Content Pipeline Workflow

Daily read of Kevin's LinkedIn and X activity, engagement analysis, gap detection, and actionable post recommendations.

The content-pipeline automation feeds Writing and Content Skills and the social backlog in wiki/career/content-backlog.md. Source: automations/content-pipeline.md, 2026-05-31

Inputs

• LinkedIn: recent posts, comments, connection growth
• X: tweets, threads, replies, and raw/x-bookmarks/ for saved signal
• Targets from Social Content OS content mix (5 technical : 3 product : 2 human per 10 posts)

Analysis steps

1. Performance review : top/worst posts, engagement trend, best posting time, audience growth (7-day window)
2. Gap analysis : compare output vs Writing and Content Skills arcs and backlog rows not yet Posted
3. Context gate : verify brand, product, ICP, and offer context files are current before campaign-style generation
4. Freshness gate : refresh socials, recent blogs, YouTube, and /last30days inputs when the source set is not fresh this month
5. Recommendations : 3-5 draft ideas with platform, format, and hook; flag Ready to post rows for Writing and Content Skills + Writing and Content Skills pass

Market and campaign research lane

Use this lane before campaign strategy when the question is market structure, positioning, customer pain, or the validity of a proposed GTM assumption. It is research, not lead-generation or send authority.

1. Freeze a corpus ledger with source kind, company/product/entity, date, freshness, URL or artifact hash, selection reason, and rights/access basis. A strong pass deliberately mixes competitor sites, incumbent earnings calls, customer reviews, community complaints, primary product documentation, and relevant performance evidence instead of treating thirty similar pages as independent confirmation.
2. Extract cited latent consensus, market assumptions and their falsifiers, strongest adversarial or investor case, contradictions, missing evidence, and the customer questions that would change the decision. Keep source disagreement visible; model synthesis is not a market fact.
3. Validate the resulting hypotheses with real, consented customer or operator conversations. Record sampling limits, disconfirming answers, and what changed in the product, positioning, workflow, or stop decision.
4. If a bounded list or enrichment job is still needed, Origami may be evaluated as an optional provider after current pricing, terms, privacy, source accuracy, credits, connector custody, suppression, opt-out, export, and deletion checks. Default to export into Kevin-owned systems and no-send; the customer remains responsible for legal basis, consent, platform compliance, and outreach behavior.

For multi-specialist content work, Vercel's Eve template is a useful orchestration pattern rather than an automatic stack install: one lead routes and reviews but does not author every deliverable; fresh specialists receive self-contained briefs; one shared brand-context document has a single owner; each specialist gets only its needed connectors; irreversible sends, deletes, moves, or publishes require exact approval. Adopt the hosted dependencies only when their operational and integration cost is justified. Source: X 2084284424233828532; Origami official pricing/terms/privacy reviewed 2026-08-12; X 2086939268136755285; vercel-labs/marketing-team-eve-template@7ccaa5f4

Illustration branch

When an approved article, script, or post needs illustrations, keep the images inside the same content workflow instead of creating a detached art task:

1. Extract the narrative beats, insertion points, intended reader takeaway, aspect ratio, count, title behavior, and accessibility text.
2. Write a shot-config-style strategy before generation: each image names its source passage, visual promise, composition, reading order, density, style, and whether a character/reference is necessary.
3. Use the existing approved image/illustration capability. GiMi's MIT skill at commit 0f602eaf is retained as a pattern source for strategy-first article illustration, three calibrated style modes, custom-character enrollment, reference checks, QA, and delivering the strategy beside the images.
4. Do not use or imply ownership of the Gimi horse-hat character: its IP is explicitly outside the repository's MIT license. Use IP=none or a Kevin-owned, consented character/style reference unless separately licensed.
5. Review every image in context for claim fidelity, source/character/style rights, text accuracy, visual coherence, repetition, crop, placement, contrast, alt text, provider/output terms, and complete-set quality. The image provider remains a separate cost/privacy/rights boundary.

This imports a useful procedure without installing a second overlapping illustration authority or copying protected character assets.

Media production branch

Open this branch only after a content recommendation or approved source becomes a concrete video brief. Choose the engine by artifact shape:

• use HyperFrames first for a bespoke launch, product/demo walkthrough, exact brand motion, code/data animation, or any frame-reproducible authored scene;
• use MoneyPrinterTurbo for high-throughput topic/script-to-stock-footage narrated shorts whose value comes from assembly, voice, captions, and variant production rather than unique scene design;
• use Remotion when the product already owns a React composition system, and a stochastic media provider only when generated footage is the actual brief.

For ChatGPT Images 2.0, Nano Banana, Higgsfield, or another generative-media candidate, preserve the exact model/provider/mode and input assets, then test prompt adherence, text and factual accuracy, edit locality, layer/timeline or other downstream editability, identity/likeness and trademark handling, provenance metadata, output rights/terms, cost, retries, and representative failure cases. A launch reel or one-shot ad proves candidate range, not that every static ad is reproducible or production-ready. Adobe integration improves the edit handoff but does not clear source rights or publication authority. Source: OpenAI ChatGPT Images 2.0 release and system card; Higgsfield Adobe plugin launch; saved Nano Banana/ChatGPT Images/Higgsfield signals, reviewed 2026-08-12

For the stock-footage assembly route:

1. Freeze the brief — owner, audience, platform, approved script or claim sources, aspect/duration, language, voice, disclosure, variant hypothesis, and success measure.
2. Freeze the runtime — exact source commit/tree, lockfile, Python/runtime, provider/model/voice IDs, local bind and access policy, queue bounds, and publishing disabled. Do not use an unpinned agent installer.
3. Clear inputs — record each stock/local asset, creator/provider/asset ID, license/terms snapshot, attribution, releases/trademarks, music, font, voice, and model/output terms. “Free stock” is not “copyright-free”; do not use MoneyPrinterTurbo's bundled YouTube-sourced songs.
4. Render local drafts — use a stable task ID and preserve the input brief, approved script, command/config-field manifest, asset ledger, logs, provider calls/cost, output hashes, and failures. Generated claims must return to primary evidence before review.
5. Inspect the complete artifact — frames and safe areas, pacing, footage relevance, captions, audio, identity, trademarks, claims, synthetic-media disclosure, accessibility, and actual platform playback. Variants change one declared hypothesis rather than multiplying near-duplicates.
6. Approve, then publish separately — Kevin or the named owner approves the exact file, caption, disclosure, accounts, platforms, visibility, and schedule. Generation credentials or a configured Upload-Post account are not publication authority. Retain platform IDs, response receipts, errors, and a takedown owner.

Minimum production receipt: brief/source IDs; engine and frozen revision; script and claim evidence; provider/model/voice; asset/license/attribution ledger; task/config manifest; render command/log/output hashes; reviewer and QA result; approval; exact account/platform/visibility/disclosure; publish receipt; measurement window, outcome, incident/takedown state, and writeback owner.

Campaign branch: clipping and derivative distribution

Open this branch only when the request names a source recording, launch, founder-content campaign, clip series, or coordinated short-form distribution. Ordinary post repurposing stays in the lighter social-content path.

1. Freeze the source — preserve recording/transcript, consent, rights, confidentiality classification, speaker/project IDs, source claims, and proof artifacts.
2. Curate before editing — a human or named reviewer watches/reads the whole source and records candidate segments with timestamps, standalone hook, complete payoff, target audience, and rejection reason.
3. Declare variants — every derivative names the source segment and one primary hypothesis: hook, caption, crop, cut, duration, platform, or account. Do not create minimally changed reposts merely to increase volume.
4. Produce and inspect — use HyperFrames or the appropriate media route, then review context fidelity, identity, audio, captions, safe areas, accessibility, claims, and platform-native rendering.
5. Approve and disclose — record who may publish, who may take down, brand rules, paid/incentivized relationship disclosures, music/media rights, and platform policy. No approval means local drafts only.
6. Publish through the service boundary — require explicit authorization for each account or approved batch. Never treat this workflow as standing auto-post authority.
7. Observe for at least the declared window — join variant IDs to platform metrics, spend, watch/retention, saves/shares, qualified visits, direct/search lift, and incidents. The article suggests a 48-hour early read; the campaign defines the real observation window.
8. Graduate, revise, or stop — scale only repeatable winners that also pass brand, policy, and business-value checks. Preserve losing variants so the same hypothesis is not unknowingly rerun.

Minimum receipt: source ID; segment timestamp; variant ID and hypothesis; editor/reviewer; rights and disclosure state; account/platform; approval; publish timestamp; spend; observation window; metrics; incident/takedown state; decision and writeback owner. Source: Clipping as Distribution Strategy, reviewed 2026-08-10

Growth research and voice-model boundary

High-signal growth playbooks, emerging-account monitoring, and personal-voice analysis are retained as research inputs, but none creates standing outreach or imitation authority.

• Preserve the complete playbook or post, named examples, platform/date, and engagement as evidence; convert each tactic into a measurable hypothesis with audience, channel, cost, consent/policy, attribution, and stop rule.
• Monitor public posts within platform rules and rate limits. Do not copy a creator's identity, protected expression, media, private data, or distinctive voice. Use relative performance as a discovery signal, then create original Kevin-aligned work with source/rights/disclosure review.
• Building Kevin's own voice guide from messages requires account and audience separation, participant consent/expectation review, minimization, local raw custody, correction/deletion, and a clear ban on exposing private contacts or sending as Kevin. A style guide is a drafting aid, not identity delegation.
• Auto-GTM products remain local-draft pilots until product terms, data sources, lawful outreach basis, suppression/opt-out, credential custody, sender/domain reputation, anti-spam controls, pricing, booking semantics, and qualified outcome evidence are verified. No-send is the default.

Agent-operated growth-loop admission

The saved 23-agent growth catalog is a useful problem inventory, not permission to install 23 autonomous actors. Admit each idea by the authority and reversibility of its action:

• Read, analyze, alert, or draft: classify cancellation reasons and draft a routing recommendation; draft closed-lost follow-up; prepare changelog copy; monitor competitor documentation and changelogs; measure how often named buyer questions receive cited answers across multiple models; analyze activation sequences; or alert a human when product usage grows materially. These are the default admissible lanes when source access and audience boundaries are declared.
• Experiment only after exact approval: live page or campaign variants, customer-message sequencing, and other externally visible tests require a named owner, frozen baseline and hypothesis, target/account/audience, bounded spend and rate, exact proposal digest, success and harm measures, stop rule, rollback, and dated result receipt. Approval expires when any material input changes.
• Never autonomous: do not let an agent kill production variants, buy ads opportunistically during a competitor outage, request reviews or referrals without context, enrich and contact pricing-page abandoners, scrape a competitor's likers or emails, move customers into harder-CTA sequences, or make high-impact lead decisions from an opaque score. A useful underlying signal may be retained as research while the proposed mutation is refused.

Every admitted loop records its data authority, objective, baseline, counterfactual or comparison, metric and guardrail, idempotency key, spend/rate limit, approval class, suppression and opt-out behavior, rollback and incident owner, and canonical writeback target. “Agent” is an execution role inside this contract, not a waiver of product, privacy, platform, or human-review rules. Source: X 2086534549341610457, complete 23-item thread, reviewed 2026-08-12

Campaign mutation and audience-research contract

Marketing agents, ad connectors, creative skills, synthetic-customer methods, and AI-generated campaign variants are useful only when the experiment and the external mutation are separately inspectable:

1. Freeze the source data, site/property, brand/identity, audience, channel, account, campaign/ad IDs, baseline, objective, budget, dates, and claims.
2. Record the exact skill/repository/model/provider revision and every input, asset, license, consent, disclosure, and platform-policy boundary. A tool's login or MCP connection is access, not permission to change spend or publish.
3. Generate local drafts or a provider preview. Each variant names one hypothesis and preserves its copy/creative/config digest, predicted outcome, cost, and failure state.
4. Require a target-bound approval for the exact audience, account, spend cap, schedule, assets, claims, disclosures, and proposal digest. Any material edit invalidates the approval.
5. Publish through an idempotent service boundary; retain provider/platform IDs, responses, cost, rollback/takedown state, and the human actor.
6. Join the result to dated impressions, clicks, qualified conversions, incrementality or a declared counterfactual, complaints, policy incidents, and the next decision. Headline speed, output, or conversion claims from a referring post remain hypotheses until the underlying case is reproducible.

Synthetic audience research is a hypothesis generator, never a substitute for consented human evidence. Preserve the prompting/persona construction, source population, model and embedding versions, scoring procedure, comparison set, segment errors, calibration, and uncertainty; validate material product, pricing, identity, or health decisions against a representative human baseline. Likewise, ad libraries, Gooseworks, Higgsfield, MeTube, marketing-skill collections, and similar sources remain useful candidates while licenses, source-media rights, identity, credential custody, output ownership, cost, complete-artifact quality, and publication authority are evaluated per job. Source: X 2032863625128587534, 2042994619982471371, 2058247099544948837, 2050697067882746209, 2065118982659883350, 2069942953255252089, 2051792721871004002, 2056411641592287435, 2076797371661643927, 2032675970889691655, 2056427804531773598, 2079381852390117853, 2076637902050844736, 2076841666921869794, 2076841678909124622; current repository receipts, reviewed 2026-08-12

Viral-launch and solo-business case-study receipt

A high-engagement launch thread or revenue screenshot is useful discovery evidence, not proof of a repeatable growth system. Preserve the complete case, then require the exact product, audience, channel, account age, creative and claim revisions, launch window, impressions, clicks, qualified conversions, attribution method, spend, refunds, retention, support load, payroll or contractor costs, founder time, platform-policy incidents, and failed variants. Separate observed facts from the author's explanation and from the hypothesis Kevin will test. For an agency or one-person-company claim, report gross versus net revenue, concentration, delivery workload, churn, rights, and sustainability before using it for planning. The output is a bounded campaign experiment with a stop rule—not an instruction to copy the post or business. Source: X 2044502868556992937, 2054989194414465222, 2056580180211093571, 2070652345562923299, 2067580119930028041, 2010042119121957316, 2076749775043690715, 2078525474410569955, reviewed 2026-08-12

Investor data-room packet lane

When the deliverable is an investor packet rather than a public campaign, freeze a dated, access-controlled evidence room and derive the narrative from it. The saved seed-round post contributes a useful five-artifact checklist: customer/contracts/revenue cube, one-page vision memo, designed deck, operating model, and complete pipeline including won, lost, and open stages. Preserve the source schema, as-of dates, formulas, units, owners, contracts, assumptions, and access log; link memo/deck claims back to exact rows instead of copying numbers by hand. Source: X bookmark, reviewed 2026-08-12

The anecdotal “five days” and valuation claim are not a fundraising benchmark. Before sharing, reconcile customer consent/confidentiality, contract and revenue definitions, realized versus projected revenue, pipeline stage and probability, margins and hiring assumptions, cap table/legal review, PII redaction, viewer permissions, download/watermark policy, and an expiry/revoke path. The agent may assemble, calculate, lint, and generate a claim-to-evidence matrix; a named founder/finance/legal reviewer approves the exact packet and recipient access.

Human gate

Agents draft; Kevin publishes. Never auto-post. A delegated clipper or agency does not inherit authority to make unsupported claims, hide a material connection, use unlicensed media, or multiply an unapproved clip. Update Writing and Content Skills, Writing and Content Skills, the content backlog, and the campaign receipt after a decision.

---

Timeline

• 2026-08-12 | Added the evidence-first market and campaign research lane: diverse dated corpora, latent-consensus/assumption/falsifier analysis, adversarial review, and real customer validation precede GTM action. Retained Origami only as an optional no-send list/enrichment provider under current terms and data checks, and Eve as a bounded orchestration pattern with a non-authoring lead, fresh specialists, one owned brand context, least-capability connectors, and exact approval for irreversible actions. Source: X 2084284424233828532, 2086939268136755285; Origami official sources; vercel-labs/marketing-team-eve-template@7ccaa5f4

• 2026-08-12 | Converted the saved 23-agent growth catalog into an authority-based admission ladder: read/alert/draft lanes are separable from exact-approved experiments, while opportunistic ads, scraping-based outreach, unreviewed asks, automatic production kills, and opaque high-impact scoring are explicitly non-autonomous. Source: X 2086534549341610457, complete thread

• 2026-08-12 | Added an investor data-room lane: five useful artifact families, dated claim-to-row lineage, access control, confidentiality/PII handling, metric definitions, assumption review, and human approval; the author's speed/valuation result remains an anecdote rather than a benchmark. Source: X bookmark

• 2026-08-12 | Added a viral-launch and solo-business case-study receipt so high-signal posts retain their ideas while campaign attribution, unit economics, workload, rights, failures, and repeatability stay explicit. Source: final frontier/career/design cohort

• 2026-08-12 | Added a campaign-mutation receipt and separated synthetic audience hypothesis generation from consented human validation. Marketing, media, and ad tools remain retained candidates, but access never grants spend, send, or publish authority. Source: content/HITL deep-adoption cohort

• 2026-08-12 | Added the generative-media candidate gate: exact model and inputs, prompt/text/factual accuracy, local editability, downstream artifact, identity/rights/provenance, cost, retries, and failures. Launch reels and one-shots remain range evidence, not production acceptance. Source: OpenAI ChatGPT Images 2.0 primary materials; Higgsfield Adobe launch; three saved X signals

• 2026-08-12 | Added the growth-research and voice-model boundary: retain useful playbooks and public performance signals while requiring originality, platform/rights proof, participant/audience separation, private-message minimization, and no-send defaults for auto-GTM. Source: compiled from X bookmarks 2066589201085370482, 2076701436029915208, 2079565340472701152, 2078016376535445940

• 2026-08-12 | Integrated GiMi's strongest transferable signal into the content workflow: strategy and insertion points before generation, explicit character enrollment/calibration, paired shot-config delivery, and complete-set QA. Kept the MIT code/procedure distinct from the excluded Gimi character IP and the external image provider. Source: X 2078632913076195649; GiMi-Xiaomi/gimi-illustration-skill@0f602eaf

• 2026-08-11 | Added a job-shaped media production branch: HyperFrames remains the designed deterministic-video default, while MoneyPrinterTurbo is recommended for high-throughput stock-footage narrated shorts only after frozen runtime, rights, claims, local draft, complete inspection, and explicit publication gates. Source: X/@Atenov_D; current harry0703/MoneyPrinterTurbo replay

• 2026-08-10 | Added the clipping campaign branch: freeze one canonical source, curate story-complete segments, vary one hypothesis at a time, inspect, approve/disclose, publish only through explicit account authority, measure business outcomes, and graduate only replayable policy-safe winners. Source: X/@subahwadhwani X Article; Clipping as Distribution Strategy

• 2026-07-04 | Added the agentic marketing gate from the reviewed EXM7777 artifact: context files and fresh research must be current before content recommendations become drafts. Source: X/@EXM7777, 2026-06-28

• 2026-05-31 | Workflow page created from content-pipeline automation. Source: automations/content-pipeline.md, 2026-05-31