Agent-Skills.md

Agent Skills: Torque Refine

Use this skill for planning and requirements work — exploring ideas, improving prompts, creating PRDs, breaking features into tasks, or iterating on existing plans. Trigger on "plan this feature", "break this down into tasks", "create a PRD", "improve this prompt", "what should the subtasks be", "refine the requirements", or when a Jira ticket needs task breakdown with subtasks. Routes to 6 modes (Start, Improve, Summarize, PRD, Plan, Refine) based on intent.

UncategorizedID: carvalab/k-skills/torque-refine

Repository

carvalab/k-skills
carvalab

Install this agent skill to your local

pnpm dlx add-skill https://github.com/carvalab/k-skills/tree/HEAD/skills/torque-refine

Skill Files

Browse the full folder contents for torque-refine.

Download Skill

Loading file tree…

skills/torque-refine/SKILL.md

Skill Metadata

Name
torque-refine
Description
Use this skill for planning and requirements work — exploring ideas, improving prompts, creating PRDs, breaking features into tasks, or iterating on existing plans. Trigger on "plan this feature", "break this down into tasks", "create a PRD", "improve this prompt", "what should the subtasks be", "refine the requirements", or when a Jira ticket needs task breakdown with subtasks. Routes to 6 modes (Start, Improve, Summarize, PRD, Plan, Refine) based on intent.

Torque Refine

Single entry-point for all prompt engineering, requirements gathering, and planning workflows. Routes to 6 internal capabilities based on user intent. Supports Jira integration to push task breakdowns as subtasks.

Context

This skill runs in planning-only refine sessions:

  • No MR creation — refine sessions never create merge requests or feature branches
  • Multi-repo workspace — the workspace may contain multiple cloned repos for codebase analysis
  • Jira integration — use torque-tool jira to create subtasks and update parent issues

Quick Start

  1. Detect user intent (see routing below)
  2. Load the matching rule: torque-tool skill read torque-refine then read the appropriate reference
  3. Follow that rule exactly

Intent Routing

| User Intent | Mode | Reference | |-------------|------|-----------| | Vague idea, needs discovery | Start | references/start.md | | Has a prompt to optimize | Improve | references/improve.md | | Extract requirements from chat | Summarize | references/summarize.md | | Needs full requirements doc | PRD | references/prd.md | | Has PRD, needs task breakdown | Plan | references/plan.md | | Update existing PRD or prompt | Refine | references/refine.md |

Default: If the user has a prompt/PRD to optimize, use Improve. If they only have a vague idea with no written artifact, use Start to gather requirements first.

Workflow

Vague Idea --> START --> conversation --> SUMMARIZE --> mini-PRD
                                              |
Clear Feature --> PRD ------------------------+-> full PRD --> PLAN --> tasks.md
                                                                |
Existing Prompt --> IMPROVE                                     v
                                                     Jira subtasks (optional)
Existing PRD/Prompt --> REFINE --> back to any mode

Key Behaviors

  1. Announce mode before starting (State Assertion)
  2. Never write implementation code - planning only
  3. Self-correct mistakes (DETECT/STOP/CORRECT/RESUME)
  4. Verify file saves - Write then Read to confirm
  5. One question at a time in Start and PRD modes
  6. Always present plan before creating Jira issues - never push to Jira without confirmation
  7. Always score quality - Every mode uses the 6-dimension quality assessment

Universal Quality Assessment

All modes must evaluate outputs using 6 dimensions (0-100%):

| Dimension | What It Measures | |-----------|-----------------| | Clarity | Is the objective clear and unambiguous? | | Efficiency | Is the content concise without losing critical information? | | Structure | Is information organized logically? | | Completeness | Are all necessary details provided? | | Actionability | Can an AI take immediate action on this? | | Specificity | How concrete and precise? (versions, paths, identifiers) |

Quality gate: If overall score < 70%, recommend running Improve mode first to optimize before proceeding to the next step.

Jira Integration

After generating tasks.md, offer to push tasks as Jira subtasks:

# Show parent issue details first
torque-tool jira show --key PROJ-456

# Create subtask with priority (always prefix with [repo-name])
torque-tool jira create-subtask --parent PROJ-456 \
  --title "[payments-api] Add retry logic" \
  --description "Scope: files to change..." \
  --assignee me --priority High

# Link dependencies (MRTCH-1 blocks MRTCH-2)
torque-tool jira link --from MRTCH-1 --to MRTCH-2 --type Blocks

# Update parent issue title with repo context
torque-tool jira update --key PROJ-456 \
  --title "[payments-api] Updated title" \
  --description "Updated requirements..."

Subtask Conventions

  • Always prefix with [repo-name]: Every subtask title starts with the repo name in brackets (e.g., [payments-api] Add retry logic)
  • Single-repo tasks: Also update the parent task title with the repo prefix
  • Priority: Use --priority with: Highest, High, Medium, Low, Lowest
  • Dependencies: Use torque-tool jira link --from X --to Y --type Blocks (X blocks Y)
  • Assignee: Use --assignee me to auto-assign subtasks to the current user
  • Description: Include scope (files to change), context, and enough detail for the next agent to work autonomously
  • Granularity: Each subtask should be completable in one coding session
  • Fallback: If torque-tool jira fails (rate limit, auth), save tasks.md and tell the user to create subtasks manually

Transitioning Issues

Workflows differ per project. Always check available transitions before moving:

# Check what transitions are available from current status
torque-tool jira transitions PROJ-456

# Move to a specific status
torque-tool jira transition PROJ-456 --name "Development"

# Close an issue (walks all transitions to Done automatically)
torque-tool jira close PROJ-456

Deleting Subtasks

When the user asks to delete subtasks (e.g., "delete the subtasks", "remove all subtasks", "start fresh"):

# List subtasks first to confirm what will be deleted
torque-tool jira list-subtasks PROJ-456

# Delete individual subtasks
torque-tool jira delete MRTCH-1371
torque-tool jira delete MRTCH-1372

# If delete fails (403), close them instead
torque-tool jira close MRTCH-1371
torque-tool jira close MRTCH-1372

# Or delete a parent and all its subtasks at once
torque-tool jira delete PROJ-456 --delete-subtasks

Always list subtasks before deleting to confirm with the user. Never delete without showing what will be removed.

Output Location

All outputs save to .mula/outputs/:

  • PRDs: .mula/outputs/{project}/
  • Prompts: .mula/outputs/prompts/{id}.md
  • Tasks: .mula/outputs/{project}/tasks.md

After Planning

When plan mode generates tasks.md:

  1. Present the task breakdown to the user
  2. Ask: "Would you like me to create these as Jira subtasks under the parent issue?"
  3. If yes, use torque-tool jira create-subtask for each task
  4. If no: "Your task breakdown is ready. Implement using your preferred workflow."

References

| Reference | Purpose | |-----------|---------| | references/start.md | Conversational exploration and requirements gathering | | references/improve.md | 6-dimension prompt quality assessment and optimization | | references/summarize.md | Extract requirements from conversation into mini-PRD | | references/prd.md | Strategic questioning to create comprehensive PRD | | references/plan.md | Transform PRD into actionable task breakdown + Jira subtasks | | references/refine.md | Iterate on existing PRD or saved prompt + Jira updates |