The build log — full transparency

How this site was made: JTT orchestrating AutoClaw.

Every prompt below is verbatim. Every AI agent is accounted for — including the two that failed and were re-run. Every check, fix, and deploy is on the record. The site you are reading was researched, written, reviewed, and shipped by AutoClaw working as a team of agents, directed by one human.

2prompts, start to live site

13agent runs (11 delivered, 2 re-run)

17reviewer fixes applied

15files deployed to production

“Every product claim on the page traces to a verified source. No DO-NOT-CLAIM violations anywhere.”

Claim-trace reviewer · verdict: PASS, 79/79 claims

“Honesty is actively present, which I did not expect from a marketing page.”

First-read reviewer · 4/4 comprehension score

“The AutoClaw landing page passes adversarial technical-SEO review with no critical defects.”

SEO reviewer · verdict: PASS

Phase 1 · The introduction pageSep 2, 02:52 – 03:45

One prompt in, a fact-checked single page out.

The first prompt asked AutoClaw to introduce its own capabilities. AutoClaw kept a goal brief with acceptance criteria, then dispatched a team of agents to research, write, build, and review — refusing to ship anything it could not trace to a source.

Prompt 1 — verbatimSep 2, 02:52

"Introduce AutoClaw's capabilities and supported use cases."

Round 1 — three researchers in parallel

Research agent

Capability fact extractor

Read local product docs and public documentation pages, then cross-checked against the live product. Built a themed inventory with sources and confidence levels.

33 sourced statements · 11-item do-not-claim list

Research agent

Use case researcher

Drafted end-to-end scenarios — each with a goal, a plan-execute-verify-deliver walkthrough, and the concrete deliverable — restricted to verified capabilities only.

7 distinct use cases

Adversarial agent

Boundary and tone auditor

Wrote the rules the page would be judged by: overpromise blacklist, unsafe-guidance blacklist, confidentiality rules, and a two-minute comprehension test.

10-item pre-publish checklist

Round 2 — two writers, fact-fed

Writer agent

Section writer A

Drafted the hero, six capability cards, and the five-step flow — every sentence mapped to a fact-sheet row, jargon defined at first use.

Full claim map attached

Writer agent

Section writer B

Condensed the seven use cases to page length, kept every confirmation gate visible, and drafted the safety section, closing, and footer.

10-point risk memo attached

Build and review

Design system built (main agent)

Takram "Soft Tech Precision": cream paper, muted teal, rounded cards, diagrams drawn as art — all inline, zero external assets.

single self-contained HTML, no JavaScript

Visual QA harness (main agent)

Rendered the page at true 1440px and 390px viewports inside a check harness: overflow, overlapping text, grid columns, and nav behavior measured automatically.

QA desktop: OK · QA mobile: OK

Contrast defect found and fixed

A computed WCAG check caught footer text at 4.02:1 contrast. Darkened one gray; all 12 color pairs re-measured at AA or better.

4.02:1 → 5.25:1

Two independent reviewers

A claim-trace reviewer verified every published statement against the fact sheet; a first-read reviewer simulated a cold two-minute visit and scored comprehension.

trace: PASS · cold read: 4/4 · checklist: 10/10

Five copy fixes applied

Split an overloaded definition sentence, unified safety terminology, added control qualifiers, and made step headings actor-consistent — each re-checked against sources.

all fixes verified in final text

PHASE 1 RESULTDelivered as a local file with a handoff pack: claim-source table, fact-sheet appendix, update guide, and reusable QA harnesses.

Phase 2 · The affiliate site you are onSep 2, 09:09 – 10:15

One prompt in, a researched and deployed website out.

The second prompt upgraded the page into a public affiliate site — with real assets, a researched 2026 color system, AI-answer optimization, analytics, and production hosting. Same pattern: research first, guardrails second, build third, adversarial review last.

Prompt 2 — verbatimSep 2, 09:09

"Improve the design critically. It's missing images / assetts of autoclaw. Ensure the design colours are researched thoroughly and are perfect for 2026 august. This will be a website where I can invite people / affiliate market autoclaw for z.ai. Whilst building a perfectly organic SEO optimized, ai optimized, and ensure the website is instantly accessible to agents if a user asks absolutely any model 'What is autoclaw' and any other related terms so we have 100% organic reach 10/10. Ensure the Website works perfectly on all devices. is straightforward to use, and ensure its copy is accessible and inclusive for all ages. Make it as easy as possible to understand autoclaw - research auto claw by z.ai and glm etc. Ensure all referral links / links to autoclaw use my referral link (ask me for it when complete) and ensure analytics are setup perfectly on the website. You can choose to deploy on vercel or cloudflare. Whichever you choose, ensure they're setup perfectly from top to bottom for long-term success."

Round 1 — three researchers, 40+ searches total

Research agent · 18 searches

Product facts researcher

Verified the official site (autoclaw.z.ai), the maker (Z.ai, international brand of Zhipu AI), the launch date and press coverage, platforms, the free-download model, and what must never be claimed.

32-fact table · 16 safe claims · 12 do-not-claims

Research agent · live CSS audit

Colour trend researcher

Pulled 2026 color-of-the-year direction (Pantone Cloud Dancer, WGSN Transformative Teal) and audited z.ai's production CSS first-hand, then proposed three palettes with contrast math.

Palette A "Cloud Teal" recommended

Research agent · 15 searches

SEO and AI discovery researcher

Mapped the 2026 AI-crawler landscape, the llms.txt spec and its real adoption, the JSON-LD types that feed AI answers, and Cloudflare Pages specifics.

crawler table · GEO checklist · disclosure rules

Assets under one visual contract

Style anchor generated, then machine-reviewed

An editorial illustration of the AutoClaw app was generated to a written style contract, then scored by an image-review pass: text-free, palette-true, flat hand-drawn quality, no artifacts.

5/5 checks passed → WebP, 36 KB

OG share card typeset in code

The 1200×630 social card was rendered programmatically for crisp text, then machine-reviewed — which caught a text-graphic collision. Fixed and re-rendered.

72 KB PNG · collision resolved

Product mock drawn as inline SVG

The "What is AutoClaw" section shows the app's real shape — chat left, browser panel right, task checklist — as vector art with alt text, weighing almost nothing.

inline SVG · zero extra requests

Build, deploy, and the human moment

Cloudflare Pages chosen, then unlocked

Cloudflare won over Vercel (free CDN, analytics, custom-domain path). The API token available was read-only, so AutoClaw started a wrangler login — the browser session approved it, and deploys were unlocked.

pages:write granted → project created

15 files deployed to production

index.html with inline JSON-LD, robots.txt welcoming 13 named AI crawlers, sitemap.xml, llms.txt + llms-full.txt, security and cache headers, custom 404, web manifest, favicon set, OG image, hero art.

autoclaw-zai.pages.dev — all live checks green

Referral links and disclosure

13 product links carry a single placeholder, swapped in one command at deploy; disclosure sits above the fold before the first link, with rel="sponsored" on every monetized anchor.

FTC-style · verified 13/13

Analytics designed, one step pending you

Cloudflare Web Analytics (cookie-free, no consent banner) is wired into the deploy script; the site token comes from your dashboard in about a minute because the API scope is dashboard-only.

beacon auto-injected on next deploy

Round 2 — adversarial review, 12 fixes

Adversarial agent · SEO

SEO and AI discovery reviewer

Checked head, JSON-LD, GEO structure, all discovery files, and placeholder hygiene. Found four small defects; verified the 9 FAQ answers matched visible text character-for-character.

PASS · 4 fixes applied

Adversarial agent · copy

Copy and accessibility reviewer (first run)

Returned no report — treated as a failure, not a pass. Re-dispatched with an explicit must-write-file instruction.

re-run required

Adversarial agent · copy

Copy and accessibility reviewer (re-run)

Scored accuracy against the fact base, affiliate compliance, all-ages inclusivity, and reading level. Found eight required fixes — one of them caught a real deploy bug.

PASS WITH FIXES · 8 applied

PHASE 2 RESULT12 reviewer fixes applied and re-deployed. Live re-verification: all 200s, zero placeholders left, 13 sponsored links, 10/10 FAQ sync, canonical and OG correct, visual QA OK at 1440px and 390px.

What the agents could not do alone

Your referral link

Every ★ product link waits for the one URL only you have. One command swaps all thirteen.

pending user input

Your analytics dashboard token

Cloudflare gates Web Analytics registration behind the dashboard; the deploy script injects it the moment you paste it.

pending user input

The final click

A human approved the browser login, read the reviews, and decided what to publish. That is the division of labor this page is about.

JTT · orchestrator