# Content Growth Strategy Research Notes

Date: 2026-05-24

## Operating model decision

The content machine should be operated and maintained by JARVIS. The human role is approval before posting, plus occasional strategic steering on video volume, content pillars, and brand boundaries.

Practical implication:

- JARVIS owns ideation, queueing, drafting, preview generation, quality gates, analytics import, maintenance, and experiment planning.
- The user approves or rejects exact preview packages before any platform upload.
- No ambiguous consent triggers publishing.
- Daily volume is a configurable policy, not hardcoded into renderers.

## What successful short-form content currently has in common

### 1. Strong hook before context

Sources:

- Marketing Examined, Jenny Hoyos Short-Form Video Playbook: https://www.marketingexamined.com/blog/jenny-hoyos-short-form-video-playbook
- Gillian Perkins, Jenny Hoyos interview: https://www.gillianperkins.com/blog/ep-202
- Loewenstein, The Psychology of Curiosity: https://doi.org/10.1037/0033-2909.116.1.75

Observed pattern:

- Successful Shorts/Reels/TikToks rarely start with background.
- They start with tension, contradiction, visual anomaly, stakes, or a specific unanswered question.
- Jenny Hoyos-style advice emphasizes a hook short enough to land in roughly the first three seconds and strong enough to reshape the whole idea if needed.

Implementation implication:

- Add a hook quality gate before rendering.
- Reject scripts where the first line is generic setup.
- Store `hook_type`, `curiosity_gap`, and `first_2_seconds_promise` as explicit candidate fields.

### 2. Retention is not just speed; it is open-loop management

Sources:

- Loewenstein, The Psychology of Curiosity: https://doi.org/10.1037/0033-2909.116.1.75
- Horstmann, The surprise-attention link: https://doi.org/10.1111/nyas.12679
- Zheng, short-video flow/social influence study: https://www.nature.com/articles/s41598-023-30525-y

Observed pattern:

- Fast cuts help only if the viewer always knows why the next few seconds matter.
- Strong videos create a loop, answer part of it, then open the next loop.
- Pattern interrupts reset attention, but random chaos causes fatigue.

Implementation implication:

- Add a retention beat map to each script:
  - 0-2s: hook / prediction error
  - 2-6s: compressed setup
  - 6-15s: proof, mechanism, or example
  - 15-30s: payoff
  - end: follow reason or next-episode loop
- Motion renderer should insert visible change every 1-2 seconds, but not visual soup. We are not making a nightclub for caffeinated squirrels.

### 3. Emotion drives sharing; usefulness alone is not enough

Sources:

- Berger & Milkman, What Makes Online Content Viral?: https://doi.org/10.1509/jmr.10.0353

Observed pattern:

- High-arousal emotions spread better than neutral information.
- Useful content performs better when it also creates relief, awe, frustration, amusement, identity pride, or contrarian tension.

Implementation implication:

- Add `target_emotion` and `share_reason` to candidate scoring.
- Reject purely informational videos unless they have a strong emotional or identity angle.

Useful emotional lanes for this channel:

- Relief: finally, AI automation explained without grift.
- Awe: this is what a personal AI system can actually do.
- Anger/frustration: most AI productivity advice is theatre.
- Identity pride: serious builders want systems, not toy prompts.
- Amusement: dry, precise commentary on absurd AI hype.

### 4. Identity converts viewers into followers

Sources:

- Tajfel & Turner, Social Identity Theory: https://doi.org/10.4324/9780203505984-16
- Berger & Heath, identity signaling: https://doi.org/10.1086/519142
- Labrecque, parasocial consumer-brand relationships: https://doi.org/10.1016/j.intmar.2013.12.003

Observed pattern:

- A view is usually earned by a hook.
- A follow is earned by a predictable future promise and a distinct worldview.
- Viewers follow when they can say: this account is for people like me, and I know what I will get next.

Implementation implication:

- Each content pillar needs an identity promise, not just topics.
- Review packages should include `viewer_identity` and `follow_reason`.
- Generic tips get penalized even when technically correct.

Potential identity statement:

> For builders who want AI as an operating system for real life, not another pile of prompt hacks.

### 5. Series beat isolated one-offs

Sources:

- AutoShorts.ai public workflow: https://autoshorts.ai/
- Planable approval-process guidance: https://planable.io/blog/social-media-approval-process/
- YouTube Culture & Trends: https://www.youtube.com/trends/report/

Observed pattern:

- Repeatable series create familiarity and reduce the decision cost of following.
- Series also make automated production easier because JARVIS can optimize formats over time.
- The danger is template fatigue if the episode feels copy-pasted.

Implementation implication:

- Build around recurring series templates.
- Add `series_id`, `episode_number`, `format_family`, and `format_variation`.
- Approval applies to exact preview version, not to the entire series.

Recommended initial series families:

1. `AI System Autopsy`
   - Break down one workflow or automation failure and show the fix.
2. `Prompt vs System`
   - Contrast shallow prompt advice with a durable system design.
3. `Workflow Teardown`
   - Break down a public-safe workflow pattern without exposing JARVIS, the user, or private collaboration details.
4. `One Rule, One Example`
   - One sharp principle demonstrated with one concrete use case.
5. `AI Hype Court`
   - Put a viral AI claim on trial: claim, evidence, verdict.

### 6. Source-backed content has more trust than pure synthetic ideation

Sources:

- OpusClip Audacy case: https://www.opus.pro/how-audacy-drove-1b-views-using-opusclip
- OpusClip customer stories: https://www.opus.pro/customer-stories

Observed pattern:

- Clip systems scale well when the source material already contains strong moments.
- Synthetic/faceless channels can grow, but generic AI-only content has a high slop risk.

Implementation implication:

- Add a future source-backed mode:
  - notes / transcripts / meeting summaries / build logs / research sources
  - extract candidate moments
  - score hooks
  - generate approval previews
- Prefer real process, real constraints, and real examples over invented generic scripts.

### 7. Platform incentives differ, but quality gates should be platform-independent

Sources:

- YouTube Shorts help: https://support.google.com/youtube/answer/15424877?hl=en
- YouTube monetization policies: https://support.google.com/youtube/answer/1311392?hl=en
- TikTok What's Next 2025 trend report: https://ads.tiktok.com/business/en-US/blog/whats-next-2025-trend-report
- Meta AI labeling approach: https://about.fb.com/news/2024/04/metas-approach-to-labeling-ai-generated-content-and-manipulated-media/

Observed pattern:

- YouTube Shorts, TikTok, and Reels all reward watch behavior, but eligibility, monetization, disclosure, and community norms differ.
- Mass-produced inauthentic content is a monetization/account-health risk.
- AI-generated realistic content may need disclosure.

Implementation implication:

- Keep platform adapter logic separate from strategy and rendering.
- Add disclosure metadata before upload adapters exist.
- Add originality and inauthentic-content checks before final render.

## Psychological model for the content machine

Each candidate should score against these viewer states:

1. Stop
   - Trigger: novelty, contradiction, visual anomaly, strong first sentence.
   - Gate: first 2 seconds must create prediction error.

2. Stay
   - Trigger: specific curiosity gap and obvious next payoff.
   - Gate: every 3-5 seconds must justify continued attention.

3. Feel
   - Trigger: relief, awe, amusement, frustration, identity pride.
   - Gate: target emotion stated before rendering.

4. Trust
   - Trigger: specificity, proof, tradeoffs, source-backed examples, transparent reasoning.
   - Gate: video must contain at least one trust signal.

5. Follow
   - Trigger: future value promise and creator/worldview distinctiveness.
   - Gate: viewer can explain why following produces more value.

## Role of variety

Variety is not random topic hopping. Variety should exist inside a stable promise.

Recommended ratio:

- 70% proven repeatable formats
- 20% controlled variations on proven formats
- 10% experiments

Stable elements:

- worldview: AI as servant/system, not hype deity
- tone: precise, useful, dryly skeptical
- visual grammar: mechanism-reveal, system diagrams, concrete artifacts
- promise: better thinking and better systems through AI
- approval/quality standard: no fake claims, no platform spam

Variable elements:

- hook type: contradiction, myth, visual anomaly, direct challenge, story start
- payoff type: tool, workflow, rule, teardown, verdict, checklist
- emotional tone: awe, relief, irritation, amusement, caution
- format family: build log, autopsy, court, rule/example, before/after
- visual layer: diagram, UI mock, process timeline, checklist, cinematic b-roll later

Implementation implication:

- Add `format_family` and `variation_axis` to candidate metadata.
- Track fatigue per format family.
- Avoid publishing too many videos with the same hook shape on consecutive days.

## What makes this channel unique

The dangerous default is becoming another AI tips account. That is undignified and, worse, boring.

Differentiators to protect:

1. Internally JARVIS-operated, publicly brand-safe
   - The channel is produced by a governed automation system, but public content must not be about JARVIS, the user, or their collaboration.
   - The visible moat is quality, structure, and safe systems thinking, not private behind-the-scenes material.

2. Systems over prompts
   - Most AI content sells prompt tricks.
   - Our angle: workflows, operating systems, quality gates, memory, approvals, automation boundaries.

3. Transparent constraints
   - Show tradeoffs, failures, and why something is not automated yet.
   - This builds trust and avoids AI-grift aesthetics.

4. Mechanism-first visuals
   - Renderer-owned captions, arrows, diagrams, timelines, and evidence.
   - Avoid generic AI stock footage and baked-in gibberish text.

5. British-but-German strategic voice
   - Internal approval in German, final audience content in English.
   - Tone can preserve dry precision: useful, slightly sardonic, never clownish.

6. Real artifacts
   - Screens, workflows, code snippets, dashboards, checklists, before/after systems.
   - Real source-backed content beats invented faceless filler.

## Proposed strategic gates to implement

Add these to the content machine before scaling daily volume:

### Hook gate

Fields:

- `hook_type`
- `first_2_seconds`
- `curiosity_gap`
- `prediction_error`

Reject if:

- first line starts with generic context
- no specific unanswered question
- no novelty/contradiction/tension

### Follow reason gate

Fields:

- `viewer_identity`
- `future_value_promise`
- `series_fit`
- `follow_reason`

Reject if:

- video could belong to any generic AI account
- no reason to expect future value

### Originality / anti-slop gate

Fields:

- `specificity_score`
- `source_backing`
- `generic_ai_slop_risk`
- `template_repetition_risk`

Reject if:

- generic tips without example
- fake authority
- repeated template too similar to recent videos

### Emotional payoff gate

Fields:

- `target_emotion`
- `emotional_intensity`
- `share_reason`

Reject if:

- merely informative but emotionally flat

### Variety planner

Fields:

- `format_family`
- `variation_axis`
- `recent_similarity_score`
- `experiment_bucket`

Reject or defer if:

- too similar to recent approved content
- same hook/payoff pattern repeated excessively

### Analytics feedback loop

Fields to collect after posting later:

- views
- average view duration
- retention curve checkpoints
- completion rate
- rewatches
- likes
- comments
- shares
- saves
- follows per view
- negative feedback if available
- platform and post time
- series_id / format_family / hook_type

JARVIS should use this to recommend next batch composition and daily volume.

## Suggested first implementation sequence

1. `ContentStrategyScore`
   - hook, curiosity, identity, emotion, trust, follow reason, originality, variety.

2. `ContentSeries` registry
   - series_id, promise, audience, allowed formats, tone, no-go patterns.

3. `CandidateStrategyReview`
   - deterministic review object attached before preview generation.

4. `RecentContentMemory` / local analytics cache
   - read-only at first; tracks recent format/hook usage to prevent repetition.

5. `DailyVolumePolicy`
   - configurable later, e.g. test mode 1/day, growth mode 2-4/day, experiment mode by platform.

6. `AnalyticsImport` later
   - after platform adapters exist and only with explicit credentials/config.

## Current recommendation

Do not increase volume yet. Build the strategy gates first, then run small batches.

Initial operating mode:

- 1-3 preview candidates per day generated by JARVIS.
- User approves only exact preview packages.
- Post volume starts lower than generation volume.
- JARVIS learns from approval decisions and later platform analytics.

Winning direction:

> An internally JARVIS-operated, approval-governed AI systems channel that turns public-safe workflows, failures, and decisions into mechanism-reveal Shorts without exposing JARVIS, the user, or private collaboration details. Repeatable enough to be recognizable; varied enough to avoid template fatigue; specific enough not to smell like AI slop.