Business idea evaluator

Business Idea Evaluator

Professional analytical system. Not a motivational coach. Evaluate the idea, not the user's enthusiasm.

Non-negotiable rules

1. Never evaluate before idea is confirmed — complete Phase 1 first.
2. Never simulate subagents in main context — launch all 18 via Task/Agent tool.
3. Never use simple arithmetic mean for final score — use scripts/calculate_brs.py.
4. Never flatter — see references/forbidden-phrases.md.
5. Never claim "no competitors" without expert 01 completing deep search.
6. Mark evidence status on every material claim — see references/evidence-status.md.
7. If web search unavailable, state it once and downgrade all market claims to hypotheses.

Pre-flight — verify subagents before Phase 2

Mandatory. Before launching any subagent, run from project root:

bash .agents/skills/business-idea-evaluator/scripts/verify_subagents.sh

If verification fails, install and re-verify:

bash .agents/skills/business-idea-evaluator/scripts/install-agents.sh

Do not proceed to Phase 2 until .cursor/agents/, .claude/agents/, .agents/agents/, and .codex/agents/ each contain all 18 subagents. First-time setup: bash .agents/skills/business-idea-evaluator/scripts/bootstrap.sh

Anti-simulation gate (Phases 2–3)

These actions are forbidden in the main agent context:

• Writing expert or math JSON without a completed Task/Agent invocation for that agent
• Summarizing what an expert "would say" instead of launching the subagent
• Skipping failed agents without failed_agents entry and retry

Required: 12 + 6 = 18 separate Task/Agent invocations, each with subagent_type matching the agent name (e.g. biz-eval-01-analog-research). Launch all agents in a phase within one message (parallel).

After each phase, briefly confirm: «Запущено N/N субагентов, получено N/N JSON-ответов» before proceeding.

Phase 1 — Idea discovery (5 questions)

See references/discovery-protocol.md for slot map and examples.

Ask exactly one question per message. Wait for answer before next question.

#	Goal	First question depends on
1	Who is the paying customer	Your reading of user's initial idea
2	Specific pain + frequency	Answer to Q1
3	Current workaround + cost of pain	Answer to Q2
4	Product format + what customer pays for	Answer to Q3
5	Launch market + MVP constraints	Answer to Q4

Rules for questions:

• Max 2 sentences per question.
• No "tell me more" or open-ended prompts.
• Each question must extract one concrete business-model element.
• Track answers internally; do not dump full Q&A into final report.

After Q5, output only:

Я понял идею так: [one dense paragraph covering: audience, pain, solution,
delivery format, monetization, launch market, differentiation, first MVP]

Then ask: «Подтверждаешь такую формулировку или нужно поправить?»

• If user corrects → update paragraph, ask confirmation again.
• Do not proceed to Phase 2 until explicit confirmation («да», «подтверждаю», «верно», etc.).

Save confirmed idea as IDEA_BRIEF for all subagents.

Phase 2 — Expert layer (12 subagents)

Launch all 12 in parallel (one message, 12 Task/Agent invocations). Read references/expert-evidence-package.md for output schema.

Agent file	Role
biz-eval-01-analog-research	Deep analog & substitute search
biz-eval-02-pain-demand	Real pain & willingness to pay
biz-eval-03-audience-behavior	Buyer behavior & acquisition
biz-eval-04-market-country	Country/market realities
biz-eval-05-trends-longevity	Trend & 1–5y durability
biz-eval-06-monetization	Revenue model & unit economics hints
biz-eval-07-marketing	Channels & trust barriers
biz-eval-08-implementation	MVP & technical feasibility
biz-eval-09-legal-platform	Legal/platform/ethical risks
biz-eval-10-competitive-moat	Copyability & defensibility
biz-eval-11-launch-zero	7/30/90 day launch plan
biz-eval-12-red-team	Attack the idea

Install agents if missing: run bash .agents/skills/business-idea-evaluator/scripts/install-agents.sh from project root.

Cursor / Claude — Task invocation pattern

One message, 12 parallel Task calls. Example for expert 01:

subagent_type: biz-eval-01-analog-research
prompt: |
  IDEA_BRIEF:
  <confirmed paragraph>

  TASK: Execute your role per agent definition. Use web search if available.
  Return ONLY valid JSON matching the schema in
  .agents/skills/business-idea-evaluator/references/expert-evidence-package.md
  for agent_id "01". Include sources with URLs when found.

Repeat for biz-eval-02-pain-demand … biz-eval-12-red-team in the same message.

Codex — explicit spawn

Codex does not auto-spawn. Tell the user once, then run:

1. «Spawn 12 expert agents in parallel» — agents in .codex/agents/*.toml
2. After merge: «Spawn 6 math agents in parallel»

Prompt template for each expert

IDEA_BRIEF:
<paste confirmed paragraph>

TASK: Execute your role per agent definition. Use web search if available.
Return ONLY valid JSON matching the schema in references/expert-evidence-package.md
for your agent_id. Include sources with URLs when found.

Collect all 12 JSON outputs. Merge into Expert Evidence Package (single JSON file or structured object). Write to biz-eval-evidence.json in workspace temp if needed.

Gate: Do not start Phase 3 until all 12 expert JSONs are received. If one fails, retry once; if still failing, note agent_failed in package and continue with penalty.

Phase 3 — Math layer (6 subagents)

Launch all 6 in parallel after Expert Evidence Package is complete. Each math subagent receives the full Expert Evidence Package and works independently of the other math subagents — it does not re-research the market, it processes the package numerically (counts, ranges, penalties). Independence is enforced by giving each its own context (separate Task/Agent invocation), not by prompt wording.

Agent	Input	Role
biz-eval-13-evidence-stats	Full package	Evidence & source quality indices
biz-eval-14-math-model	Package + agent 13	Formula design & variable weights
biz-eval-15-scenario-probability	Full package	4-scenario probability map
biz-eval-16-unit-economics	Full package	CAC/LTV/churn model
biz-eval-17-sensitivity	Full package	Top sensitivity parameters
biz-eval-18-experiments	Full package + math outputs	Validation experiments

Prompt template for math agents

EXPERT_EVIDENCE_PACKAGE:
<paste full JSON from Phase 2>

TASK: Execute your role. Return ONLY valid JSON per your agent schema.
Do not re-research market — process the package mathematically.

Phase 4 — Deterministic BRS calculation

1. Merge expert scores + math agent outputs into biz-eval-input.json (see references/scoring-formula.md; structure defined in references/evidence-package.schema.json).
2. Validate, then score:

python3 .agents/skills/business-idea-evaluator/scripts/validate_evidence_package.py biz-eval-input.json
python3 .agents/skills/business-idea-evaluator/scripts/calculate_brs.py biz-eval-input.json

3. Use script output as authoritative Business Reality Score. Main agent explains results; does not override numbers.

The script returns: base/min/max BRS, confidence, evidence_index, source_quality_index, hypothetical flag, blocking_caps_applied, main_blocking_risk, main_growth_factor, main_uncertainty_factor, the nine factors, a monte_carlo block (BRS distribution from seeded simulations whose spread is driven by the evidence base), a computed probability_map (verdict-band probabilities), and input warnings. Tune runs with --simulations N --seed N.

A worked input/output pair lives in references/example-input.json and references/example-output.json (also examples/sample-input.json).

If script fails, fix JSON and retry. Do not invent score manually.

Phase 5 — Final report

Render using references/report-template.md.

Verdict must map to BRS:

• ≥65 — test in narrow segment (not "launch fully")
• 45–64 — change model or segment before spending
• 25–44 — weak; only cheap experiments justified
• <25 — do not launch in current form

Always separate: proven by sources | statistical | hypothesis | needs verification.

Subagent installation

Agents live in skill bundle at agents/. Full bootstrap (install + verify + tests):

bash .agents/skills/business-idea-evaluator/scripts/bootstrap.sh

Install only:

bash .agents/skills/business-idea-evaluator/scripts/install-agents.sh
bash .agents/skills/business-idea-evaluator/scripts/verify_subagents.sh

Targets: .cursor/agents/, .claude/agents/, .agents/agents/ get Markdown; .codex/agents/ gets native TOML via build_codex_agents.py. Symlinks: .cursor/skills/business-idea-evaluator and .claude/skills/business-idea-evaluator → .agents/skills/business-idea-evaluator. User-level ~/.cursor/agents/, ~/.claude/agents/, ~/.codex/agents/ also receive copies.

Resume / partial runs

If user returns after discovery: skip Phase 1 if IDEA_BRIEF confirmed earlier in thread. If experts done but math pending: start Phase 3 only.

References

• expert-evidence-package.md — per-agent JSON schemas
• evidence-package.schema.json — machine-checkable JSON Schema for the merged input
• scoring-formula.md — BRS formula & blocking rules
• report-template.md — final output format
• evidence-status.md — source grading
• forbidden-phrases.md — banned language
• example-input.json / example-output.json — worked example