AI Guides › Playbooks
By Nigel Guy · 8 min read
The usual way people overspend on AI is to buy their way past a problem they have not yet hit. A second chat subscription "to compare", a top tier "so I never run out", a prompt pack from someone with a big following. Each one feels like progress on the day you buy it, and the bank statement three months later tells a different story.
The rule: pay for a tool only after you have hit a specific limit doing real work, and write that limit down before you pay.
This playbook gives you one named mechanism, the Spend Ledger, and runs five common purchases through it. Each comes with the free or cheaper route that usually does the same job.
The Spend Ledger is a five-column table you keep in a note or spreadsheet. Every AI cost you pay, or are about to pay, gets one row.
| Column | What you write |
|---|---|
| 1. Item and price | Name, monthly cost in £, renewal date |
| 2. The limit it removes | The exact wall you hit on the free route ("ran out of messages mid-draft three times last week") |
| 3. Last real use | Date and task, from your history, not from memory |
| 4. Overlap | Anything else on the ledger that already does this |
| 5. Verdict | Keep, downgrade, cancel, or "not yet" |
How to run it:
New purchases go through the same table before you pay. If column 2 is blank, the verdict is "not yet".
Paying for two or three general-purpose assistants at once is the most common leak. They overlap heavily, and you end up pasting the same question into each and reading three answers instead of acting on one.
The cheaper route: pick one paid assistant and keep the others on their free tiers for an occasional second opinion. The free tiers are more capable than many people assume. At time of writing, Claude's free plan includes web search, file creation, code execution, memory, artifacts and up to five projects, and ChatGPT's free plan includes projects, search, voice and limited deep research. Both are rate-limited, which is exactly the point: you find out what you actually run into.
Higher tiers exist for people who hit usage caps every week. At time of writing, Anthropic lists Claude Pro at US$20 a month billed monthly (US$17 a month on annual billing) and Max "from" US$100 a month; ChatGPT has Go, Plus and a Pro tier with three usage levels. These are USD list prices before tax; your UK checkout shows the pound figure including VAT, so check it there before you commit.
The cheaper route: run the entry paid tier monthly for at least a month and note each time you hit a limit (date, task). Upgrade only if the notes show it happening on work that matters. Take annual billing only once a tool has survived several Ledger reviews, because annual saves money only on a tool you would have kept anyway.
This covers the expensive kit and stacks: a new machine bought to run models locally, a paid automation platform, several connected agent tools wired together before you know what they should do. The setup becomes the hobby, and the actual work waits.
The cheaper route: do the task by hand in a normal chat at least three times. If it is still the same task on the third go, write it down as steps, then automate the single step that costs you most time. Local models and automation platforms are reasonable choices once you can name the task, the volume and the privacy need that justify them. Before that, they are setup theatre.
Many "AI writer" and content-generator products are a template layer over the same underlying models you can reach directly. Some add genuine value (a workflow, an integration your team needs). Many add a monthly fee and output that sounds like everyone else's, because the template is shared by every customer.
The cheaper route: before subscribing, ask the vendor or check their documentation for what model sits underneath and what they add on top. If the answer is "templates", write your own template once in the assistant you already pay for. Your voice, your examples and your audience are what make content stand out, and no template can supply those.
A pack of hundreds of prompts written for nobody in particular will rarely beat one prompt written for your task, your audience and your standard. Packs also go stale as models and features change, and you pay for the 95% you never open.
The cheaper route: the prompt interview below. It replaces every prompt pack with one prompt per recurring task, built from your own answers.
Use this once for each task you do every week. The model asks you questions first, then writes a reusable prompt with your standard built in, so quality does not depend on how alert you are that morning.
You are a prompt engineer helping me build a reusable prompt for a task I repeat every week: [WEEKLY_TASK].
Goal: a prompt I can paste into a fresh chat each week, changing only one or two clearly marked inputs, that produces work meeting my standard without needing this conversation.
Steps:
1. Before writing anything, ask me five questions, one at a time, covering: who the output is for, my constraints (length, format, deadlines, things to avoid), my voice with an example if I have one, what an excellent result looks like, and what a weak result usually gets wrong.
2. If any answer is vague, ask one follow-up rather than filling the gap yourself.
3. Then write the reusable prompt. It must include: a role, the context from my answers, the goal, numbered steps, the output format, a short quality standard written as checkable criteria, and the variable inputs in [SQUARE_BRACKETS] at the top.
4. Make the prompt fully self-contained. It must not refer to this chat or assume any memory.
Output: the finished prompt in one code block, followed by a three-line note naming the variables I change each week.
Before you reply with the final prompt, check: does it contain everything from my answers, are the variables clearly marked, and would a stranger with no context get a good result from it? Fix anything that fails.
Fill in: replace [WEEKLY_TASK] with the task in plain words, for example "a summary of this week's customer emails for my business partner".
Save the result in a project, a note or a document you can find again. In both Claude and ChatGPT you can store it in a project's instructions so it applies to every chat in that project.
Imagine a sole trader, Priya, who runs a small bookkeeping practice. Her Ledger shows three general chat subscriptions, an AI writing tool and a prompt pack bought last spring.
The figures here are illustrative; your Ledger will look different, and that is the point of filling it in.