The real cost of AI tools isn't the subscription — it's the repetition
I've been thinking about why AI tools feel expensive even when the per-token cost seems low.
The answer: repetition overhead.
Every time you run a task, the AI re-reads the context, re-plans the steps, re-explores the environment. Even if you've done the exact same thing 50 times before. The AI doesn't remember. You pay full price every time.
For a solo founder running the same research, scraping, or content workflows weekly, this adds up fast.
The fix isn't a cheaper model. It's eliminating the repetition overhead entirely.
That's the core of what I've been building with AllyHub:
- Manuals: The agent learns a website once, saves it, reuses it forever
- Playbooks: Multi-step workflows saved as one-click pipelines
- Skills: Domain knowledge that compounds over time
Real example: a task that cost 123 credits the first time (posts + author profiles on X) cost 32 credits the fourth time — same job, 4x more output.
The subscription cost is visible. The repetition tax is invisible. But it's the bigger number.
Would love to hear how other indie hackers are thinking about AI cost efficiency.
this is a really sharp way to frame it
what stood out to me is that most people try to solve this at the “AI level” (better prompts, better models), but the repetition usually comes from the workflow itself
like the AI isn’t really the problem, it’s that the system around it isn’t structured to reuse anything properly
we had something similar where even with saved prompts, things still felt repetitive because every run was slightly disconnected from the previous one
what helped more was stepping back and defining the flow itself as something reusable, not just the prompt or the task
i mapped a few of these flows in something like stackely just to see the full chain, and it made it obvious that a lot of repetition wasn’t necessary, it was just baked into how the process was set up
feels like once the flow compounds, the AI part naturally becomes cheaper without changing the model at all
Exactly right - the repetition is baked into the process, not the AI. That's why fixing it at the prompt level never fully works. The flow itself has to be reusable. That's the core of what we built with AllyHub - the workflow structure compounds, not just the individual prompts. Once the flow is defined, every run gets faster without any extra effort from the user.
The repetition overhead is real. I run automated workflows on the same tasks weekly — scraping, analysis, summarization — and the cost isn't the tokens, it's rebuilding context every single run. Here's what this site does, here's what to look for, here are the edge cases. The AI re-learns all of it from scratch each time.
My workaround has been manual: context docs the agent reads at session start, pre-built prompts for recurring tasks. It works but it's fragile — any change in the workflow and the docs go stale faster than you update them.
The compounding knowledge angle is what I haven't solved well yet. Re-running the same job and having it cost a fraction on the fourth run would genuinely change which tasks feel worth automating.
The context docs approach is exactly the right instinct - you've essentially built a manual version of what AllyHub automates. The fragility you're describing (docs going stale when the workflow changes) is the exact problem we solve. In AllyHub, the 'context docs' are built automatically by the agent as it works, and they update when the environment changes. The fourth run being a fraction of the first run cost is exactly what we're seeing in our data too.