Port 2 Home

What made the project unusual is what the research phase found on the other side of the brief. Sixty-three questions went to the client across two rounds, covering the legal entity, whether the company is trading, what it actually sells, who it is for and how to reach it. Sixty-three of sixty-three came back unanswered, twelve of them blocking. So the design question was not how to present this business well. It was what a page is allowed to say about a business that has told nobody anything, and how to make that page persuasive anyway.

Underneath both sits the unanswered brief. What the business sells and who it buys from were open questions, so the page's proposition, its audience and even its vocabulary rested on positions taken by the project rather than by the client. A page written that way is not automatically dishonest, but it becomes so the moment a working position hardens into a public claim without anyone noticing the promotion.

There was also a discipline problem inherent to the workflow. An assistant fills silence, confidently and instantly, and on a project this short of facts that is the failure mode that matters. Ask for a logistics landing page with nothing to go on and you get a founder quote, a customer count, a percentage and three testimonials, none of it flagged and every line a claim about a real company that nobody there has made. The same speed makes a second complete design direction nearly free, which sounds like an advantage and is how projects drift into taste-driven churn instead of decisions.

So the journey section is the spine: five stages, five handovers, and a rail drawn as one continuous element. A page arguing that accountability does not break at the handovers cannot have its own line break at them, which is also why one of the automated checks does nothing but assert that the rail stays unbroken at every width.

Specificity carries the credibility that proof normally would. The copy works in trade nouns rather than adjectives, demurrage, deconsolidation, gate-out, packing list, proof of delivery, on the reasoning that knowing the vocabulary of the job is the one competence a page can demonstrate without a customer to vouch for it. Against that, a banned vocabulary list: trusted, proven, leading, seamless, guarantee, 24/7 and the percent sign are absent from all 109 copy slots because each claims something the client has not established, and platform, software, dashboard and API are absent because whether P2H is software or a service is an open question and the page must not answer it by accident.

Three complete design directions were built rather than one refined, each owning every file under its own directory and sharing no token name, class name, component or stylesheet rule, deliberately, so a rule copied from one into another fails loudly instead of half-working. The first is editorial and deliberately unfashionable, light paper and a serif display with one anchor hue and one signal colour used exactly once, at the terminus of the route motif; typography and hairlines carry the page and the single drawn mark on it is structural. The second is a dark navy canvas using one variable grotesque through its width axis, with ornament tessellated from the client's own mark. The third is a two-colour press job on bone stock, the client's navy and the client's sky with tone made by CSS halftone screens rather than a third ink, crop marks on open elements and a 45 degree chamfer on filled ones, one board that indexes the whole argument before seven ruled bands print it. Promoting any of them to the site root is a single line, so the choice stays the client's and stays cheap.

Everything above had to survive being made quickly by one person with an assistant, and that is what the second half of the system is for. Evidence was classified before a word was written, 26 facts verified against the served HTML and the site's own WordPress API, held in a register that separates fact from inference from guesswork so nothing gets quietly promoted. All sixty-three open questions sit on one severity ladder with stable IDs, and every load-bearing sentence names the assumption it rests on together with that assumption's reversal cost, which is the number that decides whether the page can ship without an answer. Real copy and invented placeholder copy live in separate files, and in the honest state a gap renders as a bracketed placeholder naming the client question that unblocks it, with a build gate that failed while invented content reached the output. Five review rules became executable checks, and each was run against a deliberately broken build and confirmed to fail before it was trusted.

The workflow's speed has to be pointed somewhere or it turns into churn. A second complete design direction is nearly free here, and free variation with nowhere to go becomes argument about one page. Putting each direction in its own isolated folder, sharing nothing, deliberately naming the tokens differently so a copied rule breaks loudly, made three parallel designs cost less than one defended design and gave the client a choice instead of a defence. I would not share components across design directions again, even where it is easy, because the sharing is what makes them variants of one idea rather than three ideas.

The same speed is also why the honesty machinery had to be mechanical. An assistant fills silence convincingly and instantly, and on a project this short of facts that is the failure worth engineering against: a page written that way is not a draft with gaps, it is a set of claims about a real company that nobody there has made. Rules in a document are advice, so the ones that could become assertions did, and a gap became a bracketed placeholder naming the question that unblocks it rather than a sentence covering for it.

The limit of that is worth saying plainly, because it is the most useful thing the project taught me. When the client asked for a presentable page rather than a specification with holes in it, the invented content the system had quarantined was exactly what filled those holes, and the gate did its job: the first deployment attempt failed on sixty findings reaching the output. The decision to show that content anyway was the client's, it was raised and reaffirmed, and it is recorded in the repository as a decision with the objection attached rather than as an oversight. Where the choice was mine, it was mitigated: the faces beside the invented quotes are AI-generated rather than photographs of real people, so no living person's likeness is attached to words they never said, and reverting the whole thing remains a one-word change. A gate can make invention visible, detectable and cheap to undo. It cannot make the call.