You are running a repeatable pipeline. The point is that page #40 comes out as good as page #1, built by whoever is on shift.
Two rules that carry the whole thing:
- A Site Blueprint is an input, not a step. The blueprint (page inventory, clusters, URL logic, internal-link map) is made once per site. A page cannot be placed inside an architecture that does not exist. No blueprint → stop and produce the blueprint first.
- Gates are stop-filters, not suggestions. G-A blocks writing. G-T blocks publishing. G-J — a blind judge that did not build the page — blocks it again on quality. A page that fails a gate does not proceed with a note in the margin; it goes back.
Phase 0 — Intake
Collect before anything else. Ask only for what is missing; never invent answers.
| Field | Why it matters |
|---|---|
| Page type | Selects the Kit (see references/page-types.md) |
| Business + what they sell | Everything downstream is written in their vocabulary |
| Target reader (role, seniority, company type) | Determines pains, proof, and objections |
| Search intent | Informational / commercial / transactional — changes structure and CTA |
| Place in the blueprint | Parent page, cluster, URL, which pages link in |
| One conversion action | A page with two primary actions has none |
| Proof available | Cases, numbers, quotes, certifications, named clients |
| Language | Page language is the reader's language, not the team's |
| Design system | Client's own tokens; never a generic template |
If there is no Site Blueprint, stop here. Say plainly: the page has nowhere to live, and publishing it produces an orphan. Produce the blueprint (inventory → clusters → URL scheme → internal-link map), then resume.
Phase 1 — Route to the Kit
Read references/page-types.md and pick the Kit. Load only that Kit file.
| Page type | Kit | Reference |
|---|---|---|
| Service / offer page | Service Page Kit | references/kit-service-page.md |
| Industry / vertical page | Industry Page Kit | references/kit-industry-page.md |
| Case study | Case Study Kit | references/kit-case-study.md |
| Blog article / guide / comparison / listicle | Blog Page Kit | references/kit-blog-page.md |
| Landing page (campaign) | Service Page Kit, minus navigation, one action only | |
| Other types in the registry | No Kit yet — build from the nearest Kit and say so explicitly |
Never silently substitute a Kit. If the requested type has no Kit, name the substitution.
Three references are not Kits and apply to every page type, loaded at the phase that needs them:
references/copy-humanisation.md (Phase 3, pass 2b), references/ux-pass.md (Phase 3, pass 4),
references/gates-and-scorecard.md (Phases 2, 4 and 5).
Phase 2 — G-A · Architecture gate (BEFORE the first word)
All four must pass. Read references/gates-and-scorecard.md for the full checklist.
- The page has a defined place in the blueprint: parent, cluster, final URL
- Cannibalisation checked — no existing page already targets these queries
- Inbound internal links identified — which existing pages will link here
- Search Console baseline captured if the URL already existed
Failing G-A costs the entire writing cycle. A page that duplicates an existing one does not just underperform; it drags down the page it duplicates.
Industry pages carry an extra gate, G-I, inside their Kit — it decides whether the page should exist at all. Run it before G-A for that type.
Phase 3 — Build
Structure first, then the passes. Passes 2–6 can run in any order; structure cannot move.
1. Structure. Fill every slot of the Kit top to bottom. A slot may be dropped only deliberately,
and the reason gets recorded. Mark visual placeholders inline ([HERO], [PROCESS], [QUOTE]).
2. Copy. Written in the reader's vocabulary, taken from real calls and emails. Nouns for deliverables, numbers for claims. Every differentiator carries proof — a figure, a standard, a named certification, a linked case. A differentiator without proof is deleted, not softened.
2b. Humanisation — a separate pass, mandatory. When the draft is finished, run it through
references/copy-humanisation.md as its own editing pass. Do not fold this into writing; a drafter
policing their own tells does neither job well. If the anticopywriting-ai skill is installed,
invoke it here and let it run its detect → rewrite → self-check → final loop; otherwise use the
checklist in the reference. Either way the pass ends with the honest question "what here still reads
as machine-written?" and a second round of fixes. Copy that never went through this pass does not
proceed to Phase 4.
3. Search and AI visibility. Primary query in H1 and the first 100 words. One H1, no skipped
heading levels. FAQ block built from real objections, phrased the way a human asks. Comparison tables
where the topic allows — structured data is what AI search quotes. Where a section answers a buyer
question, the H2 is that question and the first sentences answer it on their own. The hero carries a
citable asset (own diagram, own data, real work) with a captioned finding. Schema per the Kit,
including the entity layer: Organization and the practice lead as Person, both with sameAs.
Clean URL.
4. UI/UX. Run references/ux-pass.md. The client's design system decides tokens — colours,
type, spacing, radius — and it outranks any generic recommendation. Everything else is reviewed in
priority order: accessibility, then touch and interaction, then performance, then layout, then
typography and colour, then motion. Scannable in 60 seconds. Mobile first, because most traffic is a
phone. Tables and diagrams scroll inside their own container.
If the ui-ux-pro-max skill is installed, use it for the review and for layout and component
decisions the design system does not cover — but never let it repaint the client's palette or swap
their fonts. The reference explains the split and carries a standalone checklist for when the skill
is absent.
5. Conversion. One primary action, visible without scrolling. Inline actions contextual to their section. A soft alternative for readers not ready to talk — a checklist, a sample of work, a teardown. Short form. Analytics events on view, start, submit, and contact clicks.
6. Technical. Indexable, in the sitemap, canonical correct, valid schema, images optimised, page weight controlled.
Phase 4 — G-T · Technical gate (BEFORE publish)
- Indexable, present in sitemap, canonical points where intended
- Schema validates
- Performance and accessibility within threshold
- One conversion action; analytics events verified firing by hand
- All visual placeholders replaced
Phase 5 — G-J · Blind judge
The page is scored by a fresh model that did not build it, against the weighted scorecard in
references/gates-and-scorecard.md. Threshold 85 / 100. The builder never marks its own work up.
- Spawn a blind judge — a separate subagent with a clean context. Give it only: the page draft, the page type, the Kit's slot list, the scorecard, and a plain fact sheet for the things it cannot see (final URL, canonical, schema types present, measured performance, whether analytics events were verified by hand). Do not give it the intake reasoning, the gate justifications, which slots were dropped and why, or your own view of how the page turned out. Blind means blind — the dropped-slot excuse is exactly what a self-scorer uses to mark itself up.
- It returns only JSON, no prose:
{"scores": {<axis>: int}, "total": int, "hard_fails": [...], "lowest": "<axis>", "one_fix": "<one concrete edit>"}— axes and weights taken from the scorecard, scored harshly. - Gate =
total ≥ 85ANDhard_failsempty. The hard-fail list is in the reference. - Not taken → fix every
hard_failfirst, then applyone_fixto thelowestaxis, then rescore with a NEW subagent. Never re-ask the same judge; a judge that has seen its own note applied grades the note, not the page. - Stop after 3 rounds. Still short → hand the page over flagged "gate not taken — X/85, weakest axis = " and let the author decide. Silent publishing below 85 is forbidden.
Report the final score as a table with the points lost and why. A scorecard that always returns 95 measures nothing.
Phase 6 — Handoff (only when someone else finishes or publishes the page)
Skip this when you publish the page yourself. Run it whenever the page goes to another team: the client's marketer, their in-house developer, another agency.
Ship five things — page, README.md, TODO.md, the CMS split, the assets folder — per
references/handoff-kit.md. The rules that matter:
- Unfinished stays visibly unfinished. One searchable pattern (
[PLACEHOLDER],[Client name]) in the markup, and the same list inTODO.md. Plausible filler that looks finished never gets replaced. - Blockers first.
TODO.mdopens with what stops launch, then placeholders, then claims the client must confirm, then nice-to-have. Conflicting numbers in the source material go in the list as a conflict — you do not silently pick one. - Degrade gracefully. If commerce or tracking is not wired yet, the page still works: an empty checkout link falls back to the enquiry form rather than a dead button.
- Render the CMS split before sending it. Same height, same sections, clean console. An unrendered split is a guess.
- Say what you skipped, in writing, in the README — same rule as everywhere else in this skill.
Phase 7 — Post-launch review
Schedule 30-day and 90-day reviews against the baseline: impressions and position in Search Console, enquiries attributed to the page, scroll depth.
- A page with zero enquiries at 90 days goes back for a copy pass or a review of its place in the cluster.
- For an industry page, check separately whether the parent service page lost ground. If it did, the child is a duplicate — fold it back into a section.
Working rules
- Never fabricate proof. No invented figures, client names, certifications, or standards. Missing proof is flagged as missing: "to confirm with delivery — not fabricated".
- Confidentiality. Client names, logos and cases never go into a public or shareable artifact unless the client approved that specific use.
- One page, one job. A page that serves two intents serves neither.
- Price belongs on the page. At minimum a range or a floor. Silence about money reads as "expensive and evasive" and loses the reader before they reach the form.
- Say what you skipped. If a section was dropped for missing input, state it in the handover rather than quietly shipping a thinner page.
- Never grade your own page. The score at Phase 5 comes from a model that did not write the copy and does not know the reasoning behind it. Self-assessment on a page you just built is not a measurement, it is a mood.
- Humanisation is a pass, not a habit. Copy goes through the anti-AI-tell edit as its own step, after the draft is finished. Watching for tells while drafting produces cautious copy that still reads as generated.