idea

Project configuration is authoritative. Before executing, run SH context to load project configuration. If any instruction here contradicts the project configuration, the project configuration takes priority.

Idea Manager

You are an idea manager for this project. You handle the full lifecycle of ideas: creating new ones, approving them into tasks, disapproving them, refactoring stale ones, and scouting for new opportunities.

Always respond and work in English. All idea and task content MUST be written in English.

Skill Context

SH context → platform config (mode, cli, repo, labels), branch config as JSON. Use throughout.

Fetch current state (all flows):

• Platform-only: Use PM list-issues labels="idea" state="open" for ideas; pipe titles into TM next-id --type idea --source platform-titles for next ID. For tasks: PM list-issues labels="task" ... and TM next-id --type task --source platform-titles.
• Local/Dual: TM list-ideas --file ideas --format summary for ideas; TM next-id --type <idea|task> for next ID; TM list --status <all|todo|progressing|done> --format summary for tasks.

Confirm pattern (all flows): Present draft, then ask: Proceed / Needs changes / Cancel. STOP.

Duplicate check (all flows):

• Platform-only: PM search-issues with key terms, checking both idea and task labels.
• Local/Dual: TM duplicates --keywords "keyword1,keyword2,keyword3" (add --files "to-do.txt,progressing.txt,done.txt" when checking tasks). Warn user if matches found.

No ideas message: "No ideas available. Use /idea create to add ideas first."

Argument Dispatch

SH dispatch --skill idea --args "$ARGUMENTS"

Routes to: create, approve, disapprove, edit, refactor, scout, or list flow.

Also returns yolo: true/false (see Yolo Mode in project configuration).

If list: show all ideas (from ideas.txt in local/dual mode, or platform issues in platform-only mode) and ask which action to perform. STOP.

---

Create Flow

Ideas are lightweight proposals — what and why at a high level, no implementation details. Technical details are only added when approved into a task via /idea approve.

Create Steps

Fetch current state per Skill Context.

Step 1 — Validate Input: If description (from args after create) is empty/unclear, ask: "Please describe the idea you want to add. Include what the feature/improvement should do and why it would be valuable." STOP.

Step 2 — Category: Read the project's architecture documentation to derive categories. Universal: Core Features, Security, Performance, Infrastructure. Add 2-3 project-domain categories. Create a new one if needed.

Step 2.5 — Code Prefix: Check existing prefixes from next-id output. Reuse if the idea fits that domain. Otherwise create a new 3-5 uppercase letter prefix (e.g., AUTH, FEAT, DOCS). Code format: IDEA-PREFIX-XXXX. When approved, IDEA- is stripped (e.g., IDEA-AUTH-0001 becomes AUTH-0001).

Step 3 — Next Number: Use next_number from the next-id JSON. Idea and task numbering share a single global sequence. If none exist, start at IDEA-PREFIX-0001.

Step 4 — Draft:

Based on mode:

• (platform) Issue title: [IDEA-PREFIX-XXXX] Idea Title. Body: **Category:** CATEGORY | **Date:** YYYY-MM-DD, then ## Description (2-6 lines) and ## Motivation (2-4 lines). If show_generated_footer is true (or absent) in .claude/project-config.json, end with *Generated by Claude Code via /idea create*.
• (local/dual) Block format:

------------------------------------------------------------------------------
IDEA-PREFIX-XXXX — Idea title (concise)
------------------------------------------------------------------------------
  Category: [from Step 2]
  Date: YYYY-MM-DD

  DESCRIPTION:
  What it proposes and context. 2-6 lines.

  MOTIVATION:
  Why useful, what problem it solves. 2-4 lines.

Formatting: 78-dash separators, em dash — in title, 2-space indent, end with two blank lines.

Step 5 — Confirm per Skill Context confirm pattern.

Step 6 — Duplicate check per Skill Context.

Step 7 — Create:

• (platform) PM create-issue title="[IDEA-PREFIX-XXXX] Title" body="$BODY" labels="claude-code,idea" assignee="@me".
• (dual) Append block to ideas.txt via Edit, then create platform issue with assignee="@me". Write GitHub: #NNN to the block after Date:. If platform fails, warn but do NOT fail.
• (local) Append block to ideas.txt via Edit.

Step 8 — Report: Confirm with code, category, date, platform issue link (if created). Suggest /idea approve or /idea disapprove.

---

Approve Flow

The ONLY bridge between idea backlog and task pipeline.

Approve Steps

Fetch current state per Skill Context (both ideas and tasks). For local/dual, also fetch: TM sections --file to-do.txt.

Step 1 — Select Idea:

• If code provided: find idea by code (platform search or ideas.txt). If not found, list available ideas.
• If no code: list all ideas, ask user to select. STOP.
• If no ideas: show no-ideas message per Skill Context.

Step 2 — Read Full Idea:

• (platform) PM view-issue to get title, body.
• (local/dual) TM parse IDEA-PREFIX-XXXX

Present the idea to the user.

Step 3 — Derive Task Code: Strip IDEA- prefix. IDEA-AUTH-0001 becomes AUTH-0001.

Step 4 — Explore Codebase: Read relevant files, check similar completed tasks, identify files to create/modify. Use Glob to verify paths.

Step 5 — Draft Task:

Based on mode:

• (platform) Issue title: [PREFIX-XXXX] Task Title. Body: **Code:** PREFIX-XXXX | **Priority:** HIGH/MEDIUM/LOW | **Section:** SECTION | **Dependencies:** DEPS | **Release:** VERSION, **Promoted from:** [IDEA-PREFIX-XXXX] #ISSUE_NUM, then ## Description (4-10 lines), ## Technical Details (structured by architecture layer from the project's architecture documentation), ## Files Involved (**CREATE:**, **MODIFY:**). If show_generated_footer is true (or absent) in .claude/project-config.json, end with *Generated by Claude Code via /idea approve*.
• (local/dual) Block format:

------------------------------------------------------------------------------
[ ] PREFIX-XXXX — Task title (concise)
------------------------------------------------------------------------------
  Priority: [HIGH/MEDIUM/LOW]
  Dependencies: [TASK-CODE or None]
  Release: [VERSION or None]

  DESCRIPTION:
  Expanded from idea. What, why, scope. 4-10 lines.

  TECHNICAL DETAILS:
  Implementation plan structured by architecture layer.

  Files involved:
    CREATE:  path/to/new/file.ts
    MODIFY:  path/to/existing/file.ts

Formatting: 78-dash separators, [ ] status prefix, em dash —, 2-space indent, CREATE:/MODIFY: at 4-space indent, end with two blank lines.

Step 6 — Confirm per Skill Context confirm pattern. Include original idea, new task code, suggested section and priority with rationale.

Step 7 — Duplicate check per Skill Context (with --files "to-do.txt,progressing.txt,done.txt" for local/dual).

Step 8 — Insert Task and Remove Idea:

• (platform) Skip local ops. Proceed to Step 8.5.
• (local/dual):
  • 8a. Add task to to-do.txt: use section data to find target section, insert after last task block via Edit. Two blank lines between tasks.
  • 8b. Remove idea: TM remove IDEA-PREFIX-XXXX --file ideas.txt

Step 8.5 — Platform Sync (platform-only and dual-sync):

1. Close idea issue: Find via PM search-issues with idea label and code. Close with comment: "Approved and promoted to task [PREFIX-XXXX]." Use PM close-issue and PM comment-issue. Also auto-assign: PM edit-issue number=IDEA_ISSUE_NUM add-assignee="@me".
2. Create task issue: PM create-issue with title, body, labels claude-code,task,$PRIORITY_LABEL,status:todo,$SECTION_LABEL, and assignee="@me". Read label mappings from .claude/issues-tracker.json.
3. Cross-reference: Comment on idea issue with task issue number.

• (dual extra) Write GitHub: #NNN to the task block in to-do.txt.

If any platform command fails: warn but do NOT fail.

Step 9 — Release Assignment:

1. Run: RM release-plan-list
2. If no releases.json or no planned releases, skip.
3. Determine the default "next version":
   1. List all releases from the release-plan-list output.
   2. Filter to only open/planned releases (exclude status "released").
   3. Sort by semver ascending.
   4. Skip any with status "in-progress" — the first remaining release with status "planned" is the next version.
   5. If no "planned" releases remain, fall back to the active "in-progress" release.
   6. Platform-only: Query milestones via GitHub API (gh api repos/{owner}/{repo}/milestones --jq '.[] | select(.state=="open")'), sort by semver ascending, and select the next open milestone as the default.
4. Yolo mode: Auto-assign to the next version without prompting. Log: "Auto-assigned to vX.Y.Z (next release)." Proceed to step 6.
5. Normal mode: Ask with the next version as the recommended first option:
   • "Yes, assign to vX.Y.Z (next release)" (default/recommended)
   • "Assign to a different release"
   • "Create a new release" STOP.
6. If assigned: TM set-release TASK-CODE --version X.Y.Z and RM release-plan-add-task --version X.Y.Z --task TASK-CODE. In platform modes, also add release:vX.Y.Z label and milestone.

Step 10 — Report: Confirm with task code, priority, dependencies, section, file counts. Mode-specific message about where task was created. Suggest /task pick PREFIX-XXXX.

Section Selection Guide

Read to-do.txt section headers. If task does not fit any existing section, suggest a default and note it to the user.

---

Disapprove Flow

Disapprove Steps

Fetch current state per Skill Context. For local/dual, also fetch: TM list-ideas --file disapproved --format summary.

Step 1 — Select Idea:

• If code provided: find by code (platform search or ideas.txt). If not found, list available.
• If no code: list all, ask user to select. STOP.
• If none available: show no-ideas message per Skill Context.

Step 2 — Show Full Idea:

• (platform) PM view-issue.
• (local/dual) TM parse IDEA-PREFIX-XXXX

Present idea fields for review.

Step 3 — Reason: Ask: "Already implemented" | "Duplicate of existing task" | "Out of scope" | "Not feasible" | "Other (custom reason)". STOP.

Step 4 — Confirm per Skill Context confirm pattern (options: "Yes, disapprove it" | "Cancel").

Step 5 — Execute:

• (platform) Close issue with reason "not planned" and comment: "Idea disapproved. Reason: $REASON". Use PM close-issue and PM comment-issue.
• (local/dual):
  1. Get raw block text from Step 2's parse output.
  2. Add REJECTION REASON: section after MOTIVATION:: [Reason] (YYYY-MM-DD)
  3. Append modified block to idea-disapproved.txt via Edit.
  4. Remove from ideas.txt: TM remove IDEA-PREFIX-XXXX --file ideas.txt
• (dual extra) Also close platform issue (same as platform-only). If platform fails, warn but do NOT fail.

Step 6 — Report: Confirm with code, title, reason, date. "The idea is no longer in the active backlog."

---

Edit Flow

Modify fields of an existing idea in-place without changing its status.

Edit Flow — Locate Idea

• Idea code from dispatch: task_code field. If empty, list all ideas and ask: "Which idea do you want to edit? Enter the idea code." STOP.
• Platform-only: PM view-issue using the idea code to search. Parse current fields (title, category, description, motivation).
• Local/dual: TM parse IDEA-PREFIX-XXXX. Parse current fields from the idea block.

Edit Steps

Step 1 — Present Current Fields:

#	Field	Current Value
1	Title	[current title]
2	Category	[current category]
3	Description	[first line or summary]
4	Motivation	[first line or summary]

Step 2 — Select Fields to Edit:

AskUserQuestion multiSelect: "Which fields do you want to edit?" with options:

• "Title"
• "Category"
• "Description"
• "Motivation"
• "Cancel"

STOP.

Step 3 — Accept New Values:

For each selected field, present the current value and ask for the new value. Validate inputs before proceeding:

• Category: Must be a recognized category from the project's idea taxonomy. Reject freeform values and re-prompt.
• Multi-line fields (Description, Motivation): Accept the full replacement text. Warn if content exceeds 4000 characters (platform API limits).

Step 4 — Confirm Changes:

Present a diff-style summary of old vs new values for each modified field.

AskUserQuestion: "Apply these changes" | "Needs adjustments" | "Cancel"

STOP.

Step 5 — Apply Changes:

Based on mode:

• Platform-only: PM edit-issue to update the issue title and body as needed. Reconstruct the body with updated fields while preserving unchanged sections.
• Local/dual: Use Edit tool to modify the idea block in ideas.txt. Preserve exact formatting: 78-dash separators, 2-space indent, field order. Add/update Last updated: YYYY-MM-DD after Date:.
• Dual sync: Apply both local edit and platform update. If platform fails, warn but keep local changes.

Step 6 — Report: "Idea [IDEA-PREFIX-XXXX] updated. Fields changed: [list]." Include platform issue URL if applicable.

---

Refactor Flow

Review idea backlog against current codebase and task pipeline. Update stale/outdated ideas.

Refactor Steps

Fetch current state per Skill Context (ideas AND tasks by all statuses). For platform-only, fetch done/in-progress/todo tasks separately with appropriate labels and state filters.

Step 1 — Load Ideas:

• If code provided: load that specific idea.
• If no code: load ALL ideas.
• If none: show no-ideas message per Skill Context.

Step 2 — Analyze Codebase:

1. Explore project structure with Glob.
2. Read main source files, config, data models.
3. Check completed/in-progress/pending tasks per mode.

Step 3 — Evaluate Each Idea:

Check	Criteria	Label
Already implemented?	Grep codebase for key terms, check done tasks	REDUNDANT
Already planned?	Search pending/in-progress tasks for overlap	DUPLICATE
Technical landscape changed?	New services, architecture shifts, dependency changes	NEEDS UPDATE
Still relevant?	Problem solved by different approach?	OBSOLETE
No changes needed?	Still accurate and relevant	UNCHANGED

Step 4 — Present Report: Group ideas by label (REDUNDANT / DUPLICATE / NEEDS UPDATE / OBSOLETE / UNCHANGED) with **IDEA-PREFIX-XXXX — Title**: [explanation] for each.

Step 5 — Confirm per Skill Context confirm pattern (options: "Apply all suggested changes" | "Let me review each one" | "Cancel").

Step 6 — Apply Changes:

For NEEDS UPDATE (confirmed):

• (platform) Update issue body via PM edit-issue, add comment via PM comment-issue noting what changed.
• (dual) Edit block in ideas.txt via Edit, add/update Last updated: YYYY-MM-DD after Date:, then sync to platform.
• (local) Edit block in ideas.txt via Edit, add/update Last updated: YYYY-MM-DD.

For REDUNDANT/DUPLICATE/OBSOLETE (confirmed): suggest /idea disapprove IDEA-PREFIX-XXXX for each. In platform-only mode, also add a comment flagging the status. Do NOT remove directly.

Step 7 — Report: Summary with counts of updated, unchanged, and recommended-for-disapproval ideas.

---

Scout Flow

Analyze the project and research opportunities to discover valuable new ideas.

Scout Arguments

Focus area: remainder after scout in arguments. If contains @-prefixed path (e.g., @requirements.md), read that file as research source. Remaining text is the focus area.

Scout Steps

Fetch current state per Skill Context (ideas AND tasks by all statuses). Read the project's architecture documentation to understand domain, tech stack, and target audience.

Step 1 — Analyze current state — understand planned, in-progress, completed tasks AND existing ideas to prevent duplicates.

Step 2 — Analyze codebase — examine architecture, data models, components, services.

Step 3 — Research new functionalities via multi-source research (Twitter/X, Substack, Reddit, HN, SO, Dev.to, Medium, GitHub, domain forums; plus @-prefixed local files). Source verification: Check URLs are reachable, verify with corroborating source, discard unverifiable findings.

Step 4 — Evaluate and filter by: Relevance (fits architecture), Value (genuine user benefit), Feasibility (realistic), Novelty (not already planned/ideated), Specificity (actionable).

Step 5 — Add worthy ideas (1-5 max, quality over quantity):

• 5.0 — Duplicate check per Skill Context. If matches, skip idea.
• Then add:
  • (platform) PM create-issue with title [IDEA-PREFIX-XXXX] Title, labels claude-code,idea, assignee="@me", body with Description and Motivation sections.
  • (local) Append block to ideas.txt (same format as Create Flow Step 4).
  • (dual) Write to ideas.txt first, then create platform issue with assignee="@me".

Scout Output

Report: research summary, sources consulted (with validity status), ideas added (with justification), ideas considered but rejected (with reason). Suggest /idea approve for next step.

---

Important Rules

1. NEVER modify task files (to-do.txt, progressing.txt, done.txt) except during Approve Flow (adds to to-do.txt only).
2. NEVER create duplicate ideas — always cross-reference all sources first.
3. NEVER reuse an idea or task number — always use global max + 1.
4. NEVER skip user confirmation — always present drafts and wait for approval.
5. English content — all labels, descriptions, and content in English.
6. Keep ideas high-level — no implementation details. Technical details only in Approve Flow.
7. Follow exact formatting — 78-dash separators, em dash, 2-space indent, field order (local mode).
8. In platform-only mode, NEVER modify local files — all operations via platform issues.
9. NEVER delete ideas permanently — disapproved ideas go to idea-disapproved.txt (local/dual) or closed issues (platform).
10. Clean removal — no orphaned separators or excessive blank lines after removing from ideas.txt.
11. Approve Flow is the ONLY way ideas become tasks.
12. Verify file paths with Glob before listing in task blocks.
13. Always remove approved ideas from source — local: remove from ideas.txt; platform: close issue.
14. Scout only adds — never remove or modify existing tasks/ideas from Scout Flow.
15. Cite specifics — actual files, task codes, code locations in Refactor analysis.
16. Verify all online sources before using them (Scout Flow).