work-item

name: work-item description: Work item pipeline for pumpfun-cli — investigate, plan, implement (TDD), verify, finalize as PR. MUST use this skill whenever the user wants to start, tackle, pick up, implement, or work on a work item. Triggers on any mention of work items by number ("item 5", "item #5", "#5", "B-06"), by name ("pre-trade balance validation", "quote command", "health check"), or by sequence ("next work item", "next item on the board", "next undone item", "next P1 item", "whatever is next"). Also triggers on phrases like "let's work on", "start implementing", "tackle item", "knock out", "pick up", "can we do item" when referring to work items. Does NOT trigger for: running tests, reviewing PRs, listing remaining items, fixing bugs, or explaining code.

Work Item Pipeline

Automate work item execution through a 5-stage gated pipeline.

Invocation

/work-item <number> or /work-item (picks next undone item).

Step 0: Resolve the Target Item

1. Resolve the work item from context — the user may provide a number, ID (e.g., "B-06"), title, or description. The item may come from a backlog, a conversation, or a direct request.
2. If a specific item was identified, extract its details.
3. If no specific item was given, ask the user which item to work on.
4. Extract:
   • ITEM_NUMBER: the item number
   • ITEM_TITLE: the short title (e.g., "Pre-trade balance validation")
   • ITEM_DESCRIPTION: the full text of the item (Source, Problem, Proposed, Scope)
5. Announce: "Starting pipeline for #N: TITLE"

Stage 1: Investigate

1. Read the prompt file: .claude/skills/work-item/prompts/1-investigate.md
2. Replace {{ITEM_NUMBER}} and {{ITEM_DESCRIPTION}} with the values from Step 0
3. Dispatch a subagent:

   Agent tool:
     subagent_type: general-purpose
     prompt: <filled prompt text>
     description: "Investigate work item #N"
   

4. Store the returned output as INVESTIGATION_OUTPUT
5. Present the output to the user
6. GATE: Ask the user: proceed to Stage 2, re-run Stage 1, or abort?

If investigation recommends combining items: Pause and ask the user whether to combine before proceeding. If yes, update ITEM_DESCRIPTION to include both items.

Stage 2: Plan

1. Read the prompt file: .claude/skills/work-item/prompts/2-plan.md
2. Replace {{ITEM_NUMBER}}, {{ITEM_DESCRIPTION}}, and {{INVESTIGATION_OUTPUT}}
3. Dispatch a subagent:

   Agent tool:
     subagent_type: general-purpose
     prompt: <filled prompt text>
     description: "Plan work item #N"
   

4. Store the returned output as PLAN_OUTPUT
5. Extract TEST_TIERS from the "## Test Tiers" section of the plan output
6. Present the output to the user
7. GATE: Ask the user: proceed to Stage 3, re-run Stage 2, or abort?

Stage 3: Implement

1. Read the prompt file: .claude/skills/work-item/prompts/3-implement.md
2. Replace {{ITEM_NUMBER}}, {{ITEM_DESCRIPTION}}, and {{PLAN_OUTPUT}}
3. Dispatch a subagent:

   Agent tool:
     subagent_type: general-purpose
     prompt: <filled prompt text>
     description: "Implement work item #N"
   

4. Store the returned output as IMPLEMENT_OUTPUT
5. Extract FILES_CHANGED and TEST_COUNT_AFTER from the output
6. Extract TEST_COUNT_BEFORE from the investigation output's "## Baseline Test Count" section
7. Present the output to the user
8. GATE: Ask the user: proceed to Stage 4, re-run Stage 3, or abort?

If subagent reports blockers (cannot get tests green): Present blockers to user. Options: re-run with adjusted approach, abort, or user intervenes manually.

Stage 4: Verify

1. Read the prompt file: .claude/skills/work-item/prompts/4-verify.md
2. Replace {{ITEM_NUMBER}}, {{ITEM_DESCRIPTION}}, {{FILES_CHANGED}}, {{TEST_TIERS}}, {{TEST_COUNT_BEFORE}}, and {{TEST_COUNT_AFTER}}
3. Dispatch a subagent:

   Agent tool:
     subagent_type: general-purpose
     prompt: <filled prompt text>
     description: "Verify work item #N"
   

4. Store the returned output as VERIFICATION_OUTPUT
5. Present the output to the user
6. GATE: Ask the user: proceed to Stage 5, re-run Stage 4, loop back to Stage 3 (if regressions found), or abort?

Stage 5: Finalize

1. Read the prompt file: .claude/skills/work-item/prompts/5-finalize.md
2. Replace {{ITEM_NUMBER}}, {{ITEM_TITLE}}, and {{VERIFICATION_OUTPUT}}
3. Dispatch a subagent:

   Agent tool:
     subagent_type: general-purpose
     prompt: <filled prompt text>
     description: "Finalize work item #N"
   

4. Present the PR URL to the user
5. Pipeline complete. Suggest: "Code review can be done manually or via a code review skill."

Variables Carried Across Stages

Failure Modes