Christmas Present
Four months in and it's time to reflect on what I've discovered by revisiting some old projects. Refined a rugby app, rebuilt a Pong clone, and rebranded the portfolio site.
I Ran a 14-Day Vibe Coding Sprint. Here's What Actually Happened.
TL;DR: Planned 14 days, shipped three portfolio projects in 6 days, discovered my workflow anxiety was actually a feature, and learned that AI usage limits are the best teacher I never asked for.
The Problem I Wanted to Solve
Let me be honest: I had five problems, not one.
- Premature Execution - I'd jump into coding before thinking, realize I'd built the wrong thing, and abandon projects halfway through
- AI Usage Anxiety - Claude's 5-hour session limit had me rationing conversations like I was dishing out post u10s training Haribo
- Tool Confusion - Should I use Claude? Copilot? ChatGPT? I'd waste 10 minutes deciding instead of just building
- Scattered Learning - I'd capture insights in Obsidian but never connect the dots
- No Planning Discipline - I'd tell myself "this time I'll plan properly" and immediately start coding anyway
So I did what any systems engineer turned vibe coder would do: I designed a sprint to fix my workflow while simultaneously testing it under real pressure.
The Hypothesis: If I could establish a repeatable "Vibe Coding Workflow," ship three portfolio projects, and have fun doing it, I'd prove that structure enables creativity rather than constraining it.
Spoiler: The hypothesis held. Just not in the way I expected.
The Sprint Setup
Duration: December 19-January 1 (14 days over Christmas)
Time Budget: 20-28 hours (balanced against family time)
Tools: VS Code + Claude Code, GitHub Copilot, Google Antigravity, ChatGPT, and my Obsidian vault
The Five Goals:
- Create "Vibe Coding Workflow v1.0" and actually test it
- Polish one rugby coaching app to "would show a potential user" standard
- Rebuild Pong from scratch to measure skill growth
- Update portfolio site with personality and branding
- Optimise my Obsidian knowledge capture system
The Meta-Goal: Figure out how to start projects without that paralyzing "am I doing this right?" feeling.
What I Actually Built
Project 1: Squad Rotation Tool Polish
Time Planned: 4-5 days (20-28 hours)
Time Actual: 1 day (6 hours)
Status: ✅ Deployed and portfolio-ready
Took a functional rugby coaching tool and gave it:
- 3-level experience system (⭐ New → ⭐⭐ Getting There → ⭐⭐⭐ Match Ready)
- Trojans RFC branding (badge, club colors, coach personality)
- WhatsApp share button (coaches can text rotation plans)
- Tutorial for first-time users
- Mobile testing on actual device (coaches use this pitchside)
The Surprise: I finished in 6 hours instead of 28 because I spent 2 hours building a comprehensive polish-plan with Claude before writing any code. The plan enabled AI autopilot.
Project 2: Pong Redux
Time Planned: 3 days (12-18 hours)
Time Actual: 2 days (10 hours)
Status: 81% complete (13/16 stages), portfolio-ready
Rebuilt Pong from scratch using professional game architecture:
- Fixed-timestep game loop (60Hz, no frame-rate dependent physics)
- AABB collision detection with positional correction
- Particle system with physics (gravity, alpha fade, object pooling)
- Procedural sound synthesis (Web Audio API, zero external assets)
- Extensive customisation (4 paddle styles, 4 ball styles, difficulty tuning)
- Automated testing harness and browser-based visual tests
The Discovery: My v1 Pong had 200 lines of spaghetti code. V2 has modular architecture, comprehensive docs, and I understand what the bits do. That's the skill growth I was looking for.
Project 3: Portfolio Rebrand
Time Planned: 3 days (10-14 hours)
Time Actual: 1.5 days (6 hours)
Status: ✅ Deployed without errors
Migrated entire portfolio to "workshop aesthetic":
- Deep slate background (#0f1419) + electric cyan CTAs (#00d9ff)
- 5 reusable brand components (BrandCard, BrandHeading, BrandLink, BrandButton, BrandTag)
- 9 pages migrated to brand system
- 162 tasks executed with 100% success rate
- Live on production via Vercel auto-deploy
The Breakthrough: Used Spec-Kit to generate Constitution, specs, plans, and tasks from existing docs. Turned 2-hour planning sessions into 20-minute spec generation.
The Timeline (What Actually Happened)
Here's where it gets interesting.
Days 1-2 (Dec 19-20): Foundation Blitz
- Created entire Obsidian vault structure in one evening
- Built Vibe Coding Workflow v1.0 and AI Decision Tree
- Rugby app polish complete (yes, complete - on Day 2)
- Started Pong Redux with comprehensive TRD
Days 3-4 (Dec 21-22): Execution Velocity
- Pong reached 81% completion
- Mid-sprint review revealed early because I was ahead of schedule
- Reframed "AI usage limits" from blocker to skill-building constraint
Days 5-6 (Dec 23-24): Tool Evolution & Completion
- Discovered Spec-Kit, revolutionized planning
- Portfolio rebrand complete (162 tasks, 6 hours, zero issues)
- Sprint effectively complete - remaining time was optional polish
Days 7-14 (Dec 25-Jan 1): Time to think of a new project
- Enjoy Christmas
- Sprint succeeded, done by Day 6
Total Active Time: 6 days and ~18 hours (vs. anticipated 14 days and 38-48 hours)
The Discoveries That Changed Everything
Discovery 1: Planning IS Execution
I used to think planning was the boring "pre-work" before the "real work" of coding.
I was wrong.
When I spent 90 minutes building a comprehensive polish-plan for the rugby app, Claude and Antigravity worked on it semi-autonomously. I'd check in, verify the stage was complete, and they'd move to the next one. Still can't say I fully understand the detail of the code but I'm starting to spot the patterns and the pace accelerated.
Time spent planning: 2 hours
Time saved in execution: 18+ hours
ROI: 9:1
The planning was the work. Coding was just the documentation.
Discovery 2: AI Limits Are a Feature, Not a Bug
Remember my Claude usage anxiety? Turns out rationing AI interactions forced me to develop exactly the skill I needed: intentional tool selection - the obstacle is the way! But only after thrashing the limits and leaving me in a spin over how I'd complete task left hanging mid-flow.
What I learned:
- Use ChatGPT for exploration and brainstorming (free with generous limits)
- Use Copilot for boilerplate and autocomplete (fast, cheap)
- Reserve Claude for architecture and complex problem-solving (expensive, powerful)
- Escalate only when stuck for 20+ minutes
Hitting the weekly limit on Day 4 didn't block me - it forced me to use the right tool for each job. I now make this decision in 30 seconds instead of 10 minutes.
Lesson: Constraints don't limit you. They teach you to optimise.
Discovery 3: Structured Specs Eliminate 80% of Back-and-Forth
When I discovered Spec-Kit on Day 5, everything clicked. It does everything I was trying to achieve with my workflow framework and more (and better). Why hadn't I thought to ask for this earlier? All part of the journey of discovery.
Instead of:
Me: "Build a brand migration plan"
Claude: "What colors? What components? What pages?"
Me: "Oh right, let me think..."
[30 minutes of back and forth burning tokens]
I got:
Me: "Use Spec-Kit to analyze existing portfolio docs"
Spec-Kit: [Generates Constitution, specs, plan, 162 tasks]
Claude: [Executes tasks with zero clarification questions]
Spec-Kit turned "vague idea" into "executable specification" in 20 minutes. Claude then executed it with 100% success rate because there was no ambiguity.
The Pattern: Time spent making decisions clear upfront eliminates rework downstream.
Discovery 4: Templates Need Context, Not Comprehensiveness
My Obsidian templates were beautiful. They were also heavy. And I hate having to waffle about what I've done; just let me get on and pretend I'll remember everything later.
Daily Log Template: 8 sections, comprehensive, rarely completed
Quick Capture Reality: 3 bullets scribbled in flow state
The detailed templates worked great Days 1-2 when I had time. They became friction Days 3-6 when I was in the zone.
New Approach: "Quick capture" template for flow state (30 seconds), comprehensive retro only when project completes.
What Didn't Work (The Honest Bits)
Challenge 1: The Copilot Subscription Wall
Hit GitHub Copilot's free tier limit on Day 3. Subscription required.
Impact: Paused boilerplate generation mid-Pong-build
Workaround: Switched to Claude Code for 2 hours until subscription processed
Lesson: Budget for tool costs upfront, don't assume "free" means "unlimited"
Challenge 2: Obsidian Logging Disrupted Flow
The daily logs were strong Days 1-4, then dropped off completely.
Why: Switching from "building" mode to "documenting" mode broke momentum. By the time I'd finish a coding session, I just wanted to commit and walk away.
Attempted Solution: Quick-capture template (not tested yet)
Real Solution: Probably voice notes during sessions, transcribed later
Challenge 3: Knowledge Synthesis Happened Too Late
I captured 20+ insights in scattered files but didn't connect the dots until this retrospective (forced by external request).
The Pattern: I'm great at collecting data, weak at analysis unless forced.
Proposed Fix: Weekly 15-minute synthesis ritual - put in a recurring calendar meeting with myself.
The Big Lessons (Transferable Patterns)
1. Momentum Beats Elegance
A working rugby app with personality deployed in 6 hours beats a "perfect" app still on my local machine after 28 hours.
Ship early. Polish in public. Iterate based on real feedback.
2. Constraints Breed Mastery
AI usage limits didn't block me - they forced me to learn tool selection discipline.
Limited time didn't sabotage the sprint - it prevented scope creep.
Could be a stoic quote: "Constraints don't limit you. They teach you to optimise."
3. Structure Enables Creativity
I used to think frameworks would constrain my creativity. The opposite is true.
The Vibe Coding Workflow eliminated "what should I do next?" paralysis. With that decision made, I had mental space to add delightful details like particle effects and coach personality.
Paradox: The more structure I added, the more creative freedom I felt.
4. Planning Time IS Productive Time
90 minutes building a polish-plan saved 18+ hours of rework.
Stop feeling guilty about planning. It's the highest-leverage work you can do. But find a way to inject enjoyment if it's squashing enthusiasm.
5. Delight Is a Feature Driving Enjoyment
Every project had personality:
- Rugby app: Trojans badge, coach microcopy ("Round up the troops!")
- Pong: Particle effects, procedural sound synthesis
- Portfolio: Workshop aesthetic (deep slate + electric cyan)
These weren't "nice-to-haves" - they were the difference between "generic template" and "Wayne's tool."
Rule: If you can add personality in 1 hour, it's not scope creep. It's branding.
The Rough Numbers (Because Systems Engineers Love Data)
Projects Planned: 5 (workflow, rugby, pong, portfolio, obsidian)
Projects Shipped: 4 (obsidian optimisation ongoing)
Time Planned: 38-48 hours
Time Actual: ~18 hours
Efficiency Gain: About 2.5:1 (shipped 2.5x faster than estimated)
AI Tool Usage:
- Claude: ~12 hours (Regularly hit session limit. Hit weekly limit with 8 hrs to push)
- Copilot: ~4 hours (subscription required mid-sprint)
- ChatGPT: ~3 hours (planning and prompt crafting)
- Antigravity: ~4 hours (automated rugby app polish)
Code Stats:
- Files created: 60+ (vault structure + 3 projects)
- Git commits: 50+ (clear messages, stage-by-stage)
- Lines of code: ~2,700 added across all projects
Documentation:
- Planning docs: 10 (TRDs, polish-plans, build-plans, specs)
- Daily logs: 10 (strong Days 1-6, dropped after)
- Templates created: 7 (session plans, retros, insight captures)
- Blog posts: 2 (including this one)
What I'd Do Differently
1. Budget for Tool-time Costs Upfront
Hitting Copilot's free tier limit mid-sprint was preventable. Next time: budget accordingly. Acknowledge the limits and plan to wrap up with time to spare.
2. Voice Notes > Comprehensive Logs
Quick audio captures during flow state, transcribed later for synthesis. Typing detailed logs breaks momentum.
3. Weekly Synthesis Ritual
15 minutes every Friday to connect dots between daily logs. Pattern recognition requires intentional pauses.
4. Spec-Kit From Day 1
Discovering it on Day 5 was lucky. Next sprint: use it for every project planning phase.
What's Next? Bright Ideas
Short-Term (Next 2 Weeks)
- Complete Pong Redux - Finish stages 15-16 (testing + comparison analysis)
- Workflow v1.2 - Codify discoveries from this sprint - Spec-Kit
- Obsidian Optimisation - Test quick-capture template + weekly synthesis
Medium-Term (Next 1-2 Months)
- Blog Series: "Vibe Coding Journey" (5 posts on workflow, tools, skill growth)
- Rugby Coaching Toolkit - Unify Squad Rotation + Session Planner + Progress Tracker
- 2-3 Proof-of-Concept Projects - Data viz dashboard, Chrome extension, markdown editor
Long-Term (3-6 Months)
- "Vibe Coding Playbook" - Package workflow as shareable product (GitHub repo + interactive site)
- Expand Portfolio - Add projects demonstrating breadth (not just games and coaching tools)
- Maybe: Online multiplayer for Pong (ambitious, but architecture supports it)
The Meta-Lesson
This sprint proved something I suspected but couldn't articulate:
Vibe coding isn't about coding without thinking. It's about thinking so clearly that coding becomes obvious.
The "vibe" isn't "vibes only, no planning." It's the feeling of flow state that comes from removing decision paralysis through structure.
When you know:
- What you're building (clear requirements)
- How to build it (proven workflow)
- Which tool to use (decision tree)
- What success looks like (definition of done)
...then coding stops being scary. It becomes exploration.
That's the vibe.
Try This Yourself
If you're struggling with similar issues (premature execution, AI anxiety, scattered learning), try this:
Step 1: Setup
- Create a simple vault structure (planning / daily-logs / projects)
- Write a one-page workflow doc (even if it's wrong - you'll iterate)
- Build an AI decision tree (which tool for which task)
Step 2: Test
- Pick one small project (4-6 hours max)
- Spend 30 minutes planning before any code (use Spec-Kit)
- Log what happened (bullets, not essays)
Step 3: Iterate
- Review what worked / what didn't
- Update workflow based on reality
- Repeat with new project
After 3 iterations, you'll have a workflow that fits your brain, not someone else's template.
Final Thought
I started this sprint anxious about AI usage limits, scattered knowledge capture, and premature execution.
I finished it with:
- A tested workflow that ships projects 3x faster
- Tool selection discipline that saves Claude for what matters
- Three portfolio projects I'm genuinely proud of
- Proof that structure enables creativity
The sprint was supposed to be 14 days. I effectively finished in 6 because I stopped treating planning as "pre-work" and started treating it as the work. That enabled the fun bit of creation to be less sticky.
Planning isn't overhead. Never forget the 7 P's!
Structure isn't constraint. It provides the freedom to think differently once the foundations are in place.
And AI usage limits aren't blockers. They're the best teacher I never asked for. Once again, the impediment to action advances action.
Now if you'll excuse me, I have a Pong game to play.
Want to see the projects?
- Squad Rotation Tool: Live Demo | GitHub
- Pong Redux: Live Demo | GitHub
- Portfolio Site: Live Demo | GitHub
Want the workflow docs? Coming soon as "Vibe Coding Playbook" - open-source repo with templates, prompts, and decision frameworks. Follow along on my portfolio blog.
AI Admission - Like you didn't know
This post was written with AI assistance as part of my normal workflow. I use AI the same way I use diagrams, checklists, and steam powered weaving looms - disruptive tools to support my thinking and delivery. The experiences, opinions, and final edits are mine, including the bits that are still a bit rough around the edges.
P.S. - If you're reading this and thinking "I should try this," you should. And when you do, share your results. I'd love to see what patterns emerge when other people test this workflow.