executing-plans
Executing Plans
Overview
Load plan, review critically, execute tasks in batches, report for review between batches.
Core principle: Batch execution with checkpoints for architect review.
Announce at start: "I'm using the executing-plans skill to implement this plan."
The Process
Step 1: Load and Review Plan
- Read the plan file completely
- Review it critically — this is where you catch flaws before they become expensive rework. Look for:
- Steps that seem vague or contradictory
- Missing dependencies or setup
- Assumptions that don't hold in your actual codebase
- Technical choices you'd question
- If concerns exist: Raise them with your partner before proceeding. Getting alignment now saves many hours of wrong execution
- If the plan looks solid: Create a todo tracker and proceed to Step 2
Step 2: Execute Batch
Default: First 3 tasks per batch
The checkpoint-driven approach (3 tasks, pause for feedback, continue) gives your partner visibility and prevents you from building on a flawed foundation.
For each task in the batch:
- Mark it as in_progress in your tracker
- Follow the plan's steps exactly — they're designed to be small and verifiable
- Run the specified verifications before marking done — this catches mistakes early
- If verification fails: stop, report it, and wait for guidance
- If verification passes: mark as completed and move to the next task
Keep focus tight. Executing 3 tasks well and getting feedback is faster than executing 10 tasks blind and discovering you went the wrong direction halfway through.
Step 3: Report
When the batch is done, report clearly:
- Show what was implemented (what changed, new files, modified code)
- Show verification output (test results, error messages if any)
- Summarize any unexpected findings or minor deviations from the plan
- End with: "Ready for feedback."
This pause-to-report is critical. It gives your partner a chance to course-correct before you invest more work.
Step 4: Continue
Your partner will respond. They might:
- Approve and ask you to continue (execute next batch)
- Point out an issue and ask for changes (go back to Step 1 or Step 2)
- Change the plan based on new information (return to Step 1)
Follow their guidance. The loop (Execute → Report → Feedback → Continue) prevents you from diverging.
Once you've completed all batches and your partner approves, move to Step 5 (Complete Development).
Step 5: Complete Development
After all tasks complete and verified, finalize the branch. Verify all tests pass, ensure the working tree is clean, and then present options to the user: merge to main, open a PR, or keep the branch for further review.
- Run the full test suite one final time
- Confirm no untracked or uncommitted changes remain
- Ask the user: merge now, open a PR, or leave for manual review
When to Stop and Ask for Help
Execution should pause for clarification when:
- You hit a blocker mid-batch (missing dependency, test fails, instruction unclear) — continuing blind wastes effort
- The plan has critical gaps preventing starting — you can't execute what doesn't exist
- You don't understand an instruction — guessing is slower than asking
- Verification fails repeatedly — this signals the plan may have flawed assumptions
Why this matters: Stopping early prevents wasted work and lets your partner give you the right guidance. It's faster to pause and clarify than to build on a wrong foundation.
When to Revisit Earlier Steps
Return to Review (Step 1) when:
- Your partner updates the plan based on your feedback — old review isn't valid anymore
- Fundamental approach needs rethinking — new information changes strategy
Forcing through blockers creates debt. A pause to reconsider is faster than fixing mistakes later.
Remember
- Review plan critically first
- Follow plan steps exactly
- Don't skip verifications
- Reference skills when plan says to
- Between batches: just report and wait
- Stop when blocked, don't guess
- Never start implementation on main/master branch without explicit user consent
Common Pitfalls
Agent Skips Review and Starts Executing Immediately
The problem: Jumping straight to Step 2 (Execute Batch) without doing Step 1 (Load and Review Plan). This skips your chance to catch plan flaws before wasting work.
Why it happens: Execution feels more productive than reading; the urge to "just start" is strong.
How to avoid it: Force yourself to read the plan, write down questions, and raise them before marking the first task in_progress. Your review step is the highest-leverage time to prevent rework.
Batch Size Too Large Causes Confusion
The problem: Trying to execute 10+ tasks in a batch makes it hard to track what's done, what's blocked, and where verification failed. You lose the checkpoint benefit.
Why it happens: Wanting to move fast, or misunderstanding "default: first 3 tasks" as a minimum instead of a safe default.
How to avoid it: Stick to 3 tasks per batch unless the plan explicitly says otherwise. After you report and get feedback, you'll know if you should do bigger batches. Smaller batches = clearer reporting = faster feedback loops.
Verification Failures Ignored
The problem: A verification fails (test doesn't pass, output doesn't match expected), but you mark the task completed anyway and move on. This compounds — later tasks depend on earlier ones working.
Why it happens: The temptation to "mark it done" and push forward. Verification feels like optional polish.
How to avoid it: If verification fails, that task is not done. Stop the batch, report the failure, and wait for guidance. The plan may have a flaw or you may have misunderstood a step. Your partner needs to know.
Plan Becomes Stale Mid-Execution
The problem: You're halfway through executing, and your partner realizes the plan needs updating (scope changed, tech choice was wrong, new blocker discovered). But you keep executing the old plan.
Why it happens: You're focused on the tasks in front of you; you don't realize the plan has been superseded.
How to avoid it: After reporting each batch, read your partner's response carefully. If they mention plan changes, go back to Step 1 (Review) even if you're already partway through. Fresh context beats forward momentum.
Integration
Related workflow skills:
- writing-plans — Generates the plan that this skill executes. This skill assumes a well-structured plan exists; if you need to create one, refer to that skill first.
- brainstorming — Produces the design that feeds into writing-plans. Run brainstorming → writing-plans → executing-plans for the full workflow.
Git worktree tip: Before executing a plan, consider running in an isolated git worktree so your work is sandboxed from main:
git worktree add ../feature-work -b feature/my-feature
This prevents accidents on main and lets you abandon bad experiments cleanly.
More from ckorhonen/claude-skills
video-editor
Expert guidance for video editing with ffmpeg, encoding best practices, and quality optimization. Use when working with video files, transcoding, remuxing, encoding settings, color spaces, or troubleshooting video quality issues.
63tui-designer
Design and implement retro/cyberpunk/hacker-style terminal UIs. Covers React (Tuimorphic), SwiftUI (Metal shaders), and CSS approaches. Use when creating terminal aesthetics, CRT effects, neon glow, scanlines, phosphor green displays, or retro-futuristic interfaces.
35practical-typography
Professional typography guidance based on Matthew Butterick's Practical Typography. Use when evaluating, critiquing, or improving document formatting, text layout, font choices, punctuation, spacing, or any typography-related decisions for print or web content.
34app-marketing-copy
Write marketing copy and App Store / Google Play listings (ASO keywords, titles, subtitles, short+long descriptions, feature bullets, release notes), plus screenshot caption sets and text-to-image prompt templates for generating store screenshot backgrounds/promo visuals. Use when asked to: write/refresh app marketing copy, craft app store metadata, brainstorm taglines/value props, produce ad/landing/email copy, or generate prompts for screenshot/creative generation.
33markdown-fetch
Fetch and extract web content as clean Markdown when provided with URLs. Use this skill whenever a user provides a URL (http/https link) that needs to be read, analyzed, summarized, or extracted. Converts web pages to Markdown with 80% fewer tokens than raw HTML. Handles all content types including JS-heavy sites, documentation, articles, and blog posts. Supports three conversion methods (auto, AI, browser rendering). Always use this instead of web_fetch when working with URLs - it's more efficient and provides cleaner output.
26llm-advisor
Consult other LLMs (GPT-4.1, o4-mini, Gemini 2.5 Pro, Claude Opus) for second opinions on complex bugs, hard problems, planning, and architecture decisions. Use proactively when stuck for 15+ minutes or facing complex debugging. Use when user says 'ask Gemini/GPT/Claude about X' or 'get a second opinion'.
22