Trovella Wiki

Skills Overview

How Trovella's research skills orchestrate structured, multi-step research via prompt-driven workflows and MCP tool integration.

Skills are Trovella's mechanism for teaching an external AI assistant how to conduct structured research. A skill is a markdown file installed in the user's Claude Code environment that instructs the AI platform on interview techniques, plan design, step execution, and output delivery. Skills contain no executable code -- they are prompts that shape the AI platform's behavior.

The Three-Layer Architecture

Trovella's research system separates concerns across three layers:

LayerRoleContains code?
Skills (the instructor)Teach the AI platform how to research -- interview patterns, plan templates, execution strategies, output formattingNo -- markdown prompts only
AI platform (the researcher)Does the actual work -- searches, reads, reasons, analyzes, writesExternal (Claude Code, etc.)
MCP tools (the project manager)Stores plans, tracks progress, enforces structure, enables resumeYes -- @repo/mcp TypeScript

This separation means Trovella's server makes zero LLM API calls for the primary research use case. All reasoning happens externally on the AI platform, subsidized by the user's own subscription. Trovella provides the structure, persistence, and observability.

Skill Inventory

Four skills compose the research system:

SkillFilePurpose
/research.claude/skills/research/SKILL.mdRouter skill -- interviews the user, classifies depth, delegates to scan or deep
/research-scan.claude/skills/research-scan/SKILL.mdQuick scan -- 2-4 step plans, fast turnaround, no checkpoints
/research-deep.claude/skills/research-deep/SKILL.mdDeep dive -- 5-8+ step plans, checkpoints, branching, critique
/research-output.claude/skills/research-output/SKILL.mdDeliverable generation -- formats results, stores output, captures feedback

The user only sees /research. The routing to scan or deep is invisible -- the experience is seamless.

How Skills Relate to Other Components

Skills sit at the top of the research stack, driving everything below:

  • Plan Orchestration -- skills design research plans by calling create_research_plan, then execute them step-by-step through the pull-based loop (get_next_step / submit_step_result). See Plan Orchestration for the state machine, branching, and stall detection.
  • Reasoning -- skills instruct the AI platform on how to reason (analyst perspective, critic perspective, synthesizer perspective), but the actual LLM calls happen externally. The extract_data tool is the one exception where Trovella makes a server-side AI call. See Reasoning for the AI integration layer.
  • Tool Protocol -- skills reference MCP tools by name in their instructions. The AI platform calls these tools via the MCP HTTP transport. See Tool Protocol for authentication, the tool catalog, and logging middleware.
  • Admin Dashboard -- every skill invocation is tracked in the skill_execution table and visible at /admin/skill-executions. See Skill Executions View for the admin UI.

Pages in This Topic

Skill Definitions

What each skill does, its internal structure, and the prompt patterns it uses.

Execution Flow

The end-to-end lifecycle from user prompt through interview, routing, planning, execution, output delivery, and feedback capture.

Routing Logic

How the router skill decides between quick scan and deep dive, including the classification criteria, depth signals, and default behavior.

Lifecycle Tracking

How skill executions are tracked in the database, the log_skill_execution MCP tool, the tRPC admin router, and metadata conventions.

On this page