AI Guides › Workbench
By Nigel Guy · 6 min read
Most first agents fail before anyone opens a tool. The usual default is to pick a platform, then ask it to "automate my admin", and end up with a vague assistant that does nothing reliably. "Seven agents, zero code" makes a good headline, but the durable part is smaller and less glamorous: one agent, one job, written down so clearly that a stranger could do it.
The rule: write the job on a card first, then choose the cheapest tool that can do exactly that job, and keep a human on every step that sends, spends or deletes.
A note on the claim behind this topic: we could not verify any specific "seven agents" build, so this guide does not rely on it. It covers what is checkable today, which is how no-code agents are set up and where they trip people up.
In practice, a no-code agent is an AI model plus three things you write in plain English: instructions, reference material, and (optionally) permission to act in other apps. The first two need no automation platform at all. The third is where cost, limits and risk appear.
Prices and limits move often. Check each vendor's current page, and check the £ price at checkout.
| Tool | What it does | Free tier (verified 2026-10-04) | Best for | Catch |
|---|---|---|---|---|
| Claude Projects | A workspace with your instructions, uploaded files and scoped chats | Free accounts can create up to 5 projects | A "standing brief" agent: drafting, reviewing, summarising in your style | It answers when you ask. It does not run on a schedule or act in other apps by itself |
| Zapier Agents | An agent you describe, connected to apps, with triggers and web lookups | 400 activities a month on Free, capped at 10 activities per run; testing counts toward the limit | Agents that react to an event, such as a new email or form entry | Every action counts as an "activity". Hit the limit and agents stop running (you can still chat). Zapier's help page notes agents are moving to "AI by Zapier", so menus may differ from older tutorials |
| ChatGPT agent mode | Was a built-in mode that browsed and acted for you | Not applicable | Nothing for now | OpenAI's help page for it now says ChatGPT agent is no longer available and points to ChatGPT Work. Do not follow older tutorials that start there |
Paid Zapier and Claude plans are priced in dollars on their sites. Look at the £ figure at checkout rather than trusting a converted number from a blog.
Fill this in before opening any tool. If you cannot fill a line, you are not ready to build.
| Line | What goes in it | Example (hypothetical) |
|---|---|---|
| Job | One sentence, one verb | "Turn each new enquiry into a draft reply for me to review" |
| Trigger | What starts it | "I paste the enquiry" or "a new form entry arrives" |
| Inputs | What it reads | "Enquiry text, my price list, my tone notes" |
| Output | Exact format | "Draft email, under 120 words, plus a one-line summary" |
| Stop point | What it must never do alone | "Never send. Never quote a price not on the list" |
| Test set | 5 real past examples | "Five old enquiries and the replies I actually sent" |
Use any chat assistant to interrogate you into a clear card.
You are a careful workflow analyst helping a non-technical person design their first AI agent.
Context: I run [YOUR_BUSINESS_OR_ROLE]. The task I repeat is: [THE_REPEATED_TASK].
Goal: produce a one-page "One-Job Card" with these lines: Job (one sentence, one verb), Trigger, Inputs, Output format, Stop point (actions the agent must never take without my approval), and a Test set (five example cases I should collect).
Process:
1. Ask me up to six questions, one at a time, about the task. Do not guess missing details.
2. Once you have enough, write the card in a table.
3. Flag anything on the card that is really two jobs and suggest splitting it.
Constraints: no jargon, no tool recommendations yet, no invented details about my business.
Before you answer, check that the Job line contains exactly one verb and that the Stop point names at least one concrete action.
Fill in: your role and the task you repeat.
Per Claude's help centre, Projects are self-contained workspaces with their own chats, uploaded knowledge and custom instructions, available to all users, with free accounts limited to five.
You are [AGENT_NAME], an assistant for [YOUR_ROLE] at [YOUR_BUSINESS].
Your single job: [JOB_LINE_FROM_CARD].
Use only the uploaded files and what I paste into the chat. If a fact you need (price, date, name) is not there, say "I need: [what]" and stop. Do not guess.
For each input, reply in this format:
[OUTPUT_FORMAT_FROM_CARD]
Style: [TONE_NOTES], British English.
Never: [STOP_POINT_ACTIONS]. If a request falls outside your job, say so and suggest what I should do instead.
Before replying, check that you used only supplied facts, matched the format, and avoided every "Never" item. If a check fails, fix it before answering.
Fill in: the bracketed items from your card.
Run all five test cases and compare against what you actually sent. Score each as usable, needs edits or wrong. Change the instructions, not the model, when it fails. Do this before connecting anything to an app.
Here are five test inputs and the replies I actually sent: [PASTE_PAIRS].
Compare your outputs with mine. List each difference that matters (facts, tone, format, missing caveats) and propose one edit to your instructions that would fix it. Do not rewrite the whole instruction set.
If you want it to run without you pasting, move the card to Zapier Agents. Create the agent, give it the same instructions, attach the apps it needs, and choose a trigger. Keep the first version read-only or draft-only: create drafts, never send. Watch the activity count; on the free plan testing counts, and one run is capped at 10 activities, so a long job may pause and ask for input.