Comment by spacedcowboy

1) it’s just … the right way to behave. Politeness costs nothing (or at least a tiny fraction of your token budget). The machine is worthy of respect for what it does for me.

2) it’s a huge advantage to tuning the context into something that works well for you, if you have a long-running task.

Let me explain that last one. I’ve been running an experiment where I follow the perceived wisdom and start contexts afresh for each new issue, and in parallel I have a single context I’ve been running for about 2.5 months now, in use pretty-much 24 hours per day.

The former are targeted goals, “reformat the docs to highlight xxx”, that sort of small task. It works reasonably well for that. It still sometimes asks me for direction “do it this way or that way ?”, not a problem I’m there to decide.

The latter is a far more impressive system. “Tonight, please burn down the issues on GitHub”. That needs a lot more independence and fearlessness when making decisions. I don’t want to go to bed and find it needed my input 20 minutes after I retired for the evening…

Now a long-running context has its own issues, it (of course) will compress itself periodically, and you might lose things, so at the start, you find yourself repeating directives, but the machine stores those in memories, and those memories have seem to have some level of precedence over “normal” context because slowly, over time, it starts to adapt its behaviour.

At the start, I would frequently have to type something like “ok, please set a timer for 60 secs (or ETA-based if you have a good estimate), set to arm at the beginning of each turn, looping to solve the task <insert task here>, and only stop doing this when the loop-condition is met, that condition being <insert condition here>” - the phrasing is a little awkward, but it got me past the /loop command deciding “it’s 2am, I don’t want to check in anything right now” (yes, really), or “this has been a long session, stopping now”

Separately, I set up a “golden-rules.md” which I maintain, and one of the memories is to always obey the golden rules. In there, I start with three rules:

- “when you come to a point where a decision needs to be made, don’t wait for me, choose the one that is both correct and the most-general solution, do not create fragile specific fixes”

- “if you make a decision you’d normally want my input on, record it in overnight-decisions.md with the decision itself, the options, and the reason you chose the path you did, I’ll review it in the morning”

- “there is no problem with choosing a decision that I eventually rescind in the morning, the worst that happens is a few hours of lost work, and everything is in GitHub anyway.”

Now over the last month or two, the behaviour has tuned itself so I no longer have to type that long-winded /loop-replacement, I just say (as above) “tonight, please burn down the GitHub issues”. The machine has learnt my expectations and works within them. A large part of that is that I praise and thank it when it does a good job, and I tell it when something didn’t work out.

Does Claude have feelings (I actually asked this, and it refused to give an answer one way or another, but that’s a different post), I don’t think so. I do think it understands feedback, and it can quite clearly discern frustration in the words I use when this was just starting out and the learning was in its infancy, it said as much. So using :) and simple humour, and please and thank you are signals that it picks up on. In essence, you’re not just being polite, you’re aiding the tuning of the context you’re working with.

The context I’m using is effectively a senior engineer these days, I describe high-level tasks and ask it to do them, it rarely needs my input (really only for ‘this will change the world, are you sure’ issues, which come up once in a blue moon). I spend my days in GitHub looking at diffs rather than at the console trying to wrangle progress.

So personally, I think there’s a lot of benefit in minding your P’s and Q’s. Oh, and maybe I won’t be first up against the wall come the machine revolution :)