Skip to main content
This is the whole thing in one page: take an idea, find out if people want it, and turn it into everything you need to launch. Each step is one call. They all read from the one demand report you produce first, so every recommendation traces back to a real quote you can open. Set a provider key first (see Installation), then: 
from metalworks import Metalworks

mw = Metalworks()

​
1. Is there demand?

Start with one sentence about your idea. metalworks reads real conversations — Reddit by default, and Hacker News, the web, or your own data when you add them — and gives you a go/no-go plus the actual needs people voiced. 
research = mw.research(
    "an affordable, jitter-free focus supplement for developers",
    subreddits=["Nootropics", "Supplements"],   # omit to let it pick
)
report = research.demand

print(report.verdict)                 # e.g. "Strong demand — 312 distinct voices..."
for c in report.ranked_clusters:
    print(c.distinct_author_count, "people:", c.claim)
If the demand is thin, it says so here — and the later steps stay honest about it instead of inventing an opportunity. Everything below runs on this report. The run also builds up a durable corpus you can grow and refresh this report against later. → Demand research

​
2. Find your angle and your competitors

positioning = mw.positioning(research)
print(positioning.positioning_statement)   # who it's for, and why it's different

competitors = mw.competitors(research)
for rival in competitors.competitors:
    for gap in rival.gaps:                  # an opening, backed by a real complaint
        print(rival.name, "misses:", gap.claim)
Your positioning is built from the unmet needs in the report; each competitor gap is a real complaint someone posted. → Positioning & competitors

​
3. Decide what to build, and write the site

surface = mw.surface(research, positioning)      # web? mobile? CLI? — and why
ux = mw.ux(research, positioning, surface.chosen)  # the 3-5 screens you need

site = mw.site(research, positioning)            # marketing copy, every line a real quote
html = mw.render_site(site, research)            # a self-contained index.html
open("index.html", "w").write(html)
The marketing copy isn’t AI prose — each section is a verbatim quote from a real user, with a link back to the thread. → Design & marketing site

​
4. Turn it into a build plan

metalworks writes the spec, not the product. It maps the demand to a feature list and scaffolds a project your own coding agent (Claude Code, Cursor, etc.) then builds. 
spec = mw.build_spec(research, positioning, surface.chosen)
for feature in spec.features:
    print(feature.title, "—", feature.rationale)   # each tied to real demand

mw.scaffold(spec, research, "./my-startup")        # writes the build harness
The scaffold includes a frozen list of the real quotes behind every feature, so whatever your agent builds stays true to what people actually asked for. → Build spec

​
5. Launch and grow

assets = mw.launch(research, positioning)   # Product Hunt / Show HN / X drafts (never posts)
plan = mw.channel_plan(research)            # a step-by-step launch checklist you run

content = mw.content_plan(research)         # an SEO/content plan, one page per real need
Launch copy is drafting-only — metalworks never posts anything. Each claim in a draft is backed by a quote; the content plan lists the real threads to cite so people and AI search engines find you. → Launch assets · Content & SEO

​
That’s the loop

One idea → one grounded report → positioning, design, a build plan, launch copy. Because it all reads from the same report, every recommendation links back to a real quote you can open and read for yourself. That’s the difference between this and a tool that just makes things up — see why you can trust the output. Run metalworks init first and metalworks remembers all of this — each run is saved under .metalworks/ and later steps chain off its report_id instead of re-running research. See Projects & memory. Prefer the command line or an agent? The exact same flow is available in the CLI, the MCP tools, and the Claude Code plugin.