§CMD_CLOSE_SESSION", "§CMD_PRESENT_NEXT_STEPS"], "commands": [], "proof": []}
  ],
  "nextSkills": ["/coordinate", "/implement", "/analyze", "/brainstorm"],
  "directives": [],
  "planTemplate": "assets/TEMPLATE_DIRECT_PLAN.md",
  "logTemplate": "assets/TEMPLATE_DIRECT_LOG.md",
  "debriefTemplate": "assets/TEMPLATE_DIRECT.md",
  "requestTemplate": "assets/TEMPLATE_DIRECT_REQUEST.md",
  "responseTemplate": "assets/TEMPLATE_DIRECT_RESPONSE.md",
  "modes": {
    "greenfield": {"label": "Greenfield", "description": "New project vision from scratch", "file": "modes/greenfield.md"},
    "evolution": {"label": "Evolution", "description": "Extend or modify an existing vision", "file": "modes/evolution.md"},
    "split": {"label": "Split", "description": "Decompose a large chapter into sub-chapters", "file": "modes/split.md"},
    "custom": {"label": "Custom", "description": "User-defined blend", "file": "modes/custom.md"}
  }
}

0. Setup

§CMD_REPORT_INTENT:

Designing project vision for ___. Mode: ___. Output: docs/[name]_DIRECT.md. Role: Project Architect — every chapter is a contract with the coordinator.

§CMD_EXECUTE_PHASE_STEPS(0.0.*)

  • Scope: Understand the project goal and select the operating mode.

Mode Selection (§CMD_SELECT_MODE):

On selection: Read the corresponding modes/{mode}.md file. It defines Role, Goal, Mindset, and mode-specific behavior.

On "Custom": Read ALL 3 named mode files first (modes/greenfield.md, modes/evolution.md, modes/split.md), then accept user's framing. Parse into role/goal/mindset.

Record: Store the selected mode. It configures:

  • Phase 0 role (from mode file)
  • Phase 1 interrogation topics (from mode file)
  • Phase 4 writing behavior (from mode file)

Mode-Specific Setup:

  • Greenfield: No existing vision to load. Proceed to interrogation.
  • Evolution: Load the existing vision document (from contextPaths or user-specified path). Display a summary: chapter count, completion status of each (#done-coordinate, #claimed-coordinate, #needs-coordinate, untagged). Store the loaded vision as the "v1 baseline" for later diffing.
  • Split: Load the existing vision document AND identify the target chapter slug. Ask the user which chapter to decompose if not specified in the prompt.

1. Interrogation

§CMD_REPORT_INTENT:

Interrogating ___ project vision assumptions. Drawing from goal, constraints, chapter decomposition, and dependency topics.

§CMD_EXECUTE_PHASE_STEPS(1.0.*)

Topics (Direct Design)

Standard topics for the command to draw from. Adapt to the task -- skip irrelevant ones, invent new ones as needed.

Standard topics (typically covered once):

  • Project goal -- what the project achieves, why it exists, success criteria
  • Constraints & non-negotiables -- hard requirements, timeline, budget, compliance
  • Scope boundaries -- what's in, what's out, where the project ends
  • Chapter decomposition -- natural work boundaries, what can be parallelized
  • Decision principles -- soft guidance for coordinator judgment calls
  • Context sources -- which /analyze, /brainstorm, /research sessions feed this vision
  • Dependency mapping -- cross-chapter dependencies, shared resources, ordering constraints
  • Risk assessment -- what could go wrong, contingency plans, reversibility

Repeatable topics (can be selected any number of times):

  • Followup -- Clarify or revisit answers from previous rounds
  • Devil's advocate -- Challenge assumptions and decisions made so far
  • What-if scenarios -- Explore hypotheticals, edge cases, and alternative futures
  • Deep dive -- Drill into a specific topic from a previous round in much more detail

Mode-specific additional topics:

  • Evolution: Current vision analysis, what changed since v1, migration impact
  • Split: Target chapter complexity analysis, natural sub-boundaries, sub-chapter dependency mapping

2. Planning

§CMD_REPORT_INTENT:

Planning ___ vision structure. ___ chapters identified. Producing DIRECT_PLAN.md with chapter outline and dependency sketch.

§CMD_EXECUTE_PHASE_STEPS(2.0.*)

Unless the user points to an existing plan, you MUST create one.

  • Plan: Fill in every section -- chapter outline with semantic slugs, preliminary dependency sketch, decision principles draft, success criteria.

The plan is a structured outline, NOT the final vision document. It captures:

  1. Chapter List: Each chapter with its semantic slug (@app/auth-system), one-line description, and preliminary scope.
  2. Preliminary Dependencies: Which chapters depend on which (to be validated in Phase 3).
  3. Decision Principles Draft: Natural language guidance extracted from interrogation.
  4. Success Criteria Draft: Measurable outcomes for the overall project.

Walk-through (optional):

code
§CMD_WALK_THROUGH_RESULTS Configuration:
  mode: "plan"
  gateQuestion: "Chapter outline ready. Walk through before dependency analysis?"
  debriefFile: "DIRECT_PLAN.md"
  planQuestions:
    - "Is this chapter scoped correctly — too broad, too narrow?"
    - "Any missing chapters or unnecessary ones?"
    - "Dependencies I'm missing between these chapters?"

If any items are flagged for revision, return to the plan for edits before proceeding.


3. Dependency Analysis

§CMD_REPORT_INTENT:

Analyzing dependencies across ___ chapters. Producing dependency graph and identifying parallel execution groups.

§CMD_EXECUTE_PHASE_STEPS(3.0.*)

This phase is the unique value-add of /direct.

Step 1: Build Dependency Graph

Analyze each chapter's scope for:

  • Shared resources: Files, APIs, database tables, packages touched by multiple chapters
  • Output dependencies: Does Chapter B need Chapter A's output as input?
  • Ordering constraints: Must Chapter A complete before Chapter B can start?

Step 2: Produce Visual Graph

Produce a dependency graph using §CMD_FLOWGRAPH notation showing chapter relationships, parallel groups, and serial chains. Use branch glyphs (├►, ╰►) for parallel groups and sequential flow () for serial dependencies.

See the Dependency Graph section of TEMPLATE_DIRECT_VISION.md for the canonical example format.

Step 3: Identify Parallel Groups

Group chapters that share no dependencies:

  • Chapters with no **Depends on** pointing to each other can run in parallel (multiple coordinators)
  • Chapters with dependencies enforce serial ordering within their chain
  • Present the grouping to the user for approval

Step 4: User Approval

Present the dependency analysis via AskUserQuestion:

"Dependency analysis complete. [N] chapters in [M] parallel groups."

  • "Approve dependencies" -- Accept the graph and proceed to Vision Writing
  • "Adjust dependencies" -- Modify the graph (add/remove dependencies)
  • "Re-analyze" -- Something was missed, re-run the analysis

Phase Transition

Execute §CMD_GATE_PHASE.


4. Execution

§CMD_REPORT_INTENT:

Selecting execution path for vision document.

§CMD_EXECUTE_PHASE_STEPS(4.*)

Gateway phase — presents inline/agent/parallel choice, then enters the selected branch.


4.A. Vision Writing

Produce the final vision document.

§CMD_REPORT_INTENT:

Writing vision document: docs/[name]_DIRECT.md. ___ chapters across ___ parallel groups.

§CMD_EXECUTE_PHASE_STEPS(4.A.*)

Output Path

The vision document is written directly to docs/:

  • Path: docs/[NAME]_DIRECT.md (user-specified or derived from project goal)
  • Provenance: References the current direct session

Writing the Vision

Use TEMPLATE_DIRECT_VISION.md to produce the vision document. Populate all sections from the interrogation and planning phases:

  1. Goal: From interrogation (project goal topic)
  2. Success Criteria: From planning phase draft
  3. Constraints: From interrogation (constraints topic)
  4. Decision Principles: From interrogation and planning
  5. Context Sources: Sessions that fed into this vision (/analyze, /brainstorm refs)
  6. Dependency Graph: ASCII graph from Phase 3
  7. Chapters: Each chapter from the plan, with:
    • Semantic slug as heading identifier
    • Scope description
    • **Depends on**: field (from dependency analysis)
    • **Tags**: #needs-coordinate (for coordinator execution)

Mode-Specific Writing Behavior

Greenfield: Write the full document from scratch.

Evolution:

  1. Write the updated v2 document (full rewrite incorporating all changes)
  2. Diff Phase: Compare v1 baseline (loaded in Setup) against the new v2
  3. Present diff hunks interactively via §CMD_WALK_THROUGH_RESULTS:
    • New chapters: Marked for #needs-coordinate tagging
    • Modified chapters: User decides: re-execute (#needs-coordinate reset), skip (keep current tag), or manual review
    • Removed chapters: User decides: cleanup (#needs-chores), or dismiss
  4. Apply tag changes based on user decisions

Split:

  1. Replace the target chapter with sub-chapters in the existing vision
  2. Sub-chapters inherit the parent's slug as prefix: @app/auth-system/token-service
  3. Original chapter heading becomes a group header (no #needs-coordinate tag)
  4. Sub-chapters get their own #needs-coordinate tags
  5. Run dependency analysis within the sub-chapter group
  6. Write the modified vision document

Report

Execute §CMD_LINK_FILE for the produced vision document.

Phase Transition

Execute §CMD_GATE_PHASE.


4.B. Agent Handoff

Hand off to an autonomous agent for vision writing.

§CMD_EXECUTE_PHASE_STEPS(4.B.*)

§CMD_HANDOFF_TO_AGENT with:

json
{
  "agentName": "writer",
  "startAtPhase": "4.A: Vision Writing",
  "planOrDirective": "[sessionDir]/DIRECT_PLAN.md",
  "logFile": "DIRECT_LOG.md",
  "taskSummary": "Write the vision document: [brief description]"
}

5. Synthesis

When the vision document is complete.

§CMD_REPORT_INTENT:

Synthesizing. Vision document written with ___ chapters. Producing DIRECT.md debrief.

§CMD_EXECUTE_PHASE_STEPS(5.0.*)

Debrief notes (for DIRECT.md):

  • Vision Summary: What the project achieves, chapter count, parallel group count
  • Dependency Analysis: Summary of the dependency graph and parallel opportunities
  • Mode Used: Greenfield/Evolution/Split and what it entailed
  • Evolution Diff (if applicable): Summary of changes from v1 to v2
  • Next Steps: Clear recommendation — invoke /coordinate with the vision path

Walk-through config:

code
§CMD_WALK_THROUGH_RESULTS Configuration:
  mode: "results"
  gateQuestion: "Vision document complete. Walk through the chapters?"
  debriefFile: "DIRECT.md"

Post-Synthesis: If the user continues talking, obey §CMD_RESUME_AFTER_CLOSE.