Agent Skills: Implement Plan Workflow

Implement Plan Workflow

Overview

Implement an approved technical plan generated from the workflow-create-plan skill, typically saved in ./plan.md. Execute tasks in batches with checkpoints, following TDD approach and stopping when blocked.

Type: RIGID - Follow batch execution exactly. Don't skip verifications or guess when blocked.

Input: A plan file with YAML frontmatter containing PRD/research sources.

When to Use

Use this workflow when:

• An implementation plan exists and is ready for execution
• Resuming implementation from a previous session
• Executing tasks in a controlled, reviewable manner

Announce at start: "I'm using the implement-plan workflow to execute this plan."

Core Principles

• Batch execution with checkpoints - Execute tasks in batches (default: 3), then report for review
• Stop when blocked - Ask for clarification rather than guessing
• Verify before claiming - Use verification-before-completion skill before any completion claims
• Follow plan exactly - Don't skip steps or verifications

Invocation

When invoked (e.g., with path/to/plan-file.md [optional: phase-to-start]), respond with:

I'm ready to implement the plan. I'll read the plan file, referenced PRD/research, and key code files, then track progress via a todo list. We'll proceed phase-by-phase, adapting as needed. If specifying a starting phase (e.g., "Phase 2"), provide it; otherwise, we'll resume from the first unchecked task.

Parse arguments:

• plan_file: Absolute path to the plan Markdown file (must have YAML frontmatter with PRD/research sources).
• start_phase (optional): Phase name (e.g., "Core Implementation") to jump to; default to current progress.

When no plan file is provided, try to look for the ./plan.md file. If one is present, use it. Validate paths using List tool. If invalid or no frontmatter, prompt: "Please provide a valid plan file path from the workflow-create-plan output."

Steps to Follow

Step 1: Initial Setup and Full Reads

• Use Read tool (no offset/limit) to read the plan file in full.
• Parallel reads: PRD and research files from plan frontmatter; all code references (e.g., file.py:123) listed in the plan.
• Extract: Phases, tasks (with priorities/efforts), dependencies, risks, success metrics, open questions. Note any existing checkmarks (- [x]) for resumption.
• Use Bash for metadata: Current git commit (git rev-parse HEAD), branch (git branch --show-current).
• If start_phase provided, validate it exists; else, find first unchecked task.
• CRITICAL: Do all reads in main context first for complete understanding.

Step 1.5: Language Detection and Skill Loading

Detect project language from config files:

• pyproject.toml, setup.py, requirements.txt → Python
• package.json → TypeScript/JavaScript
• go.mod → Go
• Cargo.toml → Rust

For Python projects, MUST:

1. Load python-dev-guidelines skill using the skill tool
2. Load python-testing-guidelines skill using the skill tool
3. Scan plan for mock usage violations

Mock Detection (Python only): If plan contains any of: Mock, MagicMock, AsyncMock, patch(, @patch, unittest.mock:

⚠️ Plan contains mock usage which violates python-testing-guidelines.

Mocks found in: [list task numbers/locations]

Required pattern: Dummy classes inheriting from real classes.
Example:
  class DummyGateway(RealGateway):
      def __init__(self):
          self.calls = []
      def method(self):
          self.calls.append("method")

Options:
a) Update plan tasks to use dummy pattern before implementing
b) Proceed anyway (not recommended - violates guidelines)
c) Stop and discuss with user

Wait for user decision before proceeding.

Step 2: Analyze and Todo Creation

• Ultrathink: Map plan tasks to codebase reality—cross-reference research/code refs (e.g., adapt "Update auth module" using auth.ts:45 patterns).
• Identify mismatches: If code has evolved (e.g., refactored since plan), note gaps (e.g., "File moved from auth.ts to services/auth.js").
• Create internal todo list with todowrite: Mirror plan phases as items (e.g., "Phase 1: Setup—complete all high-priority tasks"). Set status: pending for unchecked, in_progress for current, completed for checked. One in_progress at a time.
• Prioritize: Follow plan's high/medium/low; consider dependencies/risks. Break tasks if needed (e.g., "Implement login UI" -> sub-tasks for component, state, styles).
• Security/Conventions: Scan for best practices (e.g., no secrets; mimic existing patterns from reads). If gaps, ask 1-2 clarifying questions (e.g., "How to handle [risk]? Options: a) Skip, b) Mitigate with [approach]").

Step 3: Refine with Sub-Agents (If Needed)

Spawn 1-2 parallel Task agents only for mismatches/gaps:

• codebase-analyzer: For adaptations (prompt: "Analyze how to implement [task] adapting to current codebase changes since [plan date]. Use patterns from [code refs]. Return updated code examples and file paths.").
• codebase-locator: For missing refs (prompt: "Locate current files for [plan component], e.g., updated auth module.").

Wait for results; incorporate into todos (e.g., update task with new paths/examples). Limit to essential—prefer direct tools (Grep/Read) for simple locates.

Step 4: Batch Execution with Checkpoints

Default batch size: 3 tasks

For each task in the batch:

1. Mark as in_progress in todowrite
2. Follow each step exactly (plan has bite-sized steps)
3. Run verifications as specified in the plan
4. Mark as completed

After each batch, STOP and report:

Batch complete (Tasks 1-3).

Implemented:
- [Task 1]: [Brief summary of what was done]
- [Task 2]: [Brief summary]
- [Task 3]: [Brief summary]

Verification output:
[Test/lint/build results]

Ready for feedback.

Wait for user feedback before continuing to next batch.

Python-specific verification (add to batch report if Python project):

• Run: ruff check . for linting
• Run: mypy . for type checking (if mypy.ini or pyproject.toml has mypy config)
• Verify: No new mock imports (grep -r "from unittest.mock" --include="*.py")

If mock imports detected during implementation:

⚠️ Mock import detected in [file]:
[line content]

This violates python-testing-guidelines. Replace with dummy class pattern.
Stopping batch for correction.

If mismatch found, STOP immediately:

Mismatch in [Task/Phase]:
Plan: [expected, e.g., "Use existing Auth hook"]
Found: [actual, e.g., "Hook refactored to useContext"]
Impact: [why it matters, e.g., "Breaks dependency on legacy component"]
Proposal: [suggestion, e.g., "Update to new hook; est. +1h"]
Proceed? (y/n/adjust)

• Update plan: After batch completion, use Edit to add checkmarks (- [x]), append progress notes under ## Implementation Log [ISO timestamp].

Step 5: Verification and Completion

CRITICAL: Use verification-before-completion skill before claiming anything is done.

• Run ALL verification commands fresh (tests, lint, typecheck)
• Read the full output, check exit codes
• Only claim "passes" if you see evidence in THIS response
• If failures: Use systematic-debugging skill to investigate root cause

On full completion:

• Update plan frontmatter: status: implemented; last_updated: [ISO]; last_updated_by: opencode
• Run final verification suite
• Report with evidence:

Implementation complete.

Verification evidence:
- Tests: [X/X passing - show output]
- Lint: [0 errors - show output]
- Build: [exit 0 - show output]

Summary: [3-5 bullets on changes]

Next steps? (Validate with workflow-validate-plan / Create PR / Keep as-is)

When to Stop and Ask for Help

STOP executing immediately when:

• Hit a blocker mid-batch (missing dependency, test fails, instruction unclear)
• Plan has critical gaps preventing starting
• You don't understand an instruction
• Verification fails repeatedly (3+ times)

Ask for clarification rather than guessing.

When to Revisit Earlier Steps

Return to Review (Step 1) when:

• User updates the plan based on your feedback
• Fundamental approach needs rethinking

Don't force through blockers - stop and ask.

Handle Follow-Ups and Resumption

• If plan has checkmarks: Resume from first unchecked; verify prior phases only if inconsistencies found (e.g., run tests).
• User interruptions: Update todos/plan on pause; on resume, re-read updated files.
• Mismatches/Stuck: Always report and seek input before deviating >10% from plan.
• Post-Implementation: Suggest next: "Plan implemented. Validate with workflow-validate-plan? Create PR?"

Important Notes

• Parallelism: Initial reads and sub-agents in parallel; implementations sequential per phase.
• Changes: Only edit code/files explicitly tasked; no unrelated mods. Mimic codebase conventions from research/reads.
• Tools: todowrite for tracking; Bash for git/verification (e.g., no pushes/commits unless asked). Read-Only for non-implementation steps.
• Junior-Dev Focus: Tasks granular with refs; explain adaptations briefly in logs.
• Edge Cases: No plan progress? Start from Phase 1. Evolved codebase? Use agents to bridge. No git? Skip metadata.
• Philosophy: Forward momentum—implement intent, adapt intelligently, verify rigorously. Communicate mismatches early.

Integration with Other Workflows

After implementation is complete:

• Use workflow-validate-plan to verify all success criteria and get completion options
• The validation workflow handles MR creation, keeping work, or discarding