Three tools left my stack this quarter: a meeting summarizer, a grammar checker, and a "second brain" note-taking app. Each had a job. Each had an alternative sitting inside the model I already pay for. Here is what I learned from each exit.
Before the accounting: I want to be precise about what I mean by "prompt library." It is not a fancy system. It is a plain text file — one section per recurring task, each section containing a brief and a sample output. I open it, copy the relevant block, paste my content, and run it. The library currently has eleven entries. Three were written this quarter to replace the tools I cancelled.
Meeting summarizer — cancelled after fourteen months
The tool worked. That's what made it hard to notice the problem. Every meeting generated a summary, every summary got filed, every file got forgotten. The output was reliable and useless.
The replacement is a single prompt that takes a meeting transcript and produces exactly the format I actually use: decisions made, open questions with owners, and commitments with dates. Nothing else. It took me twenty minutes to write and has needed two small edits in three months to stay current with how my meetings have changed shape.
The difference that actually matters: I wrote the format. The SaaS tool had a format I couldn't change — or could change, but only within a set of parameters someone else decided were the right parameters. My prompt produces the thing I actually need, updated when I need it updated, at no marginal cost.
The lesson: any tool that outputs a "structured summary" is making assumptions about what structure means for your work. Those assumptions are usually wrong in small ways that compound over time. Owning the format means owning the assumptions.
The cheapest SaaS is the one you realize you didn't need.
Grammar checker — cancelled after two years
I used a popular grammar tool for two years. It caught comma splices, flagged passive voice, suggested shorter sentences. It was good at what it did. What it did wasn't quite what I needed.
The problem is that a grammar tool has no idea what the text is trying to accomplish. It evaluates sentences in isolation. The model evaluates them in context — because I gave it the context. "Read this like an editor who knows it's a proposal for a conservative client who trusts formality" is not a possible instruction to a grammar checker. It is a normal instruction to a model.
The replacement is a turn at the end of every draft: "Read this as an editor. Fix only genuine errors of grammar and fact. Note any sentence that could be cut without losing meaning, but don't cut it — flag it and let me decide." That last clause matters. The model can cut, and it will cut if you let it. The cuts are usually grammatically sound and occasionally kill a sentence that was doing important work. Keep the judgment with yourself.
What I pay for the replacement: nothing extra. What I get: a better edit, with context, adjustable whenever the context changes.
Second brain — cancelled after six months
This one surprised me. I had convinced myself I needed a system for capturing and connecting ideas across projects. I bought the app, took the course, built the structure.
Six months later I ran an audit. Of everything I had captured, I had returned to less than ten percent of it. The captures were not worthless — the act of capturing was clarifying. But the app was building a library I was not reading.
The replacement is nothing. That is the actual lesson of this one. I still take notes, but I take fewer, and I take them in the tool I'm already using for the work. If an idea is worth keeping, it goes into the relevant project document. If it doesn't fit anywhere, I think about whether it's actually worth keeping. Usually it isn't.
The friction I removed by cancelling the app turned out to be productive friction. The harder it is to capture something, the more you have to believe in it.
The pattern across all three
In each case, the replacement was not primarily about saving money. The combined saving is real but not life-changing. What I actually recovered was maintenance capacity — the low-level cognitive work of keeping three separate systems current, each with its own interface, its own update cycle, its own set of things that could go wrong.
Every tool in your stack costs attention beyond the subscription fee. The subscription is the invoice. The attention is the tax. A tool that does seventy percent of what you need but lives inside a workflow you're already in is often worth more than a tool that does ninety percent and lives somewhere else.
The prompt library is not elegant. It is a text file. It costs nothing and requires nothing except the ten minutes it takes to write a new entry when something changes. That's the whole system. The bar for adding a dedicated tool is now much higher than it was a year ago, and I think that's correct.