The AI tool directory, by category

Categories, not brands. Each category below says what kind of work it eats, what it leaves behind, and which names are worth looking at right now. The date tells you how stale the names are. The categories will outlast every one of them.

Planned for release phase 2. The page is complete as a draft; interactive parts run in the browser and do not yet save to an account.

In brief

To choose a tool, start with what goes in and what comes out; that halves the field. Then ask how much context it must hold, whether the work is iterative or one-shot, how accurate it has to be, and how often you will do it. If none of the categories fits, use the "use AI to find AI" prompt at the end of this page. Names last reviewed 12 September 2026; the next review is due in March 2027.

Categories
  1. Conversational text tools
  2. Coding assistants
  3. Image generation
  4. Video generation and editing
  5. Voice, transcription and meeting notes
  6. Research and document tools
  7. Agents
  8. Vision and analytics

Conversational text tools text → text, back and forth

Thinking out loud, drafting, summarising, asking questions you do not quite know how to ask yet, reviewing your own plans. The category most people mean when they say "AI" without thinking.

Hand it

Drafting and rewriting; summarising long material; first-pass research; flow review (Step 2 Phase B); turning your plain-language brief into an optimised prompt (Step 5).

Keep for yourself

Anything that needs context only in your head; final judgement; facts that have to be right unless the tool has web access and you check the sources.

Names to look at, September 2026: ChatGPT (OpenAI) · Claude (Anthropic) · Gemini (Google) reviewed 12 Sep 2026

For iterative work with long documents, choose on context window size. For anything client-facing and factual, use one with web access and read the sources it cites.

Coding assistants code in your editor, or an instruction → a function

Tools that live inside the editor, watch what you are writing and suggest the next piece, or take an instruction and produce a function, a test, a refactor. "For a junior developer in 2026 these aren't optional anymore."

Hand it

Boilerplate, test scaffolding, refactors with a clear spec, translating a known behaviour into a new framework, explaining unfamiliar code.

Keep for yourself

Design decisions; anything you could not defend line by line; the Tab you press without reading, which is where the book's protagonist finds his own dependency.

Names to look at, September 2026: Cursor · GitHub Copilot · Claude Code reviewed 12 Sep 2026

Whatever you use, read what it wrote. "If you can't explain it, you don't ship it" is the Avoider's line in the book, and it is also just engineering.

Image generation text → image

Describe an image, get an image. Moodboards, storyboards, campaign visuals, concept art, the stock photography nobody pays for any more.

Hand it

Moodboards and storyboards (single-shot, taste-judged by you); variations to choose between; placeholders.

Keep for yourself

The art itself, when the art is the point — Kavya's line. Anything where the maker's hand is what is being bought.

Names to look at, September 2026: Midjourney · Adobe Firefly · Stable Diffusion reviewed 12 Sep 2026

One-shot tasks suit specialised tools; a chat interface is not needed. Generate ten, choose one — the choosing is yours.

Video generation and editing text or image → video

Not as good yet as image tools, improving fast enough that the book's mentor "would bet on them".

Hand it

Rough cuts, explainer drafts, animatics, resizing and captioning at scale.

Keep for yourself

A finished piece whose tone has to land — the fintech video in the book was fixed by rewriting the end state, not by the tool.

Names to look at, September 2026: Runway · Sora (OpenAI) · Adobe Premiere's generative features reviewed 12 Sep 2026

Check the accuracy requirement: anything with real people, real claims or regulated content needs human sign-off at the frame level.

Voice, transcription and meeting notes audio → text

Record a meeting and get a clean, speaker-tagged transcript in minutes. Turn a long document into audio for the commute.

Hand it

Transcribing client calls and interviews (verifiable output, low cost of error); meeting summaries you will read and correct; dictation.

Keep for yourself

Anything where nuance of tone is the content; decisions made on a summary nobody checked against the recording.

Names to look at, September 2026: Whisper (OpenAI) · Otter · Fireflies reviewed 12 Sep 2026

High frequency: if you transcribe weekly, invest in setup and a template for what you want back from every transcript.

Research and document tools your documents → answers, summaries, audio

Tools that work over a set of documents you give them rather than the open web — a folder of reports, a course, a brief — and answer from those.

Hand it

Getting oriented in a large body of material; producing a study guide or an audio overview; finding where something is said.

Keep for yourself

Substituting the overview for having read the thing, when reading it is the job — the student trap.

Names to look at, September 2026: NotebookLM (Google) · ChatGPT / Claude projects with uploaded files reviewed 12 Sep 2026

Context requirement is the deciding characteristic here: choose the tool that can hold everything you need it to see.

Agents an instruction → actions taken on your behalf

The new category. Tools that do not just answer but go off and do: book the flight, send the email, run the analysis, work through a task list. "We're early in that, but it's where everything is heading."

Hand it

Recurring, well-specified, verifiable sequences where the cost of error is low and you review the result.

Keep for yourself

Anything irreversible, expensive or reputational without a human check — the Decision Filter's fourth question applies twice over.

Names to look at, September 2026: Agent modes inside the major conversational tools · Task-specific agents in your existing software reviewed 12 Sep 2026

Names in this category change fastest of all. Judge by the category questions, and start with something you would be comfortable undoing.

Vision and analytics image → text; data → pattern

Reading an image, a chart, a photo of a whiteboard, a screenshot; and finding patterns in data you give it.

Hand it

Extracting text and structure from images; first-pass data exploration; charts you will check.

Keep for yourself

Conclusions drawn from data you did not understand yourself — you cannot evaluate what you cannot read.

Names to look at, September 2026: Vision features in the major conversational tools · Spreadsheet and BI tools' built-in AI reviewed 12 Sep 2026

Verifiable output is the strength here; use it, and verify.

Use AI to find AI

When your task fits none of the above, or you want a second opinion on the names.

I need to do this task: [DESCRIBE IT IN A SENTENCE OR TWO]. The input is [text / image / audio / data / a set of documents] and I want back [text / an image / a transcript / a table / actions taken]. It needs [very little / a lot of] context. It is [something I will iterate on / a one-shot task]. Accuracy matters [a great deal — facts will be published / less than range and ideas]. I will do this [once / every week]. Search for current AI tools that fit, evaluate each against those characteristics, and give me a shortlist of three with one line on why each fits and one on its main limitation.

Editorial policy: names are reviewed every six months against the framework's five characteristics; a name is listed because it is a strong current example of its category, not because of any commercial relationship. The book's own list (Chapter 4, April 2026) named ChatGPT, Claude, Gemini, Cursor, GitHub Copilot, Midjourney, Stable Diffusion, Adobe Firefly, Sora, Runway, Whisper, Otter, Fireflies and NotebookLM.

Questions people ask

Which AI tool should I use?
The wrong first question. Ask what goes in and what comes out, how much context the task needs, whether it is iterative or one-shot, how accurate it must be, and how often you will do it. That points to a category; pick the strongest current name in the category and switch when something better appears.
Why does the book not recommend specific tools?
Because it would be out of date within six months of printing. The book teaches selection thinking; this page carries the names, with a review date, so the two do each other's jobs.
Is it bad to use one tool for everything?
The book calls it a screwdriver for every repair. It works for some things, badly for others, and you never find out which because you never tried anything else. The five characteristics exist to break the habit.