What AI tools should we use?

The two honest starting points

A general assistant — ChatGPT, Claude — handles reading, writing, comparing, summarising and drafting across almost any subject. It is the right first tool for most task lists because one subscription covers a dozen different jobs.

A specialist AI service is software built for one job: answering the phone, reading invoices, transcribing meetings. It costs more per job but does that job with far less setup. We only recommend one where a general assistant genuinely cannot do the work.

Where a client already runs on Microsoft 365 or Google Workspace, we check what is included before recommending anything new. Paying twice for the same capability is the most common waste we see.

Where we take it further

Choosing a tool is the easy part. The reason AI disappoints businesses is that the same request, typed slightly differently by two people on two days, produces two different answers. That is fine when you are exploring. It is useless when the output goes to a client, a principal contractor or the tax office.

So we do not hand you a subscription and a suggestion. We set your specific tasks up on the tool, with the exact instruction the task needs, so it performs your critical work correctly and the same way every time. Consistency is the deliverable.

Prompting, explained without the jargon

A prompt is just the instruction you give the tool. Left unstructured — "summarise this" — you get variability: sometimes three bullet points, sometimes a page, sometimes the wrong three points, and no way to tell why.

A structured instruction names the input, the steps, the format, the things that must always appear, and the things the tool must never do. It reads more like a work instruction than a question. Written once, it holds.

We write that instruction for you and hand it over, so it is yours. You are not renting the prompt from us, and you can read it, change it and hand it to a new staff member.

The four tool types we build with

On top of the assistant, there are four ways to package a task. Which one we use depends on how the task is used day to day — whether somebody needs a screen, adjustable settings, a page to share, or just a reliable set of rules.

• A skills.md file — a repeatable set of rules for doing a task, like an SOP the AI follows.
• Artifacts — like a skill, but with a visual screen and adjustable settings such as dates, quantities and reports.
• Sites — deployable pages that can hold apps, skills and artifacts, for when the work needs rich visual presentation.
• GEMs — mini AI helpers that do one pre-defined job, such as research, image generation or a standing report.

How we choose between them

• One person, one recurring job, no settings to change — a skill.
• Several people, numbers or dates that change each run — an artifact.
• Output that has to be shown to clients or staff in one place — a site.
• A narrow standing job that runs on request — a GEM.

What this looks like in practice

A calibration business we work with needed regulatory bulletins read every month and anything mandatory separated from anything advisory. That is a rules job with a fixed output, and it saves at least two days a month.

A builder needed quotes assembled from site notes with their own rates — numbers that change every job, so it needed a screen and settings. A consultancy needed performance figures in one place for a weekly meeting, which needed a page.

Same underlying assistant in each case. Different packaging, because the work is different.