AI Guides › Workbench
By Nigel Guy · 6 min read
Short-form videos often end with "comment a keyword and I'll send you the resources". The reply is usually a bundle of links, and the bundle is the bad default: you collect it, feel prepared, and never open half of it. A pile of someone else's links also ages quickly, because the products it points to change names, menus and prices.
We could not verify what any particular creator sends in response to such a comment, so this guide does not review one. It covers the durable part: where the official guides live, how to keep up without drowning, and how to hold the whole lot in a small library you actually use.
The rule: a resource earns a place in your library only when it has an owner, a date and a first task you will do with it this week.
Prices below are US dollars as shown by the vendor; check the £ price at checkout, since VAT and exchange rates move it.
| Resource | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| Anthropic prompt engineering overview and prompting best practices (platform.claude.com) | The vendor's reference for writing prompts: clarity, examples, structure, roles, chaining | Free to read | Anyone writing prompts for Claude | Says plainly that it assumes you already have success criteria and a first draft |
| Anthropic's interactive prompting tutorial (GitHub and a Google Sheets version) | Worked exercises on fundamentals | Free | Learning by doing | Written for Claude 3 models; Anthropic says to follow the best-practices page where they differ |
| Claude Academy (academy.claude.com) | Self-paced courses; includes "AI Fluency: Framework and foundations" built around a 4D framework (Delegation, Description, Discernment, Diligence) | Free, per the site | Building judgement, not just tricks | Courses are hours long; pick one |
| Claude Code docs (code.claude.com) | Install steps, quickstart, best practices, common workflows | Docs free; Claude Code needs a paid Claude plan or Console account. Pro is $20 a month (about £15 at time of writing; check checkout) | Anyone automating coding or file work | The docs move fast; read the date, not the memory |
| OpenAI prompt engineering guide (developers.openai.com) | Message roles, prompt anatomy, few-shot examples, formatting | Free to read | Seeing how a second vendor frames the same craft | Written for API users; translate to chat yourself |
Two of those are worth a closer look, because they are where most of the "real systems" live.
The page is built around one constraint: the context window fills quickly and performance drops as it does. Its practical advice is to give the model a check it can run (tests, a build, a screenshot comparison), to explore and plan before coding, to keep CLAUDE.md short, and to clear context between unrelated tasks. You do not need to be a developer to take the principle: ask for evidence, not assurance.
Anthropic and OpenAI describe the same shape in different words: give the model an identity and context, say what good looks like, show an example, and format the output. Read one in full and skim the other for differences.
The mechanism is a Library Card: one short entry per resource, kept in a note or spreadsheet. Five fields, no more.
Anything without a first task by the end of the week is dropped. You can always find it again.
Use this in any chat assistant with browsing enabled. Fill in the square brackets.
You are a research assistant helping me build a small, trustworthy learning library about [TOPIC, e.g. writing better prompts for my job].
Context: I work as [YOUR_ROLE] and have about [HOURS_PER_WEEK] hours a week to learn. My current level is [BEGINNER / INTERMEDIATE]. The tools I use are [TOOLS].
Goal: a shortlist of no more than [NUMBER, e.g. 5] resources, each one an official or primary source (vendor documentation, help centre, official course or official repository). Prefer these over blogs, newsletters and social posts.
Steps:
1. Search for current official sources on the topic. Open each page rather than relying on search snippets.
2. For each candidate, record: title, publisher, URL, the date shown on the page (or "no date shown"), and what it covers in one sentence.
3. Reject anything you cannot open, anything that is a summary of another source, and anything with no named publisher. List the rejects in one line each with the reason.
4. For each kept resource, propose one first task I can do within seven days using my own material, and say what a good result looks like.
Output: a table with columns Name, Publisher, URL, Date shown, Covers, First task, Good result. Then a short list of rejects.
Rules:
- Do not invent URLs, prices, course lengths or feature names. If you could not confirm something, write "not confirmed".
- If you need information from me (role, tools, level), ask before searching.
- State any price in its original currency and tell me to check the checkout price in pounds.
Before answering, check that every URL in your table came from a page you opened, and remove any row that fails.
Pick one change source per tool, and look at it on a fixed day rather than whenever a notification tempts you. For Claude Code, the docs include a page index and the install page notes that native installs update in the background while Homebrew and WinGet installs do not, so updating is your job there. For any vendor, set a monthly reminder to re-open your Library Cards and update the "Checked on" date. If a page has moved (the old Anthropic docs address now redirects to platform.claude.com, for example), update the link in the card.