The AI tools we actually use day to day (and the ones we dropped)

A month ago we were generating blog hero images with Ideogram, an image generator that had been working fine for weeks. One day it started throwing a quota error that wasn’t the usual free-tier limit — it was telling us to upgrade the plan before generating a single image more. We switched tools that same afternoon. That pretty much sums up how we pick what we use: not by what the product’s landing page promises, but by what’s still working when you actually need it.

At Sinado Apps we use Claude Code for a lot more than writing code. We have a scheduled routine that drafts a blog post several days a week, translates it into English and Galician, and publishes it on its own, without anyone reviewing it before it goes live (if something sounds off, I fix it afterward, not before). That works because instead of repeating instructions every time, we’ve got them saved as what’s called “skills” in the jargon: a checkable style guide, a log of topics already published so we don’t repeat ourselves, an exact frontmatter format to follow. The AI doesn’t magically “know” how to write like us — it reads those rules every time it’s asked to write.

The hero images do get left pending by that automated routine, and that’s where MCPs come in — Model Context Protocol, the standard that lets an assistant talk to external tools without someone coding a custom integration for each one. We use Google’s Stitch through its MCP to generate hero images in an isometric style we defined once; the assistant sticks to it every time it generates a new image, without me having to describe the style again. We also have MCPs for Trello and Slack, though we’re moving slower there — the board is old and we haven’t cleaned it up enough yet to let an agent write to it unsupervised.

The part that gets talked about least but shows up most in day-to-day work is subagents: instead of letting a single conversation thread pile up an entire large project’s context until it runs out of useful memory, we hand off specific pieces (find where a function is defined, review a diff, write a commit message) to a separate agent that does just that one task and returns a short result. Sounds like a technical detail, but it’s the difference between working on a months-long project without the AI “forgetting” what we did at the start, and having to re-explain the context every other day.

I wouldn’t recommend setting all of this up for just any agency. If the team is one person and the content volume is low, an automated publishing routine costs more in upkeep than it saves. It pays off for us because we publish regularly and because a mediocre article is cheap to fix — it gets corrected, it doesn’t cost us a client’s project.