Scout: Field to Proven Opportunities

# Title: Scout: Field to Proven Opportunities
#
# Public workflow: the phase BEFORE anyone has an idea. Take a field/market,
# mine it for real evidenced pain through three blind parallel sweeps
# (communities, incumbent reviews, demand+disconfirmation), map the pains into
# job stories, then adversarially vet the top candidates until the evidence
# survives a genuine kill attempt — ending in ranked opportunity briefs, or an
# honest no-signal verdict. Chain the winner into `launch`. Public roles only.
#
# Note: chains several research-heavy runs — token/time-heavy by design.

name        = "scout"
description = "Mines a field for evidenced pain via parallel community, review, and demand sweeps, vets candidates adversarially into ranked briefs or an honest no-signal verdict. Chain the winner into launch."

# ── 1. Parallel evidence sweep — three blind lenses ──────────────────────────
[[steps]]
name     = "sweep"
parallel = true

  [[steps.run]]
  name    = "communities"
  role    = "launch:explore"
  session = "fresh"
  retries = 1
  prompt  = """
Passive pain harvest in the communities of this field. NO product ideas yet —
you are collecting raw evidence only.

<field>
{{input}}
</field>

1. Find the watering holes: 5–10 communities (subreddits, niche forums, groups,
   Q&A sites) that are active (check member counts and recent posts) and
   on-topic for this field.
2. Mine COMPLAINT threads, not advice threads, with pain-phrase searches in the
   audience's own words: "is there a tool that", "how do I", "I hate", "why is
   it so hard to", "am I the only one who", "tired of", "workaround for", and
   manual-hack confessions ("I built a spreadsheet to…").
3. Log each hit: verbatim quote · source URL · date · tag (PAIN / JARGON /
   WORLDVIEW / RECOMMENDATION). Weight the last 6–12 months; note emotional
   relief words (easier, faster, finally).

Output the raw pain log grouped by community, plus the audience-language
glossary (their jargon, verbatim).
"""

  [[steps.run]]
  name    = "reviews"
  role    = "launch:explore"
  session = "fresh"
  retries = 1
  prompt  = """
Incumbent review mining for this field. NO product ideas yet — evidence only.

<field>
{{input}}
</field>

1. Identify the incumbent tools/services people in this field actually pay for
   (and the big free defaults).
2. Mine their reviews on G2/Capterra/Trustpilot/app stores: 3–4★ first (the
   nuanced, actionable complaints), then 1★ (rage and edge cases). 5★ tells you
   what already works — note it as the bar.
3. Cluster complaints into recurring themes with counts and the share of an
   incumbent's negative reviews each theme represents (≥15% = a real weakness).
4. Note zombie competitors (abandoned, outdated, users begging for maintenance)
   and pricing-model complaints.

Output the theme clusters with verbatim quotes, URLs, dates, counts, and
per-incumbent weakness percentages.
"""

  [[steps.run]]
  name    = "demand"
  role    = "assistant:researcher"
  session = "fresh"
  retries = 1
  prompt  = """
Demand and disconfirmation sweep for this field — cited findings only, no
opinions, no product ideas.

<field>
{{input}}
</field>

(a) DEMAND PROXIES: community sizes and growth velocity, search-trend
    trajectory for the field's core terms (flag one-off news spikes as noise),
    complaint-thread frequency. Real proxies only — no fabricated TAM.
(b) DISCONFIRMATION (the point is to find reasons this field is NOT fertile):
    the graveyard check — startups/products that already tried here and died,
    and why; the solved-problem check — dominant free or good-enough solutions
    including the status quo (spreadsheets, hiring someone, doing nothing);
    structural resistance — regulation, network effects, "big player would
    ship it as a feature" risk.
(c) TIMING: regulation, platform, or technology shifts creating fresh needs
    right now.

Cite every claim with its source and date.
"""

# ── 2. Pain map — job stories, triangulated, scored ──────────────────────────
[[steps]]
name    = "map"
role    = "launch:explore"
session = "fresh"
retries = 1
prompt  = """
Synthesize the three independent evidence streams below into a pain map, then
derive ranked opportunity candidates.

<field>
{{input}}
</field>

<communities>
{{communities}}
</communities>

<reviews>
{{reviews}}
</reviews>

<demand>
{{demand}}
</demand>

Rules:
- Cluster into 5–8 pain themes. Each theme carries: a job story ("When
  [situation], I want to [motivation], so I can [outcome]"), the exact segment,
  the current workaround, and its evidence pack (verbatim quotes with URLs,
  mention counts, review-theme percentages).
- Triangulation bar: keep only themes that appear unprompted in at least TWO
  independent source types (communities / reviews / demand), recent-weighted.
  List what you discarded and why — discards are findings too.
- Evidence hierarchy: a workaround already built or money/time already spent >
  repeated specific complaints > stated preferences. Tag each theme painkiller
  or vitamin accordingly.
- Score each surviving theme: Severity (1–5) × Frequency (1–5) × Reach (1–5),
  plus the explicit willingness-to-pay signal, noting how strongly "do nothing"
  currently wins.
- Derive the TOP 3 opportunity candidates, each stated as: SEGMENT +
  STRUGGLING MOMENT + CURRENT WORKAROUND + WHAT BETTER LOOKS LIKE, ranked with
  scores and evidence.
"""

# ── 3. Adversarial vetting loop — evidence must survive a kill attempt ───────
[[steps]]
name           = "vet"
loop           = true
max_iterations = 3
exit_when      = { output = "skeptic", matches = '(?m)^SCOUT: SOLID' }

  [[steps.run]]
  name    = "strengthen"
  role    = "launch:explore"
  session = "continue"
  retries = 1
  prompt  = """
First round: finalize the top-3 candidate dossiers from the pain map — close
the evidence gaps you can already see with targeted searches/browsing. Later
rounds: the skeptic's audit arrives as your input — fetch the exact missing
evidence it names, or honestly downgrade/replace the candidate. Never pad weak
evidence with adjectives; a downgraded candidate with true evidence beats an
inflated one.

<field>
{{input}}
</field>

<map>
{{map}}
</map>

Output the current top-3 dossiers, each with its full evidence pack.
"""

  [[steps.run]]
  name    = "skeptic"
  role    = "launch:validate"
  session = "fresh"
  prompt  = """
Refute-first audit. Your job is to find reasons these are NOT real
opportunities — if you cannot kill one with evidence, it earns PROVEN.

<field>
{{input}}
</field>

<candidates>
{{strengthen}}
</candidates>

Attack each candidate on all six:
1. Evidence quality (Mom Test) — past behavior and money/time actually spent,
   or hypothetical enthusiasm and compliments? Verbatim, dated, sourced?
2. Triangulation — unprompted across ≥2 independent source types, recent? Or a
   single viral post inflated into a trend?
3. Solved-problem check — a free or good-enough incumbent with network effects
   already owns this?
4. Graveyard collision — did predecessors die here for structural reasons that
   still hold?
5. Do-nothing test — is the status quo winning cheaply because the pain is
   tolerable?
6. Painkiller reality — an acute struggling moment, or a vitamin dressed up?

Verdict per candidate: PROVEN (survives all six) / WEAK (name the exact
missing evidence) / REFUTED (name the kill reason). Base every verdict on the
evidence in front of you, not vibes.

End with exactly one line: `SCOUT: SOLID` (at least one PROVEN candidate,
every verdict evidence-based) or `SCOUT: RETRY`. Nothing after it.
"""

# ── 4. Outcome — honest branch on the vetting verdict ────────────────────────
[[steps]]
name        = "outcome"
conditional = true
condition   = { output = "skeptic", matches = '(?m)^SCOUT: SOLID' }
on_match    = ["briefs"]
on_no_match = ["nosignal"]

  [[steps.run]]
  name    = "briefs"
  role    = "launch:validate"
  session = "fresh"
  prompt  = """
Write the final Opportunity Brief for every PROVEN candidate (one-line closing
notes for WEAK/REFUTED ones). Each brief contains:

- One-line job story, and the exact segment
- The struggling moment/trigger, the current workaround, and what they already
  spend on it (time/money)
- Why now (the timing/demand signal)
- Evidence pack: source count, top verbatim quotes with URLs, review-theme
  percentages
- Competition map including the spreadsheet, the intern, and "do nothing"
- Painkiller/vitamin verdict, opportunity score, and rank
- Pre-committed, falsifiable KILL CRITERIA with a deadline for the next
  validation stage ("X buyers doing Y within 30 days, or kill — no extensions")

<candidates>
{{strengthen}}
</candidates>

<audit>
{{skeptic}}
</audit>

Close with one line: feed the chosen brief into `octomind workflow launch` to
take it from opportunity to go-to-market.
"""

  [[steps.run]]
  name    = "nosignal"
  role    = "launch:validate"
  session = "fresh"
  prompt  = """
The vetting loop ended with no PROVEN opportunity — that is a finding, not a
failure; an honest no saves months. Report: the strongest partial candidates
with their exact evidence gaps, what was REFUTED and why (save the next person
the trip), and precisely what evidence would change each verdict.

<candidates>
{{strengthen}}
</candidates>

<audit>
{{skeptic}}
</audit>

Start with exactly: `NO PROVEN OPPORTUNITY — evidence below the bar.`
"""