Skills Overview
Specialized instruction sets that activate on triggers to guide Claude through structured research workflows.
What Are Skills?
Skills are specialized instruction sets that activate when specific triggers are detected in the conversation. Each skill contains domain-specific rules, templates, formats, and constraints that shape how Claude approaches a particular type of task.
Unlike general prompting, skills enforce structured processes — they require specific outputs, mandate checkpoints, and prevent common failure modes. A skill doesn't just suggest what to do; it defines what must happen and in what order.
Skills vs. Agents
Skills are instruction sets that shape Claude's behavior within the main conversation. Agents are separate subprocesses with their own context windows and limited tool access. Skills orchestrate when agents are dispatched.
How Skill Routing Works
The using-propel meta-skill acts as the central router. It activates before any action and checks whether a Propel skill applies to the current user request. The routing follows a strict priority order:
- Check the current mode. If the triggered skill is not active in the current mode, inform the user and suggest
/switch to the appropriate mode.
- Match against trigger patterns. Each skill has specific trigger phrases (e.g., "start an investigation", "validate this", "survey").
- Use the first match. Process skills are checked before research skills, which are checked before implementation skills.
Complete Skills Reference
Every skill in Propel, organized by category with trigger patterns and mode availability.
| Skill | Trigger | Active Modes |
|---|
| using-propel |
Always active — routes to correct skill |
Researcher Engineer Debugger Trainer |
Literature
| Skill | Trigger | Active Modes |
|---|
| deep-research |
"survey", "literature review", "what approaches exist" |
Researcher Engineer Debugger |
| paper-extraction |
"process these papers", "build paper database" |
Researcher Engineer Debugger |
Investigation
| Skill | Trigger | Active Modes |
|---|
| investigation |
"start an investigation", "trace X", "what touches X", "archeology" |
Researcher Engineer Debugger |
Design
| Skill | Trigger | Active Modes |
|---|
| research-design |
"propose how to", "design the implementation", "plan the architecture" |
Engineer |
| writing-plans |
"write the plan", "break into tasks" |
Engineer |
Implementation
| Skill | Trigger | Active Modes |
|---|
| subagent-driven-research |
"go", "implement this", "start building" (after approved plan) |
Engineer |
| research-validation |
"validate this", "test the implementation", "check if this works" |
Engineer |
| verification-before-completion |
"verify", claims of "done" or "fixed" |
Engineer Debugger |
Debugging
| Skill | Trigger | Active Modes |
|---|
| systematic-debugging |
Bug reports, training failures, unexpected behavior |
Engineer Debugger |
Learning & Reflection
| Skill | Trigger | Active Modes |
|---|
| think-deeply |
Confirmation-seeking statements, leading questions, binary framings |
Researcher Engineer Debugger Trainer |
| retrospective |
"retrospective", "capture learnings", ~20 turns without one |
Researcher Engineer Debugger Trainer |
Cross-Cutting
| Skill | Trigger | Active Modes |
|---|
| context-hygiene |
"getting long", context feels large, >15 substantive turns |
Researcher Engineer Debugger Trainer |
| using-git-worktrees |
"create a worktree", "experiment branch" |
Engineer |
Supervision & Peer Review
| Skill | Trigger | Active Modes |
|---|
| monitor |
"/monitor", "watch this run", "keep an eye on", multi-step command chains |
Engineer Debugger Trainer |
Frontend
| Skill | Trigger | Active Modes |
|---|
| frontend-design |
Build web components, pages, landing pages, dashboards, React/HTML/CSS layouts, or style/beautify any web UI |
Engineer |
Training
| Skill | Trigger | Active Modes |
|---|
| trainer-mode |
"train", "launch training", "run training", "my training crashed" |
Trainer |
Customization
| Skill | Trigger | Active Modes |
|---|
| project-customization |
"customize Propel", "analyze my project", "detect conventions" |
Researcher Engineer Debugger Trainer |
Plugin Skills
These skills bring an outside model or rubric into the pipeline. The skill owns the interaction contract — when it fires, how the question is framed, how the reply is verified and synthesized — and Claude always owns the synthesis step. External output never reaches you unfiltered, and it never writes to your repository.
The dual-model layer is no longer a plugin skill you invoke. codex-consult fires automatically at every gate and drives the Codex CLI directly — no plugin, no command. The earlier
/c-codex and
/codex-lead commands have been removed: a second opinion you have to ask for is one you get only when you already suspect you need it.
| Skill | Wraps | Purpose | Active Modes |
|---|
| codex-consult |
OpenAI Codex CLI (direct) |
Automatic at every gate. Frames a sharp question, dispatches codex-bridge, verifies every claim against the repo, and presents an attributed card. |
Researcher Engineer Debugger Trainer |
| c-review |
anthropics/claude-code code-review |
Runs the official Anthropic code-review rubric alongside Propel's own auditors, merged into one Gate 3 card. |
Engineer Debugger |
How Skills Are Injected
Skills are injected into Claude's context via hooks on SessionStart. When a Propel session begins, the hook system reads the current mode from .propel/mode.json and injects the relevant skill instruction files as system context. The using-propel meta-skill is always injected regardless of mode — it handles routing to the correct skill based on what the user asks.
Each skill file lives in the skills/ directory as a SKILL.md file with YAML frontmatter containing the skill name, description, and trigger patterns. The hook reads these files and includes them in Claude's context at session initialization.
Skill Categories Explained
| Category | Purpose | When Active |
|---|
| Meta | Routes requests to the correct skill; enforces mode boundaries | Always |
| Literature | Structured surveys, paper processing, and literature databases | All modes except Trainer |
| Investigation | Scaffolds scratch/ directories for empirical exploration | All modes except Trainer |
| Design | Converts investigation findings into implementation plans | Engineer only |
| Implementation | Executes plans with 3-stage review and validation gates | Engineer (validation also in Debugger) |
| Debugging | Root-cause analysis with evidence standards and 3-strike limit | Engineer, Debugger |
| Learning | Anti-sycophancy guards and experiment retrospectives | All modes |
| Cross-Cutting | Context management, git worktrees, session hygiene | Varies by skill |
| Frontend | Production-grade UI with distinctive design choices (vendored from anthropics/skills) | Engineer only |
| Training | Training execution, runtime bug fixing, monitoring | Trainer only |
| Customization | Project profiling and persistent convention detection | All modes |
| Dual-Model Skills | Bring an outside model or rubric in (codex-consult, c-review); Claude owns the contract and the verification | codex-consult: automatic. c-review: Gate 3 / pre-PR |
Propel