--- name: consumer-ai-factory-opportunity-evaluation description: Evaluate and log Consumer AI Company Factory venture opportunities using painpoint mining, keyword/search intent, specialist distribution, cheap tests, and Obsidian build logs. tags: - consumer-ai - venture-validation - opportunity-evaluation - happycouple - obsidian - partnerships --- # Consumer AI Factory Opportunity Evaluation ## When to use Use this skill when Antoine asks to: - Evaluate a new consumer AI app idea or niche. - Decide what is worth testing next in the Consumer AI Company Factory. - Compare venture opportunities beyond HappyCouple. - Turn painpoint mining, keyword research, or creator/specialist research into a venture decision. - Maintain factory build logs or reusable venture artifacts in Obsidian. - Reason about specialists/creators as distribution partners. ## Core doctrine The factory should not ask: ```text Is this a good idea? ``` It should ask: ```text Which pain has enough signal to deserve the next cheapest test? ``` Operating principle: ```text Research selects the next test. Tests decide the next build. Usage decides the next company. ``` A strong Consumer AI opportunity should have all five: 1. A panic/shame sentence users already say, e.g. "I don't know what to say", "I can't start", "Is this quote bullshit?", "What does school need from me this week?", "Do I need the vet now?" 2. A natural input: voice/text brain dump, forwarded email, uploaded PDF, photo/video, screenshot, quote, report, or message thread. 3. A named output artifact: next-conversation script, unstuck card, Sunday school digest, 48-hour parent care plan, quote sanity report, vet-ready symptom summary. Avoid vague "insights" or "dashboard" outputs. 4. A real non-ChatGPT advantage: expert frameworks users would not know how to prompt, persistent memory/tracking, multiplayer/collaborative loops, structured workflow completion, proprietary/domain rubrics, specialist data/context, source trails, reminders/follow-up, human/expert review, safety boundaries, or substantially better consumer UX. 5. A believable channel: Reddit/community pain clusters, search intent, creators/specialists, professional referral partners, or existing audience aggregators. 6. A credible growth/retention loop: acquisition, activation, retention, share/referral, personalization/stored value, and monetization assumptions must be explicit before moving from probe to MVP. Load `consumer-growth-model-and-loop-map` when the decision depends on web vs mobile, growth loops, retention, channel-model fit, or whether a product is just a demo. Hard gate: reject "ChatGPT with a nicer UI" ideas. If a user can plausibly paste the same context into ChatGPT and get most of the value in one turn, the idea is weak unless the product compounds through frameworks, tracking, collaboration, workflow completion, or trusted distribution. HappyCouple's bar is the reference: Gottman-informed reasoning, relationship tracking, partner/two-player loop, and a longitudinal relationship map make it more than generic advice chat. Opportunity research is a core compounding factory capability, not a one-off brainstorm. Every research run should improve the machine itself: better sources, sharper filters, stronger scoring, better cheap-test design, and clearer learning capture. For consumer-agent market maps, treat the research artifact as both an internal opportunity-selection tool and an external distribution asset when AgentAppStore is involved. The preferred pattern is: ```text research landscape → canonical vault artifact → weekly refresh → AgentAppStore category/comparison/teardown content → qualified traffic and venture signal → updated opportunity decisions ``` If Antoine asks what to do with a broad consumer-agent landscape, do not default to launching a new probe. First decide whether the better next action is to clean/ship the distribution surface itself — e.g. AgentAppStore positioning, category pages, comparison pages, snippets, and venture listings — so the research starts compounding. Direct output standard: ```text Do not return “10 cool ideas.” Return 10 ranked opportunities with evidence, cheapest next test, kill criteria, and a better research process than before. ``` ### Fast ideation mode When Antoine asks for quick idea generation such as “give me 10 consumer AI ideas around X,” do not silently turn the request into a long live-research project. Default to a concise ranked hypothesis list grounded in existing factory doctrine, with clear caveats that these are pre-research opportunities. Include for each idea: pain sentence, input, output artifact, AI/non-ChatGPT advantage, cheapest test, and kill criterion. Only run live research, save Obsidian artifacts, or launch deeper scorecards when Antoine asks to validate, compare seriously, or choose what to build next. If the user has recently complained about speed, lead with the answer and avoid multi-step tool audits unless current external facts materially affect the decision. After every opportunity research run, explicitly capture: - Which sources produced real signal vs noise. - Which pain patterns repeated across communities/search/reviews. - Which app ideas looked clever but failed evidence checks. - Which scorecard dimensions were predictive or weak. - Which creator/specialist archetypes looked distribution-rich. - Which cheap tests would expose demand fastest. - What should change in the next research prompt/process. ## Key Antoine preferences Substantial reusable research/strategy artifacts should be proactively saved or updated in Obsidian, not only summarized in memory. For every meaningful factory step, update the build log. Messaging quality is a core factory capability. Every important outbound message must pass the 3-clarity test: 1. What do we do? 2. What do we want from them? 3. How does it benefit them? Personalization supports clarity; it does not replace clarity. If a message is deeply researched but unclear on product, ask, or recipient benefit, rewrite it before saving/sending. Factory-level doctrines, playbooks, and templates should be promoted out of venture folders when they are reusable across future Consumer AI ventures. ## Athlete career / recruiting probes When evaluating athlete career-advancement ideas — recruiting profiles, player CVs, highlight-video tools, trial matching, scout/agent outreach, college recruiting, or similar — do not frame the product as a generic AI coach or “AI agent” too early. The stronger first-use artifact is usually a **career action pack**: ```text real player context + video link -> evaluable profile/CV + highlight checklist + targeted outreach + realistic level/path guidance ``` Key research lesson from Player Agent / soccer-player-profile work: the profile alone is weaker than the bundle of profile + highlight critique + outreach pack + target-level guidance. Users are often asking: “What do I send, to whom, and am I aiming at the right level?” Split language by market: - US college/parent lane: `soccer recruiting`, `soccer resume`, `college soccer recruiting`, `coach email`, `ID camp decision`. - Global/UK pro-path lane: `football trial`, `football agent`, `football CV`, `open football trials`, `trial outreach`. Trust guardrails: avoid “get signed,” “guaranteed trials,” or implying licensed representation. Add scam/red-flag checks when the product touches agents, trials, or overseas opportunities. Session-specific reference: `references/player-agent-soccer-profile-research-2026-07-07.md`. ## Existing Obsidian artifacts Canonical factory folder: ```text /Users/antoinelevy/Documents/Obsidian Vault/Consumer AI Company Factory ``` Opportunity evaluation system: ```text /Users/antoinelevy/Documents/Obsidian Vault/Consumer AI Company Factory/01 Playbooks/Consumer AI Company Factory - Opportunity Evaluation System.md ``` Build log: ```text /Users/antoinelevy/Documents/Obsidian Vault/Consumer AI Company Factory/00 Operating System/Consumer AI Company Factory - Build Log.md ``` Factory task status: ```text /Users/antoinelevy/Documents/Obsidian Vault/Consumer AI Company Factory/00 Operating System/Factory Task Status.md ``` Opportunity research artifacts: ```text /Users/antoinelevy/Documents/Obsidian Vault/Consumer AI Company Factory/04 Opportunity Research/ ``` HappyCouple organic distribution sprint: ```text /Users/antoinelevy/Documents/Obsidian Vault/Consumer AI Company Factory/05 Ventures/Venture 001 - HappyCouple/03 GTM/HappyCouple - Organic Distribution Sprint.md ``` ## Research stack discipline When Antoine asks how to make research better, first separate **missing discovery surfaces** from **existing venture-specific pipelines**. Do not recommend adding a pipeline that already exists; verify the current artifact/script/cron state, then recommend the next upgrade layer. Preferred Consumer AI Factory research stack: - **Reddit mining**: raw pain language, acute moments, current workarounds, trust objections. For HappyCouple, use the existing HappyCouple Reddit pipeline rather than creating a new one. - **App Store / Play Store reviews**: competitor complaints, retention failures, pricing objections, feature gaps, “this saved us” language. - **Exa or semantic web search**: broad company/product discovery, obscure launch pages, “companies like X,” creator/specialist discovery, and market maps. - **SERP/keyword endpoint** such as SerpAPI/DataForSEO/Brave/Tavily: search intent, People Also Ask, SEO-shaped demand probes, query clusters. - **X/Twitter search**: current founder/product discourse, launch traction, emerging complaints, creator signals. - **Product Hunt/GitHub/Hacker News**: builder/launch signal and technical competitor discovery. Default upgrade sequence for HappyCouple-style consumer venture research: 1. Verify existing Reddit/App Store/Obsidian artifacts and latest run freshness. 2. Run a small fresh sample if stale. 3. Add semantic/LLM curation on top of keyword classifiers. 4. Convert research into product/GTM artifacts: prompt chips, moments, SEO pages, shortform hooks, telemetry labels, and experiment briefs. 5. Only then add broader endpoints like Exa/SERP/X if they answer a distinct missing question. ## Evaluation inputs ### Consumer AI agent landscape scans When Antoine asks for a landscape of consumer AI agents/products, use the scope and spreadsheet pattern in `references/consumer-agent-landscape-xlsx.md`. The key correction from the AgentAppStore session: do not research "AI agents" horizontally. Focus on true consumer life-task agents — dressing/style, travel, life admin, relationship/dating, family, finance, health, learning, career, shopping, home, and companion/life coach. Keep broad directory/API matches in a separate candidate tab unless manually vetted. ### 1. Painpoint mining Use Reddit, TikTok/Reels/YouTube comments, app reviews, forums, Discords, Quora, X, Amazon reviews, niche communities, and the Consumer AI Agents Landscape / AgentAppStore research surface when available. For AgentAppStore-style market maps, treat competitor/category research as an ongoing factory input, not a one-off spreadsheet. The useful loop is: ```text weekly consumer-agent landscape refresh → category changes / new products / teardowns → opportunity scoring → AgentAppStore SEO/content backlog → venture distribution and build decisions ``` Look for: - repeated situations - exact user language - emotional intensity - current hacks/workarounds - shame/private pain - urgency - trust barriers - categories where products show true agentic depth: personal context, memory, workflow completion, reminders, collaboration, or integration - categories where products are mostly generic ChatGPT wrappers, which should lower confidence unless distribution evidence is unusually strong ### 2. Keyword / search-intent research Use Google autocomplete, People Also Ask, Keyword Planner/Ahrefs/Semrush when available, TikTok/YouTube/App Store search suggestions. Look for: - private intent - high-intent long-tail phrases - solution-aware queries - acquisition potential - language scale Important nuance: SEO/keyword research is useful upstream for demand discovery, but do not over-invest in slow SEO before product demand is validated. Early SEO-shaped pages should be lightweight demand probes and reusable landing assets. ### 3. Existing product / review mining Use App Store reviews, Chrome extension reviews, Product Hunt, Reddit competitor mentions, Trustpilot/G2/Capterra if relevant, YouTube reviews. Look for: - current solutions - complaints - willingness-to-pay clues - retention failures - feature gaps - trust objections ### 4. Specialist/creator distribution research Antoine's factory thesis: prioritize niches where domain specialists/creators have distribution advantages but low AI/technical fluency. Why this matters: - They own audience trust. - They understand the pain deeply. - They may not be able to build AI products themselves. - They can validate and distribute. - The factory can productize their frameworks into personalized workflows. Target niches should be scored for: - specialist audience size - engagement quality - specialist credibility - technical/AI fluency gap - partnership likelihood - brand safety - productization potential Partnership wedge: ```text Expert-led distribution + AI productization + low technical fluency = partnership wedge. ``` ## Factory scorecard Score each opportunity 1–5: | Dimension | Question | |---|---| | Pain frequency | Does this show up repeatedly without prompting? | | Pain intensity | Are users anxious, angry, embarrassed, desperate, or stuck? | | Moment specificity | Can the product start from a concrete moment in under 30 seconds? | | Search intent | Are people privately searching for help? | | Existing workaround | Are they already using friends, ChatGPT, forums, spreadsheets, coaches, templates, etc.? | | AI advantage | Does AI make this materially better than static content/software? | | Trust feasibility | Will users share enough context for AI to help? | | First-value speed | Can the app deliver useful value in under 2 minutes? | | Retention path | Does the first moment naturally become a recurring workflow? | | Distribution fit | Is there a believable organic channel? | | Specialist distribution | Do specialists/creators already aggregate this audience? | | Monetization proximity | Is there eventual willingness to pay? | | Safety/regulatory risk | Can we handle harm cases without becoming unsafe or over-regulated? | | Commodity risk | Can this become more than a generic ChatGPT wrapper? | ## Venture stages ### Stage 0 — Signal scan Goal: decide if the pain deserves attention. Artifacts: - painpoint mine - keyword snapshot - specialist/creator map - competitor/review scan - opportunity memo For a broad opportunity scan, run evidence streams separately before synthesis: 1. Pain/community mining: repeated user language, emotional stakes, current hacks, trust objections. 2. Competitor/App Store/review mining: existing solutions, complaints, willingness-to-pay clues, retention failures. 3. Search intent/social query mining: private high-intent moments, upload/document/decision queries, monetizable urgency. 4. Creator/specialist distribution mapping: audience aggregators, specialist credibility, technical fluency gap, partnership wedge. Then synthesize into a ranked artifact, not raw notes. Standard output for a 10-app scan: - executive recommendation: build/probe/kill posture - ranked table of all ideas - top 3 with why-now, cheapest test, kill criteria - warm/watch list with caveats - cross-run factory learnings: source quality, opportunity shapes, scoring updates, process improvements Decision: kill, keep watching, or define wedge. ### Stage 1 — Wedge definition Goal: choose one acute first-use moment. Artifacts: - target user - exact pain moment - first-use promise - user language examples - trust objections Decision: build probe, revise wedge, or kill. ### Stage 2 — Demand probe Goal: test whether users act. Before advancing beyond Stage 2, require a growth-loop pass with `consumer-growth-model-and-loop-map`. The probe should not only prove that the first output is useful; it should identify the likely loop that can create repeat usage or new users. Default to web for this validation phase. Move to mobile only when the winning loop requires native behavior such as push notifications, voice/camera capture, home-screen presence, contacts/social graph, App Store search, or high-frequency private habit. Artifacts: - landing page/input - interactive first-value prototype - content posts - creator/specialist outreach - manual user tests - fake-door or concierge flow Prototype rule: When the winning opportunity has a concrete first-use moment, build the smallest artifact that makes the promised output tangible enough to test submission/payment behavior. Do not default to a full AI agent. A rule-based or semi-manual prototype is often better if it proves the wedge faster. Good demand-probe prototype shape: ```text messy real user context -> structured useful output -> explicit paid/concierge next step ``` Example from Homeowner Triage / Home Admin OS learning: Weak broad wedge: ```text home emails/docs/bills/repairs -> persistent home admin dashboard with costs, due dates, documents, reminders, suppliers, and next actions ``` This can look useful but is a bad starting wedge: ambient pain, broad ICP, high-friction data entry, delayed value, unclear distribution, and too much overlap with ChatGPT/Notion/Drive/spreadsheets. Sharper homeowner reframe: ```text expensive homeowner decision -> document/photo/quote/report input -> second-opinion artifact -> decision/action before signing or missing a deadline ``` Best examples: - HVAC/roof Contractor Quote Sanity Check: quote + ZIP + project details -> fairness verdict, local price range, scope red flags, questions to ask, negotiation script. - Inspection Report Translator: inspection PDF under contingency deadline -> red/yellow/green risk summary, repair-cost ranges, seller-credit ask list, specialist follow-up checklist. - Water Leak/Emergency Home Triage only later, with conservative safety flows and human/provider escalation. Factory lesson: pain intensity alone is insufficient. Before building or polishing a prototype, require: 1. a clear distribution-owner map, 2. a clear non-ChatGPT advantage based on persistent memory, ingestion, structure, reminders, source trails, expert review, or workflow completion, 3. a named output artifact, not a dashboard, and 4. evidence that the user is at a high-stakes decision point rather than vaguely wanting to get organized. The prototype should test: - Will users paste/upload real context? - Does the output make the value proposition obvious? - Do users click/pay/request the concierge version? - Do they want the promised triage/planning or a different job like contractor matching? Metrics: - visit → real context submitted - sample interaction → own-context submission - CTA click - payment intent / payment conversion when appropriate - reply/comment/save/share - waitlist/signup intent - qualitative output usefulness rating Decision: build MVP, change wedge, or kill. ### Stage 3 — First-value MVP Goal: solve the moment fast. Artifacts: - usable MVP - telemetry - session recordings - qualitative reactions Metrics: - real context submitted - first useful output reached - completion rate - save/share/return intent - user quotes Decision: improve MVP, test retention, change wedge, or kill. ### Stage 4 — Retention / monetization test Goal: see whether this becomes durable. Artifacts: - return loop - account/save mechanism - pricing/payment test when appropriate Metrics: - repeat usage - saved output - invite/share - paid intent - payment conversion if ready Decision: scale channel, revise product loop, or kill. ### Stage 5 — Scale channel Goal: amplify only after activation works. Channels: - SEO - creators - community - paid ads - partnerships Rule: do not scale acquisition before first-value activation is credible. ## Standard venture opportunity brief Use this template: ```markdown # Venture Opportunity Brief - [Idea] ## User ## Pain moment ## Current behavior ## Evidence ### Painpoint examples ### Keyword/search evidence ### Specialist/creator distribution evidence ### Competitor/review evidence ## Emotional language ## First-use promise ## AI advantage ## Organic channels ## Trust/safety risks ## Commodity risk ## Scorecard | Dimension | Score | Notes | |---|---:|---| | Pain frequency | | | | Pain intensity | | | | Moment specificity | | | | Search intent | | | | Existing workaround | | | | AI advantage | | | | Trust feasibility | | | | First-value speed | | | | Retention path | | | | Distribution fit | | | | Specialist distribution | | | | Monetization proximity | | | | Safety/regulatory risk | | | | Commodity risk | | | ## Cheapest next test ## Kill criteria ## Decision ``` ## Factory command-center operating loop When Antoine asks for factory infrastructure, command centers, dashboards, venture OS work, or reusable Consumer AI Factory tooling, use the six-module operating loop rather than inventing ad hoc dashboards: ```text raw signal -> pain sentence -> pain cluster -> wedge candidate -> probe -> telemetry -> learning -> decision -> next action ``` Core modules: 1. Opportunity Pipeline — raw sources, pain observations, pain clusters, scored wedge candidates. 2. Probe Generator — wedge brief, first-value loop, CTA/payment/concierge ask, telemetry plan. 3. Distribution Experiment Ledger — channel, subchannel, tactic, audience, message, offer, CTA, result, signal quality. 4. Telemetry Snapshot — cross-venture events that make probes comparable. 5. Learning Log — scoped reusable doctrine, not vague conclusions. 6. Decision Queue — continue / revise / test distribution / park / kill decisions with evidence and deadlines. Reference detail: `references/factory-command-center-os.md` captures the session-derived command-center model, artifact locations, schema names, and implementation guidance. ## Living command-center OS When Antoine asks to make the factory command center or operating system “living,” do not stop at a dashboard/spec. Promote it into an operating loop: ```text live ledgers -> aggregator/generator -> rendered status -> scheduled pulse -> decision/action -> experiment result -> learning/ledger update ``` Minimum implementation pattern: 1. Create live records, not only schemas, for opportunity candidates, active probes, distribution experiments, telemetry snapshots, learning log, and decision queue. 2. Backfill the current active venture, one serious validation candidate, and parked candidates with explicit gates and “do not build” constraints. 3. Upgrade the status generator so it emits `today_action`, ledger counts, open decisions, stale warnings, recent experiments/learnings, and module status (`live` vs `specified`). 4. Render today’s action, ledger health, and open decisions in the command center UI. 5. Add tests around the generator contract before refactoring: temp ledger fixtures, missing-ledger stale warnings, and action selection from decision queue. 6. Add a daily Factory Pulse cron that regenerates status and reports signal, open decisions, one action, and what not to do. Reference: `references/living-os-command-center.md`. ## Factory asset organization When maintaining Consumer AI Factory artifacts, distinguish reusable factory assets from venture-specific execution assets. Factory assets: - doctrine and principles - cross-venture learnings - playbooks - templates - shared infrastructure notes - agent workflows - build logs and portfolio-level decisions - command-center ledgers and generated operating snapshots Venture assets: - venture brief - product notes - research/pain-mining outputs - GTM queues - outreach drafts - analytics/status notes - sprint logs - decision logs If a lesson learned inside HappyCouple or another venture applies to future ventures, promote it upward into a factory-level artifact or skill. ### Living OS / command-center pattern When Antoine asks to make the Consumer AI Factory OS “living,” do not stop at a spec, dashboard, or UI. Convert doctrine into an operating loop: ```text live ledgers → canonical generator → command center → daily pulse → decision/action → telemetry snapshot → weekly review ``` Minimum file-backed ledgers: - `opportunity_candidates.json` - `active_probes.json` - `distribution_experiments.json` - `telemetry_snapshots.json` - `learning_log.json` - `decision_queue.json` The generated status file is an output, not the durable source of truth. Add tests for ledger aggregation, today-action selection, stale-warning behavior, idempotent telemetry snapshot appending, and pulse/review formatting. Schedule a daily pulse and weekly decision review only after the generator has a verified CLI. Detailed reference: `references/living-factory-os-command-center.md`. Current Obsidian caveat: legacy factory notes may exist at root-level paths, but the canonical target is `/Users/antoinelevy/Documents/Obsidian Vault/Consumer AI Company Factory/` with subfolders for `00 Operating System`, `01 Playbooks`, `04 Opportunity Research`, `05 Ventures`, etc. Before reading/writing, prefer the canonical folder and use `search_files` to verify exact paths if uncertain. Do not perform a blind migration without explicit approval; create/update doctrine first, then migrate deliberately. ## Build log rule Every meaningful venture step needs a running build-log entry. Append to the canonical build log: ```text /Users/antoinelevy/Documents/Obsidian Vault/Consumer AI Company Factory/00 Operating System/Consumer AI Company Factory - Build Log.md ``` Entry template: ```markdown ### YYYY-MM-DD — [Stage] [Action] - Action: - Why: - Artifact: - Signal: - Decision: - Factory learning: ``` ## HappyCouple current lesson HappyCouple is Venture 001 and sits around Stage 2 / Stage 3 boundary: - Pain is directionally validated through Reddit mining and useful-answer engagement. - First-use wedge is relationship moments, but moments should guide internally rather than become rigid UX categories. - Do not keep optimizing the AI conversation from theory before real user feedback. - Current priority is organic distribution, creator/specialist partnership testing, and converting warm Reddit engagement into measured traffic/activation. - Paid ads are deprioritized until payment exists, and payment is not currently the priority. ## Portfolio operating model When Antoine asks about overall factory focus, do not force a one-project-only model. His current intended portfolio shape is: ```text one advanced proof-case venture + one strategic infrastructure/distribution surface + smaller category-transfer probes ``` Current roles: - HappyCouple: advanced proof-case venture for distribution, product learning, and reusable playbooks. - AgentAppStore: market-intelligence, trend-tracking, SEO/content/newsletter, and venture-distribution infrastructure. - Start Small / Hairmatch: younger probes for testing whether the factory can turn vague ideas into fleshed products while building distribution alongside product. Guardrail: multiple projects are acceptable only when each has a distinct learning role, explicit kill/continue criteria, and does not steal primary execution attention from the advanced proof case. Reference: `references/portfolio-operating-model-2026-05-17.md`. ## Anti-internet young-men platform pattern When evaluating young-men consumer AI opportunities around loneliness, porn, AI companions, social media, or algorithmic feeds, consult `references/anti-internet-young-men-platform.md`. The key product rule is: if a surface increases consumption without increasing real-world action, it is part of the problem. Prefer Reality Coach + Rep Card probes that validate pain, challenge distortion/avoidance, and route users into real-world reps. For the Dopamen / Clean Feed clipping-distribution pattern, also consult `references/dopamen-clipping-cta-distribution.md`. The durable lesson: do not turn Dopamen into a generic clip farm. Use positive/agency-building media as inputs for original commentary, permissioned clips, embeds, or licensed content; attach a concrete real-world rep and CTA into the Real World Scoreboard. Avoid straight reposting other creators' content with a commercial Dopamen CTA unless rights/campaign rules explicitly allow it. For Dopamen campaign-reward clipping, consult `references/dopamen-campaign-clipping-rewards-scouting.md`. Treat campaign payouts and Dopamen acquisition as related but separate jobs: record full campaign rules before posting, use Dopamen CTAs only when explicitly allowed, distinguish raw views from eligible/approved/paid views, and reject high-CPM campaigns that would pollute the anti-manosphere / real-world-reps thesis. Start this as a Dopamen Growth Ops module; only split into a separate ClipOps agent when repeated daily operations, campaign-rule memory, payout submissions, and compliance risk justify the extra agent boundary. When Antoine asks to create content for the first/next Dopamen clipping campaign, do **not** treat it as generic short-form ideation. Preserve the three-test frame explicitly in the artifact: (1) revenue from campaign clipping, (2) Dopamen fit/acquisition learning, and (3) reusable shortform content-machine workflow. Produce a concrete campaign content pack with clip briefs, hook overlays, captions/tags, source/timestamp placeholders, Dopamen real-world-rep mappings, compliance/CTA guardrails, first production picks, and a post ledger schema tracking campaign payout metrics separately from Dopamen signal. When Antoine says “clipping” in this context, first check whether he means the Dopamen paid campaign-clipping opportunity, not generic short-form content. For a specific Content Rewards campaign URL, produce a campaign pack and ledger row: extract the public page, pull public Google Doc requirements via `/export?format=txt` when available, capture full rules/source-library/required tags/payouts/disqualifiers, score Dopamen fit, and mark external CTA/link status conservatively. If the agent has browser access to the logged-in platform/session, use it directly to inspect/join/navigate campaign surfaces rather than asking Antoine for screenshots or copied text. Content Rewards “Join Campaign” may route into a Whop community; continue through the joined/community surface and inspect linked campaign/resource pages where possible. Example session reference: `references/dopamen-die-with-zero-campaign-intake-2026-05-22.md`. For Dopamen marketplace/account onboarding, use venture-owned account hygiene before joining campaigns: verify the AgentMail inbox, record OAuth/account gates, and do not substitute Antoine's personal accounts by default. Session reference: `references/dopamen-growth-ops-account-setup-2026-05-22.md`. When Antoine says "content farming" or "clipping" for Dopamen, preserve the three-goal framing: (1) paid campaign revenue, (2) Dopamen fit/acquisition learning, and (3) reusable shortform content machine. Do not collapse the task into generic YouTube clipping or only campaign payout. Create campaign packs plus post ledgers that track payout, platform creative performance, and Dopamen real-world-rep signal separately. Session reference: `references/dopamen-shortform-content-machine-campaign001-2026-05-22.md`. If campaign access is gated by email verification, proceed up to the code boundary, ask only for the code when Gmail is not authenticated, then continue directly into joined rules/source-library/timestamp extraction. ## Verification Before finalizing factory work: 1. Save/update the relevant Obsidian artifact. 2. Append a build-log entry if the step is meaningful. 3. If a durable doctrine changed, update memory compactly. 4. Report exact paths changed.