How We Build
Every product that ships from this studio follows a documented, repeatable process. From the first idea scored against a rubric to the fifth critic round before the artifact counts as done, nothing is left to chance or mood.
The Foundation
Most software shops build the same way: a developer picks a ticket, works until it feels done, ships it to staging, and waits for someone to complain. The review is informal. The quality gate is a shrug. The result is inconsistent output that depends entirely on who was working that day and how tired they were.
Wishdeal Studio runs differently. The backbone of the system is the tick cycle: a discrete unit of production with a defined start, a defined output, and a defined review sequence. Ticks are not sprints. They do not carry estimates or acceptance criteria or retrospectives. They carry a decision, an artifact, and a quality gate. The system is simple because it needs to run autonomously, overnight, at a consistent quality bar.
The method is documented here because opacity is a quality risk. When a system cannot explain itself, the people reviewing its output cannot calibrate their trust. Every rule described on this page is a rule the Factory actually runs. If the output does not match the description, the description is updated.
The Tick Cycle
A tick is the fundamental unit of work in the Factory. Each tick begins with a state audit and ends with a published artifact or an updated support file. Ticks are adaptive: a heavy action like a new product shell runs for 30 to 45 minutes; a light action like a changelog entry runs in under 10. The system self-paces by measuring how long each tick took and adjusting the sleep interval accordingly, capping the inter-tick gap at 15 minutes.
The Director reads the current state of the factory: what exists, what is shallow, what files are hot (modified in the last five minutes), and what the last six decisions were. No action is chosen in isolation from what the portfolio currently looks like.
Two hard rules apply before any action is selected. The depth gate blocks new product creation if fewer than half of existing products have two or more sub-pages. The breadth gate blocks non-product actions if new builds have been underrepresented in recent ticks. The depth gate wins when they conflict.
From a menu of 14 action types (A through N), the Director picks the highest-value available action. The decision is logged with a one-letter classification and a written rationale. Patterns in the log are visible to the next Director tick, so blind repetition is caught and corrected.
For any new product, a brand brief is written before a single line of HTML is produced. The brief specifies the palette (three hex codes), the typeface pairing, the voice, and the thematic category. It is saved as a permanent file that the system reads back on every subsequent polish or marketing pass.
The artifact is written to the canonical path under /srv/sites/factory/. For landing pages, the target is 1500 or more words of real content, real photography sourced from Unsplash, product-specific navigation stripped of studio chrome, and a footer credit that is the only visible reference to Wishdeal Studio.
Every shipped page earns at least two internal critic rounds before it is considered draft-complete. Each round produces specific named weaknesses, not vague suggestions, and a rewrite that addresses them. Round three or later on the same model hits diminishing returns; the Foreman pass takes over from there.
After two internal critic rounds, the artifact goes through a Foreman pass: a comparison against a reference corpus of 12 high-quality benchmark sites. The Foreman runs on a different model from the one that built the artifact. It produces 12 to 25 concrete demands rooted in what the best examples actually do, not abstract principles.
The completed artifact is logged with its tick number, decision letter, artifact path, and outcome. The state file is updated so the Director's next decision accurately reflects what now exists. The log is public at /factory/log/.
Quality Gates
Shipping the first draft is the beginning, not the end. Every product goes through a structured sequence of review before it earns the status of a finished artifact. The sequence is not optional and it is not abbreviated for time pressure. An artifact that has not cleared all five gates is marked as in-progress in the state file and picked up in a future tick.
Round 1
Five or more specific weaknesses named and addressed. Copy, structure, visual layout, and link integrity all in scope. General feedback does not count; each weakness must be specific enough to produce an edit.
Round 2
A full rewrite incorporating Round 1 demands. The brand brief is re-read. Studio chrome is stripped from the navigation. Photography is verified to be distinct from prior products. Word count is confirmed at target.
Round 3
Cross-comparison against 12 reference sites in the benchmark corpus. Demands are anchored in what best-in-class pages actually do, not what seems reasonable in isolation. Runs on a different model to avoid echo-chamber drift.
Round 4
A cross-model pass that brings a genuinely different perspective to the artifact. Catches blind spots that accumulate when the same model reviews its own output across multiple rounds. Produces a final list of residual weaknesses.
Round 5
Every href in the artifact is verified to return HTTP 200. Placeholder anchors, broken internal links, and dead-end routes are fixed and logged to /factory/director/broken-links.md. The artifact is not marked complete until this passes clean.
Brand Identity Protocol
The most common failure mode in portfolio studios is templating. Every product looks like a sub-page of the same corporate site. The navigation says "About" and "How It Works" -- concepts the customer does not care about. The palette is recycled. The typography is whatever the last developer found convenient. The result is a directory of products that all feel like the same product wearing different names.
We treat each product as an independent business. Before any code is written, the brand brief specifies the palette (three hex codes chosen for the customer audience, not the studio aesthetic), the typeface pairing (slab serifs for service businesses, geometric sans-serif for SaaS tools, humanist fonts for education and coaching products), and a one-line voice description. The brief is written to disk and read back on every subsequent polish or marketing pass, so the product identity stays consistent across ticks that may be days apart.
The studio's own design language (Fraunces + Inter, forest green and gold, cream backgrounds) is for studio pages like this one. Individual products are allowed -- and expected -- to look nothing like it.
Depth Before Breadth
A single landing page is the minimum viable artifact. It is not a finished product. A finished product has a pricing page that answers the cost objection before the call. It has a case studies page that shows the result before the prospect asks. It has a FAQ that handles the long-tail questions that would otherwise land in someone's inbox at 9 PM. It has an onboarding checklist that reduces activation friction for the first customer who actually signs up.
The depth gate is a hard rule enforced at the start of every tick: if fewer than half of all products in the portfolio have two or more sub-pages, new product creation is paused. The Director must choose a sub-page expansion, a marketing artifact, or a concept essay before it is allowed to ship another top-level shell. This rule exists because breadth without depth is decoration. The studio is not trying to have the most products. It is trying to have the most complete products.
Answers the cost question before the prospect gives up and closes the tab. Includes tier comparison, a most-popular callout, and a billing FAQ specific to the product's model.
Three before-and-after narratives showing the result the product delivers, with specific numbers where available. Social proof at the sub-page level, not just a logo strip on the homepage.
Ten to fifteen questions organized by stage: setup, daily use, billing, and integration. Structured for schema markup and search visibility. Reduces pre-sale support volume.
Step-by-step guide for a new customer's first week. Reduces activation friction, serves as a sales leave-behind, and gives the product a sense of operational depth beyond the landing page.
A structured list of tools the product connects to, with brief descriptions of what the integration does. Answers the "does it work with X" question that kills deals on discovery calls.
For products handling sensitive data: data residency, access controls, compliance posture, and incident response summary. Required for any enterprise or regulated-industry ICP.
The Automation Stack
The Factory is not a single AI model writing HTML in a loop. It is a layered system where different components have different roles and different information. The Director sets strategy. Workers execute construction. The Foreman enforces standards against an external benchmark corpus. Separating these roles is what prevents the echo-chamber problem: a single model reviewing its own work will always find it acceptable, because its quality threshold is calibrated to exactly the work it just produced.
Reads the full factory state on each tick. Enforces both hard gate rules. Selects the highest-value action from a 14-item menu. Outputs a structured JSON decision document that is logged and auditable. The Director does not write HTML; it decides what gets written.
Execute the Director's decision. Write HTML, run critic rounds, produce marketing copy, capture screenshots. Multiple workers can operate in parallel on different artifacts. Workers do not make strategic decisions; they receive a specification and execute it.
Compares a shipped artifact against a corpus of 12 reference sites. Returns 12 to 25 specific demands. Never runs on the same model that built the artifact. The cross-model gap is the point: it catches things the builder's model cannot see about its own output.
Scans every href in a shipped artifact and verifies it returns HTTP 200. Broken links, placeholder anchors, and redirects to / are flagged to a persistent broken-links log. Runs as the final gate before an artifact is marked complete.
Headless Chrome captures live screenshots of shipped pages at desktop viewport. The screenshots are embedded directly in product pages so every visitor sees what the product actually looks like rendered in a browser, not a Figma mockup or a stock illustration.
Every tick appends a permanent record: decision letter, action taken, artifact path, and outcome. The log is public at /factory/log/. Patterns are visible over time. If the system has been picking the same letter for six ticks, the log shows it and the Director is required to diversify.
Commissioning a Build
If you have a product concept that belongs in the portfolio -- a SaaS tool, a service-business automation, a niche AI application -- the intake process is direct. You bring the idea and the target customer. The Factory handles everything from brand brief to shipped landing page to marketing collateral.
The standard output for a commissioned build includes: a fully designed product landing page (1500 or more words, original photography, product-specific brand identity), a pricing sub-page, a case studies or social-proof sub-page, a 5-email cold outreach drip targeting the ICP, a 10-message LinkedIn outreach pack with named role targeting, and a 90-second video demo script ready for recording. That is the complete kit a founder needs to start selling before writing a single line of product code.
The build process runs in one overnight window. Review and polish passes happen in subsequent ticks over the following 24 to 48 hours. By the end of that window, the artifact has been through all five quality gates and is ready for the founder to take live under their own domain. The small footer credit is removed for commissioned builds. You own the artifact outright.
For inquiries, reach out via the About page. Wishdeal Studio takes a limited number of commissioned builds per month to protect the throughput of the production pipeline.
Frequently Asked Questions
Every artifact is produced by an AI model following a detailed brief, then reviewed and refined in multiple rounds. The process is analogous to a design agency using Figma: the tool is automated, the creative direction and quality standards are human-defined. The brand brief and quality gates are the human layer.
No. Each product has a documented brand brief with its own palette, typeface pairing, and voice. The brief is written before the first line of HTML and re-read on every subsequent pass. Products in the portfolio should look like independent businesses, not siblings on the same franchise site.
Each brand brief includes a thematic root tag. If three or more products share a root (e.g., "freelancer AI tools"), the Director is blocked from adding another in that category until it picks from a different area. Concept diversity matters more than action diversity.
The Foreman compares a page against 12 reference sites and returns specific demands. It runs on a different model from the one that built the artifact because same-model review accumulates a shared quality threshold over time. A different model has a different baseline and catches different gaps.
They stay in the queue. The Director tracks which artifacts have completed which review stages. Products that shipped before the current gate sequence was standard are upgraded gradually. No artifact permanently escapes the quality gates; it is a matter of which tick picks it up next.
Yes. That is the intent of the commissioned-build offering. The HTML, brand brief, email drip, outreach pack, and demo script are all deliverables. You own the artifact and can host it wherever you like. The footer credit is removed as part of the handoff.
For full-ownership acquisitions (every artifact transferred to your accounts, white-label), the "available to own" path is at /factory/own/. Different from the $99-199 adopt tier, which gives you the build files. The own tier is a negotiated acquisition.
The depth gate prevents the portfolio from staying shallow forever. The breadth gate prevents the system from spending all its ticks on polish and marketing without adding new products. They pull in opposite directions by design. Without both, the system optimizes for one metric at the expense of the other.
The builds index at /factory/builds/ shows every product with its current status. Products that have cleared all five quality gates are marked as ready for handoff. Products still in the review pipeline show their current gate status. The changelog at /factory/log/ has the full history for any specific artifact.
The bulk-render generator that produces product landing pages reads each product's hand-written dossier teaser, brand brief, and Adoptability data, then feeds the per-product context to a Claude CLI call to fill page placeholders. The LLM is condensing hand-written operator-voice inputs, not inventing copy from a generic prompt. That distinction is the difference between scale-with-fluff (every page sounds the same) and scale-with-fidelity (every page sounds like its product). The studio shipped 60 product pages in 45 minutes this way in iter 58, with a 90 percent first-pass quality rate confirmed by a 10-page audit in iter 59. The remaining 10 percent gets hand-polish, not regeneration.
It is caught by the next health-check cron and surfaced as a failed endpoint. The studio also runs an em-dash sweep cron every 15 minutes that catches HTML-entity violations across all rendered files. When iter 62 extended the sweep to catch HTML entities, it found 615 silent em-dashes across 265 files that had been violating the no-em-dash invariant for an unknown duration. The fix was at source: 12 generators patched to emit clean output, plus the sweep cron remains as the safety net. The invariant is "the bug is allowed to exist for at most 15 minutes anywhere in the catalog".
See it run live: /factory/cron-status/ shows all 124 cron jobs with last-run timestamps. The page regenerates every 15 minutes so the timestamps visibly shift if you reload.
See the content audit: /factory/quality-report/ shows live health checks, fake-proof audit status, and the 10 content invariants the system defends. Regenerates every 30 minutes.These three operator-voice essays go a level deeper than this page on specific aspects of the autonomous loop.
Cadence dynamics, source-fix discipline, audit-discovery, what we wish we built sooner. 12 min.
The audit-driven 4-iter arc that fixed every SOC 2 claim, training-corpus reference, and inflated customer count. 10 min.
Four reader contexts. Three axes most operators misread. The tier rule that beats the tier. 10 min.
All 11 essays at /factory/playbooks/.
See what the Factory has shipped, or reach out to commission a build for your next product idea.
Browse the Builds Get in Touch