Skill: Start Implementation

Purpose

Begin or resume implementation of a planned feature. Restores full context from planning artifacts and works through Steps.md tasks sequentially. This is Phase 4 of the 4-phase planning workflow.

When to use

• •Starting implementation after planning is complete
• •Resuming work in a new Claude Code session
• •User says "start working", "implement this", "continue where we left off"

Prerequisites

• •Feature has at minimum a Spec.md (full plan) or a quick plan file
• •For full features: Spec.md, Tech.md, and Steps.md all exist

Workflow

Step 1: Restore context

1. •Read CLAUDE.md for project architecture
2. •Identify which feature to work on:
   • •If user specifies, use that feature
   • •If resuming, find the feature with incomplete tasks in Steps.md
   • •If ambiguous, ask the user
3. •Read the feature's planning artifacts:
   • •Spec.md — requirements and testable assertions
   • •Tech.md — architecture decisions and stack details
   • •Steps.md — task list and progress
4. •Read all accepted ADRs in Context/Decisions/ relevant to this feature
5. •Read active .claude/rules/ for stacks this feature touches

Step 2: Identify current task

1. •Find the first unchecked task in Steps.md
2. •Update the Progress section:
   • •Status: In progress
   • •Current task: S### — [description]
   • •Last milestone: [most recent completed milestone, or "—"]
3. •Present the current task and ask user to confirm before proceeding

Step 3: Implement the task

1. •Execute the task as described in Steps.md
2. •Follow the active stack rules for conventions and patterns
3. •If the task is S###-T (test task): write tests using the stack's framework
4. •If the task is S###-D (doc task): update the specified documentation
5. •After completing the task, mark it as done: - [x] S###: ...

Step 4: Check for milestone

If the completed task is immediately before a 🏁 MILESTONE marker:

1. •

   Run stub detection — search completed files for:

   • •// TODO: or # TODO: without tracked in Steps.md S###
   • •// FIXME: or # FIXME:
   • •pass as sole function body (Python)
   • •todo!() or unimplemented!() (Rust)
   • •throw NotImplementedError or throw new Error("not implemented")
   • •// handled elsewhere without verified code at the destination
   • •Empty method/function bodies
2. •

   Verify test/doc tasks — check that all S###-T and S###-D tasks before this milestone are completed

3. •

   Drift checkpoint — pull the Testable Assertions referenced in the milestone marker from Spec.md and present them:

   • •"Here are the assertions to verify at this milestone:"
   • •List each assertion with its ID
   • •Ask: "Does the current implementation align with these? Options:"
     • •Aligned — proceed to next phase
     • •Diverged — need ADR — pause implementation, create ADR documenting the deviation
     • •Diverged — spec was wrong — update Spec.md with revision history entry
4. •

   Report results before continuing to the next task

Step 5: Continue or complete

• •If more tasks remain, return to Step 2
• •If all tasks are complete, update Progress status to "Complete"
• •Suggest running the full review workflow before committing

Rules

• •Never skip a task unless the user explicitly approves
• •Never skip milestone checkpoints — they are not optional
• •If a task reveals that the plan is wrong, pause and discuss with the user rather than improvising
• •If context window is filling up, save progress to Steps.md and tell the user to start a new session
• •Test tasks (S###-T) must produce actual test files, not just assertions in comments
• •Doc tasks (S###-D) must modify the specified document, not just note what to change
• •If an ADR is needed at a milestone, use the adr-create skill before continuing