本文へ移動
cccskills
無料GitHub で公開

create-user-story

Create user stories with duplicate checking across any project tracker (Linear, GitHub Issues, Jira)

インストール方法を見る

含まれるファイル(1)

  • SKILL.md7.7 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

COG Create User Story Skill

When to Invoke

  • User wants to create a new user story, feature request, or issue
  • User says "create story", "new story", "add issue", "write user story", "create ticket"
  • User describes a feature or requirement that should be tracked

Agent Mode Awareness

Check agent_mode in 00-inbox/MY-PROFILE.md frontmatter:

  • If agent_mode: team — use parallel agents to check for duplicates across all active trackers simultaneously, then create the story
  • If agent_mode: solo — check duplicates and create the story sequentially in the main conversation

Command: /create-user-story

Pre-Flight Check

  1. Read 00-inbox/MY-INTEGRATIONS.md to determine which project trackers are active:

    • Linear — use Linear MCP tools
    • GitHub — use gh CLI
    • Jira — use jira CLI or Jira API via WebFetch
    • If NO tracker is active, save the story as a markdown file in 04-projects/[project]/stories/ and inform the user
  2. Read 00-inbox/MY-PROFILE.md to get:

    • Active projects (to determine which project/repo/board to target)
    • User's name (for story author attribution)
  3. Ask the user (if not already provided):

    • What is the feature or requirement? (free-form description)
    • Which project does this belong to? (if multiple active projects)
    • Priority level? (optional — default to medium)

Execution Strategy

Phase 1: Understand the Request

Parse the user's input to extract:

  • Feature description — what the user wants built
  • Target users — who benefits from this
  • Problem statement — what problem this solves
  • Project context — which project this belongs to

If the user provides a brief description, ask clarifying questions before proceeding:

  • Who is the target user for this feature?
  • What problem does this solve?
  • Are there any specific acceptance criteria you have in mind?

Phase 2: Duplicate Check

CRITICAL: Always check for duplicates before creating.

Team Mode (parallel agents)

Launch duplicate-check agents in parallel using the Task tool with run_in_background: true:

Agent: "duplicate-checker-linear" (if Linear is active)

Search Linear for potential duplicate issues.
1. Use ToolSearch to load Linear tools
2. Use mcp__claude_ai_Linear_2__list_issues to search for issues with similar keywords
3. Search across all active projects/teams
4. Return any issues that match by title similarity or description overlap

Search terms: [extracted keywords from user's description]
Return: list of potential duplicates with title, status, URL, and similarity assessment

Agent: "duplicate-checker-github" (if GitHub is active)

Search GitHub Issues for potential duplicates.
Repository: [CUSTOMIZE: your-org/your-repo]

1. gh search issues "[keywords]" --repo [CUSTOMIZE: your-org/your-repo] --json number,title,state,url,body --limit 20
2. Also search closed issues to check if this was already done:
   gh search issues "[keywords]" --repo [CUSTOMIZE: your-org/your-repo] --state closed --json number,title,state,url --limit 10

Return: list of potential duplicates with title, status, URL, and similarity assessment

Agent: "duplicate-checker-jira" (if Jira is active)

Search Jira for potential duplicate issues.
Project: [CUSTOMIZE: YOUR-PROJECT-KEY]

1. Use WebFetch or jira CLI to search:
   JQL: project = "[CUSTOMIZE: YOUR-PROJECT-KEY]" AND text ~ "[keywords]" ORDER BY created DESC
2. Check both open and recently resolved issues

Return: list of potential duplicates with key, title, status, URL, and similarity assessment

Solo Mode (sequential)

Run the same duplicate checks sequentially for whichever tracker is active.

Phase 3: Report Duplicates (if found)

If potential duplicates are found, present them to the user:

I found [N] potential duplicate(s):

1. **[TITLE]** ([STATUS]) — [URL]
   Similarity: [High/Medium/Low] — [reason]

2. **[TITLE]** ([STATUS]) — [URL]
   Similarity: [High/Medium/Low] — [reason]

Options:
a) These are different — proceed with creating the new story
b) This is a duplicate of #[N] — skip creation
c) This is related to #[N] — create and link them

Wait for user confirmation before proceeding.

Phase 4: Format the User Story

Generate the story in standard user story format:

## Title
[Concise, descriptive title]

## User Story
**As a** [type of user],
**I want** [goal/desire],
**So that** [benefit/value].

## Description
[Expanded description of the feature, including context and background]

## Acceptance Criteria

### Scenario 1: [Happy path scenario name]
- **Given** [precondition]
- **When** [action]
- **Then** [expected outcome]

### Scenario 2: [Alternative scenario name]
- **Given** [precondition]
- **When** [action]
- **Then** [expected outcome]

### Scenario 3: [Edge case or error scenario name]
- **Given** [precondition]
- **When** [action]
- **Then** [expected outcome]

## Technical Notes
[Any implementation hints, constraints, or dependencies — if applicable]

## Out of Scope
[What this story explicitly does NOT cover — helps prevent scope creep]

Present the formatted story to the user for review before creating it.

Phase 5: Create in Tracker

After user approves the story content:

Linear

1. Use ToolSearch to load Linear tools
2. Use mcp__claude_ai_Linear_2__get_team to find the target team
3. Use mcp__claude_ai_Linear_2__list_issue_labels to find appropriate labels
4. Use mcp__claude_ai_Linear_2__save_issue to create the issue with:
   - title: [story title]
   - description: [full story in markdown]
   - team: [target team]
   - priority: [user-specified or default medium]
   - labels: [appropriate labels]

GitHub Issues

gh issue create \
  --repo [CUSTOMIZE: your-org/your-repo] \
  --title "[story title]" \
  --body "[full story in markdown]" \
  --label "[appropriate labels]"

Jira

Use WebFetch to POST to Jira REST API or use jira CLI:
- Project: [CUSTOMIZE: YOUR-PROJECT-KEY]
- Issue Type: Story
- Summary: [story title]
- Description: [full story in markdown/Jira wiki format]
- Priority: [user-specified or default Medium]
- Labels: [appropriate labels]

Vault Fallback (no tracker active)

Save to 04-projects/[project]/stories/story-YYYY-MM-DD-[slug].md

Phase 6: Confirm and Link

  1. Confirm creation with the user:

    Story created: **[TITLE]**
    [URL or file path]
    Priority: [priority]
    Labels: [labels]
    
  2. If the user indicated this is related to another issue (Phase 3, option c), create a link/relation between them.

  3. Also save a local copy to 04-projects/[project]/stories/ for vault reference.


Fallback Behavior

ScenarioBehavior
No tracker activeSave as markdown in 04-projects/[project]/stories/
Tracker API failsSave locally, warn user, suggest manual creation with the formatted content
Duplicate check failsWarn user that duplicate check was skipped, proceed with creation after confirmation
User doesn't specify projectList active projects from MY-PROFILE.md and ask which one
User provides minimal inputAsk clarifying questions to build a complete story

Error Handling

  • API rate limits: Wait and retry once, then fall back to local save
  • Authentication failures: Inform user their integration may need re-authentication
  • Large search results: Limit duplicate check to top 10 most relevant matches
  • Context overflow: Summarize duplicate results rather than showing full descriptions

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Deep strategic research engine — decomposes questions into parallel research threads, spawns multiple agents, and synthesizes into actionable strategic analysis

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

braindump

無料

Quick capture of raw thoughts with intelligent domain classification and competitive intelligence extraction

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

Run one task through the V-model verification loop: CP-2 plan → CP-3 build → CP-3v component verify → CP-4 integration verify (full lane) → CP-5 acceptance. The worker never grades its own homework; evidence rows trace back to AC-n. Opt-in: invoke with /closed-loop or by asking for the closed loop, proper verification, or an evidence trail. Ordinary work does not run this.

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

Deep-dive 7-day analysis across all data sources for weekly reviews, board prep, and strategic planning

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

Autonomous content pipeline - scout announcements in your field, triage by trend momentum and personal angle, produce posts/blogs/videos in your voice with ledger-based dedup, hard volume caps, and screenshot-verified publishing

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

Generate personalized news intelligence with verified sources (7-day freshness requirement)

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

huytieu のスキルをすべて見る

このスキルの問題を報告する