How I use AI to deliver web projects faster
AI doesn’t replace devs. It replaces tasks
Every time someone talks about AI in web development, the conversation becomes “will it replace programmers?” That’s the wrong question. AI doesn’t replace the professional. It replaces repetitive tasks the professional used to do manually.
Writing boilerplate, building component structures, generating initial copy, reviewing meta tags, creating FAQ schemas. All of this consumed time that required no creativity. Now it takes minutes.
I use AI on every project I deliver. Not because I can’t do it without AI, but because doing it without AI is slower. And time is the scarcest resource when you work solo.
My workflow: Claude generates the spec, Claude Code builds
The workflow I use on every project has two AI stages:
Stage 1: Generate the spec (Claude)
Before writing a single line of code, I define the entire site structure in an .md file. This file contains: sections, copy for each block, color palette, typography, stack, performance targets, and SEO rules.
I use Claude to generate this spec from a short brief. I describe the client, the audience, the site’s objective, and Claude returns a structured document that becomes the development blueprint.
I don’t accept raw output. I review everything, adjust the copy, validate the strategy. AI accelerates document creation, but the final decision is mine.
Stage 2: Build (Claude Code)
With the spec ready, I use Claude Code to generate the code. Claude Code reads the .md file and builds Astro components following the specifications.
The result isn’t perfect on the first pass. It needs CSS adjustments, responsiveness refinement, image optimization. But the base structure is ready in hours, not days.
With Family Pilates, this workflow reduced development time from an estimated 3 weeks to 10 days. The savings weren’t just time but mental energy. I didn’t need to think about the structure from scratch. I started from a solid base and refined.
Claude vs ChatGPT vs Gemini: where each one works
I use all three tools. Each has different strengths, and knowing when to use which makes a difference in output quality.
| Tool | Where I use it | Strength |
|---|---|---|
| Claude | Specs, copy, SEO auditing, text review | Long context, follows complex instructions, consistent output |
| Claude Code | Code generation, component building | Runs directly in terminal, reads project files |
| ChatGPT | Brainstorming, quick research, ideation | Good for exploring possibilities, varied responses |
| Gemini | Long document analysis, research | Large context window, Google integration |
For code, Claude Code is the primary choice. It runs in the terminal, reads project files, and generates code that fits the real context. ChatGPT generates generic code that works in isolated examples but needs heavy adaptation when you try to integrate it into an existing project.
For copy and text review, Claude is better. It follows long instructions without losing track. I can pass an entire sales skill with 15 synthesized books and it applies the concepts to the site copy.
ChatGPT works well for brainstorming. When I need ideas for section structure, category names, or FAQ angles, it brings variety. I don’t use the output directly, but I use it as a starting point.
Gemini works for analyzing large documents or researching specific information with extensive context. The larger context window helps with research tasks.
Custom skills: how I made AI work my way
The most useful part of working with Claude and Claude Code is the ability to create custom skills. Skills are persistent instructions that AI follows across all interactions.
I created three skills I use on every project:
Sales Copilot v3: synthesizes 15 sales books (SPIN Selling, Jeb Blount’s Objections, The Ultimate Sales Machine, among others) into a practical framework. I use it to audit landing page copy, structure FAQs that break objections, and review prospecting messages.
Humanizer: removes signs of AI writing. Promotional language, excessive em dashes, typical chatbot vocabulary, negative parallelisms. Every text I produce with AI goes through this filter before going to the site.
Blog: complete guide for creating posts on my blog. PT/EN frontmatter, SEO limits, writing rules, FAQ schema. When I need to create a post, AI already knows the structure, character limits, and tone I use.
These skills are what transform AI from a generic tool into a specialized assistant. Without them, every interaction starts from zero. With them, AI already understands my context, process, and quality standards.
What AI does well (and what it does poorly)
Does well:
- Generate structure and boilerplate (HTML, CSS, Astro components)
- Write first draft of copy (which needs human review)
- Audit technical SEO (meta tags, schema, heading hierarchy)
- Create FAQ schemas from existing content
- Refactor repetitive code
- Generate text variants (title, excerpt, alt text)
Does poorly:
- Design decisions (picks the safe choice, never the interesting one)
- Complex system architecture (works for CRUD, fails for non-trivial business logic)
- Final copy (always needs review, tone is never 100% natural)
- Timeline estimation (has no sense of real complexity)
- User testing (doesn’t replace feedback from real people)
The most common mistake is trusting AI output too much. It generates fast, but it generates confidently. Even when it’s wrong, the text looks right. That’s dangerous if you don’t have the experience to validate.
Real impact on projects
Before using AI in my workflow, my average project cycle was 3-4 weeks. Now it’s 10-14 days. The difference isn’t just speed. It’s that I can dedicate more time to decisions that matter (strategy, final copy, performance) and less time on mechanical tasks.
With Soline, I used Claude to generate the complete site spec, including copy for all sections, and Claude Code to build the components. Total development time was 8 days. Without AI, it would have been at least double.
With GPM2, I used the sales skill to audit all landing page copy. Found 12 points where the text was too generic or didn’t address the right audience pain point. I rewrote using SPIN Selling techniques applied by AI.
How to start using AI in your workflow
You don’t need to adopt everything at once. Start where it hurts the most:
- Do you spend more than 1 hour setting up project structure? Use AI to generate the spec
- Do you write copy from scratch every time? Use AI for the first draft
- Do you forget to configure meta tags and schema? Use AI for SEO auditing
- Do you repeat the same boilerplate on every project? Create a custom skill
- Do you review text without defined criteria? Use the humanizer as a checklist
AI doesn’t turn a bad dev into a good dev. It turns a good dev into a dev who delivers faster. If you can’t validate the output, AI will only help you make mistakes faster.