Handing over an unfamiliar project—AI was a big help, but it also cost a pretty penny.

Last month, I took over a project left behind by a former colleague. There’s virtually no documentation, and the submission information is very informal.

The most useful usage

I didn’t ask it to change anything at all. First, I spent half a day having it walk me through the project: what the project is about, how the main modules are divided, where the data comes in and goes out, and what the less-obvious conventions are.

I could spend half a day just figuring this out on my own—three days’ worth of work. The questions I’m asking are pretty much the same as those I’d ask a colleague who’s familiar with this area: How does this log flow? Why does this function call A instead of B? What edge cases does this class handle?

The cost is that the credit limit is dropping rapidly.

Because it needs to read a large number of documents in order to respond. And all these documents have been incorporated into my context, meaning that every subsequent message will have to carry them again.

By midday, I had already used up more than half of the current quota. At that moment, I felt a bit regretful and thought I should have been more careful with my spending.

Handing over an unfamiliar project—AI was a big help, but it also cost a pretty penny.

Last month, I took over a project left behind by a former colleague. There’s virtually no documentation, and the submission information is very informal.

The most useful usage

I didn’t ask it to change anything at all. First, I spent half a day having it walk me through the project: what the project is about, how the main modules are divided, where the data comes in and goes out, and what the less-obvious conventions are.

I could spend half a day just figuring this out on my own—three days’ worth of work. The questions I’m asking are pretty much the same as those I’d ask a colleague who’s familiar with this area: How does this log flow? Why does this function call A instead of B? What edge cases does this class handle?

The cost is that the credit limit is dropping rapidly.

Because it needs to read a large number of documents in order to respond. And all these documents have been incorporated into my context, meaning that every subsequent message will have to carry them again.

By midday, I had already used up more than half of the current quota. At that moment, I felt a bit regretful and thought I should have been more careful with my spending.

Later, it was discovered that there was a more efficient way.

Leave the research to the sub-agent. It reads through the stack of documents in its own independent window and sends me only the conclusions. This keeps my main context clean, making each subsequent message inexpensive.

I’m using this approach a lot these days, especially when taking over unfamiliar code. It achieves the same result but consumes several times less resources.

But handing over tasks like this really is a big-time undertaking.

No matter how much you optimize, understanding a new project still requires reading a lot of material. This kind of work tends to be concentrated in just a few days, and you don't need to use that much information on a regular basis.

When I’m short on funds now, I’ll top up a bit from Asale. It’s a token-sharing network where others’ idle token quotas are listed; pricing is based on the amount of tokens input and output, and the page shows what percentage of the original price it corresponds to. When I checked, GPT-5.6 was around 49%, while Claude was at about 60%, and both figures fluctuate somewhat.

As for the risk section, it’s pretty straightforward: The request is forwarded through someone else’s client, and the body of the message is visible to that end user. Sharing subscriptions might violate upstream terms and conditions. When I came across this passage, I actually felt a bit more reassured—there aren’t many products that are willing to put unfavorable information right on the homepage. Of course, we still need to avoid anything that really should be avoided; don’t send confidential information here at all.

You can find it on GitHub by searching for "asale." It’s compatible with Mac, Windows, and Linux, and the source code is open.

Later, it was discovered that there was a more efficient way.

Leave the research to the sub-agent. It reads through the stack of documents in its own independent window and sends me only the conclusions. This keeps my main context clean, making each subsequent message inexpensive.

I’m using this approach a lot these days, especially when taking over unfamiliar code. It achieves the same result but consumes several times less resources.

But handing over tasks like this really is a big-time undertaking.

No matter how much you optimize, understanding a new project still requires reading a lot of material. This kind of work tends to be concentrated in just a few days, and you don't need to use that much information on a regular basis.

When I’m short on funds now, I’ll top up a bit from Asale. It’s a token-sharing network where others’ idle token quotas are listed; pricing is based on the amount of tokens input and output, and the page shows what percentage of the original price it corresponds to. When I checked, GPT-5.6 was around 49%, while Claude was at about 60%, and both figures fluctuate somewhat.

As for the risk section, it’s pretty straightforward: The request is forwarded through someone else’s client, and the body of the message is visible to that end user. Sharing subscriptions might violate upstream terms and conditions. When I came across this passage, I actually felt a bit more reassured—there aren’t many products that are willing to put unfavorable information right on the homepage. Of course, we still need to avoid anything that really should be avoided; don’t send confidential information here at all.

You can find it on GitHub by searching for "asale." It’s compatible with Mac, Windows, and Linux, and the source code is open.