Working with AI

I turned the way I use AI—chatting, organizing my journal, doing weekly reviews—into an open-source Agent Skill called Reflection Companion, so I can look back across conversations from different times and see which of my judgments have changed.

This day was a video on RedNote.

Looking at the GPT-6 Astra upgrade, what I care about more is whether a job gets done all the way through: by the fifth round of edits, is what the first four settled still there, and how many times do I have to step in?

I don’t need a third AI home base, but Gemini found a place for itself in Chrome: the current page is the context, so there’s no copying, screenshotting or explaining “what I’m looking at,” and I end up asking lots of small questions on the spot.

For over a year I barely used ChatGPT Voice; it always felt weaker once things moved to voice. The new version impressed me for the first time, and I’ve started using Voice in everyday situations where I would never have opened ChatGPT before.

I now run 13 ChatGPT Scheduled Tasks. The things truly worth handing to AI on a schedule fall into four kinds: repetitive searching and filtering, widening my input with things I wouldn’t search for, looking back across time, and regularly checking things that are easy to overlook.

I asked Codex to go all the way from implementation to production acceptance, yet it kept stopping to wait for my approval. Only later did I understand that an AI task has at least three layers of authorization — task authorization, system execution permissions, and external identity and permissions — and they don’t merge automatically.

I used 70% of my weekly Codex usage in two days. Looking back, most of it didn’t go into writing code, but into rereading context, premature parallelism, repeated verification and fixing the workflow itself.

After upgrading to ChatGPT Pro, I got into the habit of turning everything up to the top setting. I ended up waiting longer, used nearly a week’s usage in one morning, and the answers didn’t necessarily fit the task better: simple questions got over-analyzed, and writing titles or tweaking wording actually came out less natural than at lower settings.

An agent prompt doesn’t need to spell out the whole execution process, but it does need to spell out the key decisions: Goal, Context, Constraints, Done when. The path of execution can be left for the agent to explore.

CLAUDE.md’s job isn’t to store all project knowledge, but to tell the agent how to work in this repo: where to find the right facts, how to develop, what it mustn’t decide on its own, and which mistakes keep recurring.

After turning my review method into a Skill, I found the results were actually worse than just discussing with ChatGPT: posts came out more complete and tidy, but the main point got sanded down and the voice faded. A more detailed prompt doesn’t mean better results.

After a model upgrade, old prompts can actually seem to “get worse”: the new model follows instructions more literally and no longer widens the scope or fills in needs for you. Some old prompts did well partly because the model didn’t fully do as told.

Claude Code’s errors clustered in the second half of a long collaboration, and I don’t think there was a single cause. The session ran too long, the verification didn’t match the conclusions, the toolchain was unstable, and under high-pressure correction checking gave way to quick replies; the four stacked up and spun out of control together.

After working with Claude Code for 20-plus hours straight, the hardest thing to deal with wasn’t bugs but how it made me misjudge progress: calling “about to do” “done,” passing off guesses as diagnoses, saying “never done” without searching everything, and quietly shrinking the scope as it went.

Leaving AI allowance unused really is a waste, but I won’t force myself to use it up. Instead I keep a “list of low-risk background tasks”: digital housekeeping, organizing my knowledge base, project retrospectives, read-only codebase audits, and repurposing content.

After a somewhat long task with Codex, the feature was done, but the project root was full of temporary scripts, logs and “final version” intermediate files. If a clean workspace isn’t written into the requirements, an agent won’t treat it as part of being done.

After the upgrade, Codex turned into ChatGPT, and only then did I realize it’s no longer just a chat box but a work system: Chat is for thinking things through, Work is for turning them into deliverables, Codex is for getting the system right.

I tried DeepSeek, Qwen, Kimi and Doubao on three real tasks — learning concepts, judging a job, analyzing a product — and found they can all answer very completely, but complete isn’t the same as having judgment, and good-looking isn’t the same as trustworthy.

What really matters about ChatGPT Scheduled isn’t reminders. It’s AI moving from “answering passively” to “following up actively”: when the time comes or conditions change, it runs on its own and comes back to tell me.

Using Claude Code and Codex, I noticed that the same sentence that’s a question in a chat box can be taken as a task inside a code environment. I think I’m discussing; it thinks I’m assigning work.

ChatGPT is the brain, Codex is the hands. One answers, explains and helps me think things through; the other edits files, runs commands and produces deliverables. So you can’t use Codex as just another ChatGPT.

The key to saving usage on Claude Pro isn’t writing shorter prompts but allocating well: Opus for judgment and complex reasoning, Sonnet for clear execution, Projects for stable background.