My Real AI Workflow as a PHP & WordPress Developer: 7 Assistants, One Stack

Plenty of articles compare AI assistants in the abstract; this one just documents what I actually use. I run a freelance practice building WordPress plugins, Laravel applications and e-commerce stores for international clients, and in 2026 AI tools are threaded through every stage of that work. Here is the real stack — what each tool does, what it costs, and the parts of the job that stayed stubbornly human.

Development: Claude Code, all day

The core of my technical work runs through Claude Code. Real examples from recent client projects: refactoring a legacy plugin to WordPress coding standards across dozens of files; building a Laravel API integration test-first; tracing a WooCommerce checkout bug through template overrides and hook ordering. The workflow is consistent — I write a clear ticket, the agent plans and edits, tests run, I review the diff like a senior reviewing a mid-level developer. Review is non-negotiable: it writes plausible bugs with total confidence, just less often than any alternative I have tested. This is the engine behind my development services being faster than they were two years ago without quality dropping.

Research: Perplexity before anything factual

Client asks whether to move from Shopify to WooCommerce? Current plugin vulnerability landscape? Hosting comparison for a Laravel app? Every factual question starts in Perplexity because the citations let me verify before I advise. The answer then feeds a Claude-drafted recommendation with my judgment layered on. Research that took an afternoon of tabs now takes twenty minutes — and is better sourced.

Client operations: ChatGPT as the office assistant

Proposals from rough scope notes, contract-clause explanations, meeting summaries, awkward-email drafting ("the polite version of no"), quick spreadsheet formulas — ChatGPT is my back office. Custom GPTs hold my proposal structure and pricing logic so repeat documents start 80% done. Could Claude do this too? Mostly yes — but ChatGPT's breadth (voice on the move, quick images for mockups, agent errands) keeps it in the stack.

Content and SEO: the pipeline, not a button

This site's articles run on the workflow from my content-and-SEO comparison: Perplexity research → written brief → Claude draft → my experience and examples injected → edit → technical SEO pass. The posts rank because the experience is real; AI just removes the blank-page tax. Client site content follows the same pipeline under my SEO service.

The supporting cast

• Gemini — whole-codebase questions on unfamiliar legacy projects (the context capacity is unmatched) and anything tied to a client's Google Workspace.
• DeepSeek — cheap API muscle inside client automations: classification, extraction, summarisation at volumes where frontier pricing would sting.
• Copilot & Grok — situational: Copilot when a client's environment is Microsoft-locked; Grok when a project needs live social sentiment.

What it costs and what it returns

Roughly $60–70/month across subscriptions plus modest API usage. Against billable time, the stack pays for itself in the first day of each month — conservatively, it compresses 25–30% of my working hours into review-and-judgment time instead of production time.

What stayed human

Client trust, scoping judgment, architecture decisions with long consequences, saying "no" to bad ideas, and accountability when something breaks. AI made these more central, not less — when production gets cheap, judgment is the product.

Mistakes I made getting here

The stack looks tidy now; the road was not. I subscribed to everything at once in the early days — five tools, none learned properly, most cancelled. One tool learned deeply beats five sampled. I trusted output speed over review discipline exactly once on client work: an AI-generated migration script with a subtle date-handling bug cost a weekend and taught me that review is not overhead, it is the job. I used free tiers for client work before reading a single data policy — nothing went wrong, which is the worst kind of lesson; the two-line data rule exists in my practice because of that unearned luck. And I resisted agentic tools for months ("faster to do it myself") until a forty-file refactor proved otherwise. If you are earlier on this road: pick two tools, give them real work, review everything, and read the data policy the same day you sign up.

Frequently asked questions

How long did it take for AI to actually speed you up?

Honestly? Two to three months per tool. The first weeks are slower — you are learning to delegate and to review. The compounding starts when project setup (instructions, conventions, custom GPTs) makes every subsequent task start warm.

What would you keep if you could keep only one tool?

Claude — my output is code and writing, and it is the best at both. Your answer follows from your output, which is the whole method of this article.

Do clients care that you use AI?

They care about outcomes, timelines and confidentiality. I am transparent about the toolchain, contractual about the data handling, and accountable for every deliverable regardless of what drafted it. Framed that way, it has only ever been received as professionalism.

What's the next tool you're watching?

Less a tool than a pattern: agents that operate across my existing tools (email, project management, invoicing) rather than living in a chat tab. The assistants above all point that direction; whichever executes it reliably first earns a slot in next year's version of this article.

Bottom line

The stack is boring on purpose: specialist for code, engine for research, generalist for ops, pipeline for content. If you are building your own, start with the full comparison, subscribe to two tools, and give them your real work for a month — the right stack reveals itself through use, never through reviews. Even mine.