Demand research

Find out whether people actually want your idea — backed by real conversations, not a guess. Give metalworks one sentence about what you want to build. It reads real conversations across your sources — Reddit, Hacker News, the web, your own data — and hands back a read on how strong the demand is, plus the specific needs people voiced, each one backed by a real quote you can open and read. (The explicit GO / PIVOT / NO-GO call is assess’s job, not this report’s.) This is the report every other step runs on, so do it first.

Run it

/demand-report an affordable, jitter-free focus supplement for developers

from metalworks import Metalworks

mw = Metalworks()                       # provider key from the env, or keyless via Claude Code login
research = mw.research(
    "an affordable, jitter-free focus supplement for developers",
    subreddits=["Nootropics", "Supplements"],   # omit to let it pick
)
report = research.demand                # .research() returns a bundle; the report is .demand

print(report.demand_summary)            # how strong is the demand? a one-line summary
for c in report.ranked_clusters:
    print(c.distinct_author_count, "people:", c.claim)
    for q in c.quotes:                  # the real comments behind the claim
        print("   ", q.source_url, "→", q.text[:80])

metalworks research run --question "an affordable, jitter-free focus supplement for developers"

A run prints something like:

Strong demand — 312 distinct voices across 9 themes.
  187 people: jitter and crashes from caffeine-based focus products
   https://reddit.com/r/Nootropics/comments/... → "every pre-workout style focus thing wrecks me by 2pm..."
   94 people: hard to find something that isn't priced like a luxury subscription

What you give it / what you get back

You give it one sentence and, optionally, a list of subreddits. You get back a report on research.demand:

Field	What it is
`report.demand_summary`	A one-line demand-strength summary (the go/no-go is `assess`’s job).
`report.ranked_clusters`	The real needs people voiced, ranked by how many different people raised each — not by how viral one post was. Each cluster has a `claim`, a `distinct_author_count`, and its `quotes`.
`report.web_findings`	Supporting facts pulled from the web, each carrying its real source URL.
`report.partial`, `report.caveat`	Set when part of the run came up short (see below).

Every quote is the exact text of a real stored comment, matched — not paraphrased. Every count is the result of counting distinct authors, not a model’s estimate. If a finding can’t be tied to a real quote or a real source URL, metalworks drops it rather than shipping it. That’s why a “go” here is one you can defend line by line — see why you can trust the output.

Want a richer brief first?

mw.research(question) works from one sentence. If you’d rather refine the question, set success criteria, and let metalworks suggest subreddits before it runs, plan first:

brief = mw.plan("Should I build a focus-supplement brand for developers?")
report = mw.research(brief).demand      # hand the brief straight back

Scope what it reads

mw.research(
    "...",
    subreddits=["Nootropics"],
    time_window_months=6,   # how far back to read (default 12)
    per_sub_limit=500,      # cap submissions pulled per subreddit
    max_findings=20,        # cap web findings
)

Read from more sources

metalworks reads Reddit by default, and can also read Hacker News, the web, or your own data. Name the sources you want for a run:

metalworks research run --question "..." --source reddit --source hackernews

See Sources to choose what to read from, or use your own data to load conversations you already have.

Update a report later

Each run saves what it read, so your research data builds up over time. Collect more, then update the report to see what changed:

metalworks corpus add --source hackernews -q "..."
metalworks research refresh <report-id>     # an updated report + what changed

When the result is thin

metalworks tells you when there isn’t a real opening instead of inventing one. If the demand is weak or a part of the run degraded — say it couldn’t reach the web — the report comes back with partial=True and a plain-English caveat saying what happened. The later steps stay honest about it too, rather than building a launch plan on signal that isn’t there.

You have the report. Decide whether it’s worth building with the validation loop (GO / PIVOT / NO-GO), or turn it into what you need to launch: → Validation loop · Positioning · Competitors · Build spec · the full walkthrough

Get started

Build a startup

Your project

Research

Design

Build

Distribution

How it works

Run it

What you give it / what you get back

Want a richer brief first?

Scope what it reads

Read from more sources

Update a report later

When the result is thin

Next

​Run it

​What you give it / what you get back

​Want a richer brief first?

​Scope what it reads

​Read from more sources

​Update a report later

​When the result is thin

​Next

Run it

What you give it / what you get back

Want a richer brief first?

Scope what it reads

Read from more sources

Update a report later

When the result is thin

Next