Skip to main content
Decide what to build and write the site — both straight from real demand. With a demand report and your positioning in hand, two more calls cover design: surface + screens (what kind of product to build and the screens it needs) and a marketing site (a page whose every claim is a verbatim quote from a real user). Both read from the same report, so nothing is made up.

​
Surface & screens — what to build

from metalworks import Metalworks

mw = Metalworks()
research = mw.research("an affordable, jitter-free focus supplement for developers")
positioning = mw.positioning(research)

surface = mw.surface(research, positioning)       # web? mobile? CLI? — and why
print(surface.chosen, "(runner-up:", surface.runner_up, ")")
for dim in surface.rubric:
    tag = "assumption" if dim.is_assumption else "backed by quotes"
    print(f"  {dim.name}: {dim.finding} [{tag}]")

ux = mw.ux(research, positioning, surface.chosen)  # the 3-5 screens you need
for s in ux.screens:
    flag = "validated" if s.validated else "HYPOTHESIS"
    star = "★" if s.serves_wedge else ""
    print(f"  {s.name}: {s.purpose} [{flag}] {star}")
From the command line: 
metalworks research surface <report-id>
mw.surface(...) picks the right kind of product to build — web, mobile, cli, browser_extension, and so on — with a runner-up and the trade-offs. It judges the call on five fixed questions: where your users already are, how technical they are, how often they’d use it, whether it needs realtime or hardware, and how you’d distribute it. mw.ux(...) then sketches the 3-5 screens you need, each with a one-line purpose and a single primary action.

​
What you give it / what you get back

FieldWhat it is
surface.chosen / runner_upThe recommended product type and the second choice.
surface.rubric[].findingWhat the evidence says on each of the five questions.
surface.rubric[].is_assumptionTrue when no real quote backs that finding — a stated guess, labeled.
surface.confidenceHow sure metalworks is, set from how many questions it could actually back with quotes.
ux.screens[].validatedTrue when at least one real voice asked for that screen; False means it’s a hypothesis to test.
ux.screens[].serves_wedgeTrue when the screen directly delivers your positioning.
Each finding is checked against the real quotes in your report. Anything a quote backs is labeled as such; anything it can’t back is flagged as an assumption rather than dressed up as a fact. Screens nobody asked for ship as honest hypotheses, never silent requirements. This module deliberately stops at text and structure — it recommends what to build and the screens, not the visual design.

​
When the result is thin

This is the step with the least to stand on, so metalworks is deliberately strict: when the signal is thin, it under-claims — marking more findings as honest assumptions and dropping the whole recommendation to partial=True — rather than pinning a confident “the users are here” on a loosely related comment.

​
Marketing site — every claim is a real quote

site = mw.site(research, positioning)             # marketing copy, every line a real quote
for sec in site.sections:
    print(sec.role, f"[{sec.provenance}]", "→", sec.copy[:60])

html = mw.render_site(site, research)             # a self-contained index.html
open("index.html", "w").write(html)
From the command line: 
metalworks research site <report-id>
mw.site(...) drafts a small marketing page; mw.render_site(...) turns it into one self-contained index.html you can open. The copy isn’t AI prose — every claim-bearing line is a word-for-word quote from a real user, and each one renders with a footnote linking back to the Reddit thread it came from, so any visitor is one click from the real comment.

​
What you give it / what you get back

FieldWhat it is
site.sections[].roleWhat the section does: hero, feature, objection, pricing, social_proof, or cta.
site.sections[].copyThe text. For a claimed section it contains a verbatim fragment of a real quote.
site.sections[].provenanceverbatim (a cited real quote) or connective (claim-free glue between sections).
The hero is built on the need the most distinct people raised — the broadest demand, not the loudest single post. metalworks may add short connective lines to bridge sections, but those carry no claims at all: any glue line that sneaks in a number or a superlative like “best” or “only” is dropped, so unsourced claims can’t slip in.

​
When the result is thin

metalworks drops anything it can’t back with a real quote. If a line doesn’t exactly match a stored comment, that section is removed rather than shipped on the model’s word. If nothing matches at all, the site comes back empty with partial=True and a caveat — never an invented section, never a crash.

​
Next

You know what to build and you have a site. Now turn it into a build plan and launch: → Build spec · Launch assets · the full walkthrough · why you can trust the output