How I work
Context engineering, not prompt tricks.
I use AI the way a good editor uses a research team. The advantage is not the model. It is the quality of the brief you give it and the judgment you apply to what comes back.
01
Market research first, system prompt second
Every system prompt I write starts from my own market research: who the buyer is, what they already believe, what they are afraid of getting wrong. A model given real market context produces work worth editing. A model given a vague instruction produces confident nonsense.
02
Context engineering as a deliverable
For Sapling I maintain structured context, the data model, the stakeholder map, the commercial constraints, so any tool or collaborator I bring in inherits the reasoning rather than the conclusion. This is the same discipline as writing a good BRD, applied to a new kind of team member.
03
Human judgment on the output
I strip the tells: the padded openers, the mathematical symmetry, the hedging. If a sentence does not carry an argument or a fact, it goes. Ogilvy's rule holds. The consumer is not a moron, and neither is the hiring manager.
04
Behavioural framing over feature lists
Rory Sutherland's point is that the perceived problem is rarely the real one. Micro Green Energy did not need more content. It needed a calculator, because a number ends the buyer's uncertainty faster than a paragraph ever will.
05
Agility with a memory
Adaptability is not starting over. It is carrying the constraint you learned last time into the next decision. The indexing defect I found in a client's React app is why Sapling's routing was correct before its first line of content was written.
The working loop
Eight steps, and the last one feeds the first.
This is the model behind every project on this site, not a slide invented for this page. Micro Green Energy and Sapling both went through it in order.
01
Market
Who is actually buying, and what do they already believe.
02
Research
Original research, not assumptions borrowed from a competitor's site.
03
Positioning
The argument a salesperson could repeat under pressure.
04
Product
What has to be true of the thing itself for the positioning to hold.
05
Build
Specified, built and shipped, not handed off as a brief.
06
Launch
Instrumented before it goes live, not after someone asks for numbers.
07
Measure
The metric that was defined before launch, not the one that looks best after.
08
Iterate
The constraint learned here becomes the starting assumption next time.
Feeds back into Market
Operating principles
Four rules I do not negotiate on.
Diagnose before you spend
Nobody gets promoted for the campaign they did not run, which is exactly why so much budget goes to the wrong constraint. I find the binding one first.
Less, but properly
Four channels done well beat fourteen half-built ones. The discipline is in refusing the other ten while everyone is still excited about them.
Write it so sales can repeat it
If the positioning cannot survive being said out loud by a salesperson under pressure, it is not positioning. It is decoration.
Instrument everything, quote carefully
Measurement goes in before launch. But I only publish numbers I can stand behind, which is why some figures on this site are offered on request instead.
Stack
Tools I use daily, not tools I have heard of.
- GA4
- Google Search Console
- Google Tag Manager
- SEMrush
- Ahrefs
- Screaming Frog
- SiteImprove
- Hotjar
- HubSpot
- Dynamics 365
- WordPress
- TerminalFour
- Node.js
- SQL
- GitHub
- Hostinger
- Figma
- Jira
- Microsoft 365