In Codex, That Row of “fix bug” Is Unrecognizable After a Week

Hamel Husain posted a very short tweet: if you’ve piled up a lot of tasks in Codex, have Codex rename the tasks across the whole project according to this article by Shreya Shankar. His assessment was “quick quality life upgrade.” In plain terms: spend a few minutes cleaning up task names, and you’ll save yourself a lot of pain later when you’re trying to find things.

The pain point is concrete. Once a task list gets long, the screen is full of fix failing test, Find issues in tagged slack thread, update config. On day one you still remember the context; a week later all you remember is that you seem to have changed something. Hamel also said in a reply that he had never thought about renaming sidebar tasks before, and only realized it was worthwhile after reading the article.

Shreya’s article is called Five Principles for Generating Readable AI Text in Bulk, and the original link is on the fsdatalab blog. She is discussing bulk AI-generated text, such as session titles in the Claude Code and Codex sidebars, iPhone notification summaries, and temporary dashboards. The problem isn’t whether an individual sentence reads smoothly; it’s that when dozens of pieces of text are placed together, it’s hard for people to scan them quickly. The original article includes five sets of before-and-after comparison images; they’re worth clicking through.

The five principles can be compressed into five sentences.

1. Write in the order people read sentences. English readers are used to looking first for the subject, verb, and object. In a list, if the subject is the same across items, omit the subject, start with a verb, and keep the tense consistent. Claude’s sidebar used to mix nouns, gerunds, and imperatives; changing them to consistent verb-first phrasing like Measuring vLLM performance makes them much easier to scan.

2. Put the most important words first. People scan web pages in an F-shaped pattern and often only look at the first two words in a list. Putting “Shipped” at the front of order emails is more useful than starting every email with “Your order.”

3. Unify structure and vocabulary. When Notion AI summarizes GitHub issues, some entries say “The issue reports that,” others say “Pressing Ctrl+C,” and it mixes “issue” and “page.” Readers have to repeatedly check whether these are the same thing. Perplexity Discover’s headings are much more consistent.

4. Use variation only to express meaningful differences. In Duolingo’s review topic cards, only the bug card included the version number 7.146.0, so readers would assume the version number was related to the bug. Once the template is unified, differences become explainable.

5. The emotional tone of verbs should align with the panel’s purpose. A complaints panel should use “complain”; a praise panel should use “love.” Readers scanning a dashboard have very low attention, so the verb needs to tell them upfront whether this is good news or bad news.

This set of ideas can be used directly to improve Codex task names. When asking Codex to rename tasks, don’t just say to come up with better names. Give it rules: start every item with a verb and keep tense consistent; put the words that distinguish the task first; include the repository, PR number, or branch; use the same structure for tasks of the same type. Someone in the replies to Hamel’s tweet added that if you put the PR number or branch in the task name, that’s what you’ll search for a week later. Others said to write the naming rules into AGENTS.md so that new tasks are named according to the rules when they are created, instead of waiting for the list to rot and then doing batch firefighting. One more reminder worth remembering: tasks drift midway through, so it’s best to have the agent rename them once more when they’re finished.

Task names seem like a small thing, but they’re actually a search index. A column of “untitled task” or “fix bug” is equivalent to losing the context. Renaming doesn’t solve code problems, but it lets you still recover the judgment you had at the time. The original article’s five principles and comparison images are here: https://fsdatalab.github.io/blog/agent-text-at-scale/ . Hamel’s original tweet is here: https://x.com/HamelHusain/status/2115918321141109029 . If you want to get started directly, you can give Codex this prompt: Read this article and rename all tasks in the current project according to its five principles, preserving repository, PR number, and branch information, and using the same structure for tasks of the same type.