All categories

Payment flows are where AI-generated code gets expensive to get wrong.

Authentication is the one place your agent should never improvise.

Schema decisions outlive almost everything else you ship, which makes them a bad place to let an agent guess.

REST, GraphQL, background jobs, rate limiting — this is the backend surface an agent touches most often, and where sloppy defaults quietly become production incidents.

A UI an agent builds unsupervised tends to drift — inconsistent spacing, missed accessibility, ad hoc breakpoints.

SEO is mostly a list of small, easy-to-forget details — meta tags, sitemaps, canonical URLs — that an agent will skip unless it's told to check them.

Security is the category where an agent's confident-sounding code is most likely to be wrong.

An agent that writes tests after the fact usually writes tests that pass without proving much.

Deployment mistakes are expensive precisely because they happen in production.

This category is about the agent itself — how it should behave, not what it should build.

Every framework has its own conventions an agent should follow instead of guessing from generic training data.

These are the decisions that get made once, on day one, and then quietly govern everything after — naming conventions, env vars, monorepo layout, git workflow.

Checkout, cart, and catalog logic are full of edge cases an agent won't think to ask about — abandoned carts, tax jurisdictions, out-of-stock states.

Content modeling and localization decisions are hard to unwind once real content exists.

Every third-party API has its own quirks — auth flows, webhooks, rate limits — that an agent will get wrong on the first try without guidance.

Building AI features into your own product is a different discipline from just using one — prompt templates, retrieval, embeddings.

Not every module here is code.

Translation quality lives or dies on details an agent won't infer on its own — a fixed glossary, a consistent register, regional spelling.

Subtitles, transcription, and video editing all have hard technical rules — line length, timing, reading speed — that an agent needs spelled out, not inferred.

Good AI image and design prompts are more precise than most people expect — style, lighting, and composition all need to be stated, not implied.

Every platform rewards a different format, length, and tone, and an agent that doesn't know that writes the same generic post everywhere.

Summarizing and studying well is a skill an agent needs instructions for, same as anything else — what to keep, what to cut, how to structure flashcards or citations.

Meeting notes, specs, and pitch decks all have a shape people expect, and an agent that doesn't know it produces something technically correct but useless.

Task lists, journaling prompts, trip planning — the smaller, personal instructions that make an agent genuinely useful day to day, not just at work.

Everything that doesn't fit neatly into the categories above still deserves a good set of instructions.