Oh, interesting. My sense is that wouldn’t fully clear the context window? Either way, I can immediately think of a project I have in Claude that could use that sort of condensing.
I use the app. For whatever reason I do all my coding in Claude, all my random fact-finding in ChatGPT. I’m not sure how other LLM’s work in terms of organization, but I like very much in Claude that if I’m doing a bunch of coding that builds off other coding, I can store it all in a project and Claude retains that very specific context.
Yeah, mine too. I’ve used it occasionally when I chained too many tasks, but haven’t finished the last one. I want the LLM to still remember the things it’s done and the feedback I’ve given it for the latest task.
Now I try to keep each task separate, then clear the context before the next task. Claude Code will read your source and docs each time so anything I want to save is should be in the repo.
Sometimes I will explore or plan a task and determine that I don’t want to do it yet, but I want to remember the final plan. In this case I’ll have the LLM summarize and save it to a markdown file. The caveat is that LLM can see the markdown and use that to influence it’s actions. Thus I only save summaries that I intend to do in the future.
Finally, this is probably common knowledge, but the agents.md is really useful for setting top-level requirements as well as instructions on what to do each iteration.
For my current project I have it run lint, unit tests, export docs, commit, etc. automatically each time. It’s pretty good about knowing when it can skip a step.
I also see it reference my requirements during planning: “you said backwards compatibility shouldn’t be supported so…”
Today I asked it to make me a timeline for a federal grant. Actually, I asked it to just give me a template and said I’d fill it in myself. It not only ignored that part, but the timeline it made included components from a different grant.
This week I was searching for clipart. I needed an image of a momma and baby deer. In the search results, most of the deer in the AI images were wearing spots… both momma and baby. Some of the images included a spotty dad. Ugh. I was able to find what I needed, though the project has been postponed.
Also (this has certainly been commented on within the last 1000 or so posts), I was in the mood for some light reading and chose to read the openAI user terms&policies. Le sigh. Their invasive use of my contact list stands out in the things I found unnegotiable. We’ve a tough road ahead of us as natural beings living in an increasingly artificial world.