Skip to content

scene-skills-protocol.md

Latest commit

 

History

History
121 lines (90 loc) · 4.46 KB

scene-skills-protocol.md

File metadata and controls

121 lines (90 loc) · 4.46 KB

Scene Skills Protocol

Standard protocol for Clawsight Scene Skills — how they consume profile data, interact with each other, and degrade gracefully.

What is a Scene Skill?

A Scene Skill is a standalone SKILL.md that consumes Clawsight's profile data (USER.md, MEMORY.md, memory/projects/*) for specialized analysis. Scene Skills are read-only consumers — they never write to Clawsight's memory files.

Design Principles

  1. Read-only access: Scene Skills read profile data but never modify it
  2. Graceful degradation: Must work without profile data (Lite Mode), better with it (Rich/Enhanced Mode)
  3. Evidence-first: All claims must cite specific data sources with confidence markers
  4. Independent installability: Each Scene Skill is a standalone SKILL.md
  5. Size budget: < 6KB per SKILL.md (methodology docs extracted to docs/)
  6. Language matching: Output in the user's language

Operating Modes

Mode Condition Experience
Enhanced Profile + upstream skill outputs available Full chain analysis with cross-validated data
Rich ≥1 Clawsight source file found with relevant data Profile-driven analysis
Lite No profile data available Interactive Q&A, still useful but less personalized

Mode Detection Logic

1. Check for USER.md, MEMORY.md, memory/projects/*.md
2. Check for upstream skill output blocks (<!-- SKILL_OUTPUT ... -->)
3. If upstream outputs found → Enhanced Mode
4. If ≥1 source found with relevant data → Rich Mode
5. If no profile data → Lite Mode

Cross-Skill Data Passing

Format

Skills pass structured data via HTML comment blocks appended to the end of reports:

<!-- SKILL_NAME_OUTPUT
key1: value1
key2:
  - item1
  - item2
nested:
  sub_key: sub_value
-->

Rules

  1. Append-only: Output blocks are appended after the visible report
  2. YAML format: Content inside comment blocks uses YAML syntax
  3. Namespaced: Each skill uses its own prefix (e.g., CAREER_MIRROR_OUTPUT, TECH_SPECTRUM_OUTPUT, CAREER_SIM_OUTPUT)
  4. Optional consumption: Downstream skills check for but don't require upstream outputs
  5. No circular dependencies: Data flows in one direction only

Current Data Flow (4-Skill Chain)

career-mirror → CAREER_MIRROR_OUTPUT
                    ↓
              tech-spectrum → TECH_SPECTRUM_OUTPUT
                                   ↓
                            career-sim → CAREER_SIM_OUTPUT
                                              ↓
                                        tech-compass (reads all upstreams)

Invocation Patterns

Slash Commands

Each Scene Skill registers a slash command:

Command Skill
/career-mirror Career introspection
/tech-spectrum AI disruption positioning
/career-sim Divergent career path simulation
/tech-compass Action planning

Natural Language Triggers

Each SKILL.md defines natural language trigger patterns. Examples:

  • "analyze my career" → career-mirror
  • "where do I stand with AI" → tech-spectrum
  • "what are my career options" → career-sim
  • "what should I learn next" → tech-compass

Chain Invocation

Users can trigger the full chain:

  • Run /career-mirror first
  • Then /tech-spectrum (auto-detects career-mirror output)
  • Then /career-sim (auto-detects both upstream outputs)
  • Then /tech-compass (auto-detects all upstream outputs, aligns to chosen path)

Or run any skill independently — each degrades gracefully without upstream data.

Constraints for All Scene Skills

  1. Evidence-first: Every claim must reference specific profile data or state the basis
  2. Observation, not prescription (career-mirror specific): Describe what IS, not what SHOULD BE
  3. Read-only: Never suggest modifying USER.md/MEMORY.md directly
  4. Constructive framing: Present gaps as opportunities, not failures
  5. Confidence marking: Use [high]/[medium]/[low] confidence indicators
  6. Scope boundary: Each skill has a defined scope — do not overlap with other skills' domains

MCP Enhancement Path

Phase Capability Data Sources
Phase 1 (current) Pure Skill LLM knowledge only
Phase 2 MCP tools Web search, job market APIs, trend APIs
Phase 3 Data layer Structured trend database, historical tracking

Phase 2+ enhancements are additive — Phase 1 functionality is always the fallback.