金曜の30分、Claudeに「私のプロンプト癖」を3つダメ出しさせた話 A personal experiment where the author spent 30 minutes on a Friday night asking Claude to…

On a Friday evening, the author carved out just thirty minutes for a small but instructive experiment: asking Claude to play the role of a critic and point out three recurring bad habits in their own prompts. The setup was simple — paste a batch of past prompts, then ask the model to analyze them meta-level rather than execute them.

The motivation is worth noting. Prompt engineering, in practice, is deeply tied to an individual's writing habits. The same author tends to embed the same implicit assumptions, the same omissions, and the same stylistic quirks across every prompt they write. Self-review rarely surfaces these patterns because the writer reads what they meant rather than what they wrote. Using an LLM as a third-party reviewer offers a cheap proxy for an outside perspective.

According to the article, Claude flagged the kinds of issues that prompt-engineering guides routinely warn about: requesting tasks without stating goals or success criteria, omitting background context the model would need, and leaving output format under-specified. These are familiar textbook points, but receiving them grounded in one's own concrete examples reportedly carries more weight than reading them in the abstract.

This approach fits into a broader industry trend. Anthropic already ships a prompt improver in the Claude Console that rewrites user prompts automatically, and OpenAI offers similar meta-prompting capabilities in its Playground. The general pattern of having an AI write or refine prompts for another AI — sometimes called meta-prompting or prompt distillation — is becoming standard practice. As agentic tools like Claude Code and Cursor take on more autonomous work, the clarity and consistency of upstream prompts increasingly determine downstream productivity, which makes auditing one's own prompt style more than a stylistic exercise.

There is also something practical about the thirty-minute framing. Many engineers struggle to find time for deliberate practice on prompting itself; bundling self-critique into a short, recurring slot — perhaps as part of a weekly retrospective — is a lightweight habit that could plausibly compound over months. Exporting chat histories and feeding them back to an LLM for pattern analysis may well become a more common personal-productivity workflow, though the privacy implications of doing so with proprietary or client data are worth keeping in mind.

The piece is less a technical breakthrough than a reminder: the models we use to produce work can also be used to inspect how we ask for that work. For individuals serious about improving with LLMs, occasional meta-reviews of one's own prompts may be one of the highest-leverage uses of a spare half hour.