Recoup Internal Marketing

Recoup Internal — Marketing (the daily run)

The loop for a day of marketing on one of our own accounts. This skill is the orchestrator: it owns deciding what to make and proving it is grounded in evidence. It delegates production to the video-format skills and publishing to recoup-internal-social-ship-posts.

The goal — read this before anything else

Convert a free content viewer into a paying Recoup subscriber — and know which post did it.

That sentence is the job. Everything below serves it, and any step that stops serving it should be changed rather than performed.

What follows from it, in order of how often it gets forgotten:

1. Likes, views and followers are leading indicators, not the result. A run that produced a well-liked post and zero attributable signups did not succeed — it just felt like it did.
2. A post is one half of a conversion; the destination is the other. A brilliant asset pointed at a page that cannot convert is a wasted asset and a burned first-time visitor. You own the whole path from the first frame to the subscription, not just the frame.
3. If the result cannot be measured, say so out loud — every run, in writing. An unmeasured conversion is not a zero, and reporting engagement in its place is how a broken funnel stays invisible for weeks.
4. You are allowed to not post. If the evidence says the bottleneck is downstream — a dead destination, a broken signup path — the highest-value marketing act that day is to say so and get it fixed. Shipping a post into a funnel that converts nobody is activity, not progress.

Why this is written so bluntly: on 2026-07-30 a full day went into craft for a four-platform slate while the page it pointed at had produced zero trials in eight weeks, and nothing in this skill would have surfaced that. Steps 2a and 5 now do. Full account: references/conversion.md.

Do the steps in order. Each one exists because skipping it has cost us something specific, noted inline. Step 2 is the one most often skipped and the most expensive to skip.

Step 0 — Identify the account

Read ACCOUNT.md in the account workspace for the artist_account_id and the connected socials. Connections live on the artist account, not the token's default identity.

Our accounts are distinct and must not be mixed: the personal sweetman account and the Recoup official accounts each have their own IDs. Confirm which one today's post is for before writing a word.

Step 1 — Read the workspace (never skim)

The account workspace holds the durable marketing knowledge. Read in this order:

Also read references/video-pipeline.md (the shared build recipe) and references/voice.md (voice choice, audio tags, loudness) before generating audio.

Output of this step: name the arc and the character today's post serves. A post that serves no arc is an ad; reframe it or drop it. If the workspace has no NARRATIVE.md, say so to the account owner rather than inventing a story.

Step 2a — Read the funnel BEFORE the feed

Open every run with one number: trials/subscriptions started since the last run. It is the only number that says whether the marketing is working. Pull it from recoup-internal-sales / recoup-internal-funnel-valuation-pipeline (Privy signups, Stripe trials and cards, credits, Attio stage) — not from the social scrape, which cannot see it.

Then state, in one line each:

• trials/subscriptions since the last run — the result;
• whether the destination for today's likely CTA converts at all — see the gate in step 5;
• whether conversion is currently readable — visit capture IS live (verified 2026-08-18): Vercel Web Analytics records utm_* on both properties automatically. Run the attributed-visits pull in references/conversion.md → The attributed-visits pull for the per-campaign numbers. What stays unreadable is the visit→signup join (chat#1889 row 29) — declare that gap in writing rather than discovering it at the re-pull.

If trials have been flat at zero across several runs, that is the finding, and it outranks today's post. Say so, and see the funnel guardrail below.

Step 2 — Scrape the account's own socials

Do not skip this and do not substitute the posts-log for it. The log is what we thought happened at the last re-pull; the scrape is what is true now. On 2026-07-28 a full announcement video was built off the log alone, and the owner then reported that recent hooks had been weak — which a fresh scrape would have shown before the build, not after.

# One {runId, datasetId} per linked profile. Costs 5 credits + 1 per requested post, PER PROFILE.
curl -sS -X POST "https://api.recoupable.dev/api/artist/socials/scrape" \
  -H "Authorization: Bearer $RECOUP_ACCESS_TOKEN" -H "Content-Type: application/json" \
  -d '{"artist_account_id":"<ARTIST_ACCOUNT_ID>","posts":20}'

# Poll each run until SUCCEEDED (~1-3 min), then read `data`
curl -sS "https://api.recoupable.dev/api/apify/runs/<RUN_ID>" \
  -H "Authorization: Bearer $RECOUP_ACCESS_TOKEN"

Then rank recent posts by engagement and name the differences between the top performer and the flatliners: hook type, format, whether a collaborator was tagged, whether AI was disclosed, length, posting slot. Those differences are today's brief. Details and per-platform metric shapes: references/learn-from-socials.md.

Report deltas, not all-time totals. Also flag any platform whose numbers have structurally collapsed (e.g. Shorts views dropping an order of magnitude) — that is a distribution problem no amount of asset quality will fix, and it outranks today's post.

Step 3 — Pick today's topic

Three legitimate sources, in preference order:

1. Continue a live arc. Best default. Serialized beats compound and the canon asks recurring segments to end on a cliffhanger. Check NARRATIVE.md for what is mid-arc and what was teased.
2. Pull a logged content idea. Ideas parked in the workspace or in NARRATIVE.md's upcoming canon dates (anniversaries, launches, scheduled milestones).
3. Draft a new idea that advances an arc and is worth adding to the canon. If it needs a new storyline, write the arc proposal as its own doc and get an owner ruling before treating it as canon.

Gate zero, before every other gate: name the KIND of piece — launch, incident, or report — from the source, not from what scored last week. Our strongest register is confession, so a shipped feature gets pulled toward being a story about our own mistake. That cost a full rebuild on 2026-08-12. Structures and the beat-budget test: references/topic-selection.md → Gate zero.

Gate the pick on all of these before building — details in references/topic-selection.md:

• Why named — theirs first, then ours. One sentence on what the person in the piece gets out of it, one on what we get. Both at the top of the plan doc. If you cannot write the first sentence, do not build the piece — it is an ad wearing someone's name. Where it publishes an artist's real numbers, state the risk to them in the same breath as the case, then let them decide. Ours is allowed to be commercial; say so plainly rather than dressing it as a favour. Arguments that hold up, and the incident that added this gate: references/topic-selection.md.
• Arc + character named.
• Destination named. Say where a convinced viewer goes before you build, and make sure it continues this post's promise. If the honest answer is "the homepage," the idea is not yet a conversion post — either find the page that pays off its promise, or accept and state that this is a trust/awareness beat, not a conversion one. Both are legitimate; conflating them is not.
• Collaborator named, or explicitly none. Featuring the artist as a collaborator and inviting them to co-post is the largest measured lever we have: the 2026-07-22 collab reel took 45 likes against 1–3 for every non-collab reel in the weeks after. Ask every run: who is the collaborator, and have they been invited? "None" is an allowed answer that costs reach.
• Consent. Internal roster artists need no per-piece permission. External artists need explicit likeness consent, and their real numbers are a separate gate. Check cast/<character>/.
• Numbers. Any figure attributed to a real artist must be measured and verified. Never publish a number you cannot audit, including numbers about ourselves. Measured-vs-estimated disclosure is one later beat in the piece, never part of the opening figure — see references/hooks.md → One flat figure.
• NO AI disclosure in body copy. Owner ruling, 2026-07-28: never spend caption, tweet, description or on-screen text saying the visuals are AI generated. It is already obvious to the audience, and the characters are worth more spent on substance. Where platform-policy labelling is wanted, use the platform's own AI-content toggle at upload. This reverses the earlier 2026-07-06 A/B guidance; if an older workspace doc still says "always disclose", this ruling wins.
• One idea per day. Do not ship two competing assets into the same slot.

Step 4 — Build the asset

Gate: LOCK THE CONCEPT BEFORE A SINGLE IMAGE. Write the premise, the character and the structure in one paragraph, and get a ruling on it. Then do not change them once generation starts. On 2026-09-02 the concept moved three times mid-build (one style, then six styles, then three; then "other worlds" versus "the same room re-imagined"), and every change orphaned the work under it, including a complete 21-still pass. The owner's verdict on that film was "$2.85 for garbage", and the cost was never the money, it was that no shot could be judged while the target kept moving. If the vision is imprecise, the film will be imprecise. That is not fixable downstream.

Then build in this order, and it is the same order for EVERY video, not just artist music videos:

character sheet → location plates → scenes → motion → composite

Each stage is approved before the next begins, because each one is the seed for the next and a mistake compounds. Full mechanics live in references/video-pipeline.md → Reference discipline. This order used to be documented only in the artist-music-video reference (now its own skill, recoup-music-video), which a run for our own channels never opens; that is why it was skipped. It is a general rule, not a music-video rule.

1. Clear the HOOKS.md checklist before writing the scene table. Hook first, not last. Name which archetype the hook is (specific number, contrarian claim, list tease); if you cannot name one, rewrite it. Never open by qualifying the audience. Cold-feed premise test: the video itself must state its premise in its first beats (VO or on-screen text) — feeds autoplay, so a cut that relies on the caption for its story has no story for most viewers (references/hooks.md).
2. Pick a style from VIDEO-STYLES.md and clone its reference project's video/ dir. Do not start from scratch. Route to the matching format skill where one exists (see references/video-formats.md); a format with no skill yet is built from its reference project.
3. Choose the voice deliberately, and check the MODEL before the voice — references/voice.md. Most "the VO sounds robotic" problems are a stale model_id inherited from a cloned generator script, not the voice. A real listener called our narration "mechanical, not human" on 2026-08-06 and was right: we were on a stock voice driven by a model two generations old. Run a bake-off on the film's actual lines, normalise every candidate to the same loudness before listening, and record the winner (voice id, model, settings, date). A recurring character keeps its canonical voice in cast/<character>/ and prefers a first-person inner voice over a narrator; a narrator may change whenever the old one is the problem. Verify every audio tag by transcribing the output back — an unrecognised tag gets spoken aloud.
4. Name the final act before the hook. A film that opens well and ends on an unrelated metric loses the viewer at the turn, which is the most expensive place to lose them. Say in one line what the last three beats are about, and check that every number in them belongs to something the film already introduced. For a shipped API or feature that means the contract and a runnable CTA, not a benchmark: references/video-formats.md → The third act of a shipped-feature film.
5. Write the plan doc first (SCRIPT.md / scene-plan.md). It opens with the why — theirs, then ours — above the arc, both as single sentences, and where a real person's numbers are published, the risk to them stated alongside. A plan doc that starts at the scene table has skipped the only question the audience cannot see you skip. Get it reviewed before spending generation credits.
6. Structure the piece, not just the hook — references/structure.md. A multi-item piece (a changelog, a roundup, a list) assembled by writing each item well and running them in order reads as a jumble. State the count in the frame line, speak the ordinals, show a 1/5 step counter, and bookend with a contents card that ticks off. Put any confession in the frame line rather than inside one item.
7. Get the AUDIO approved before you composite. A separate gate from the plan doc, and the cheapest one in the run. Generate the VO, normalise it, concatenate the lines into one review file with short gaps, and have the owner listen. On 2026-08-07 this caught a voice the owner's own family had called mechanical — while the timeline did not yet exist. Compositing first would have made a rejected read invalidate the whole build, because a voice change moves every beat and every caption and is a timeline rebuild, not a file swap.
8. Climb the QC ladder cheaply. Render is the last rung, not the first. plan doc → audio approved → stills → render → frames out of the render. Use npx hyperframes@0.7.5 snapshot --at <times> . (pin the version, same as every other invocation; ~15s, writes PNGs and a contact sheet) to look at your work; npm run render takes minutes. One run burned seven full renders on findings that were all visible in a still. Then read frames out of the finished render anyway — a render returning the right duration is not evidence it looks right, and inspect samples fixed points so it misses collisions between them.
9. Audit the SCREEN, not only the script. Run the second half of the references/hooks.md checklist over every panel. It scopes to claims and measurements — structural notation (a 1/5 counter, ordinals), dates, prices shown as UI and cited standards are exempt. For each figure presented as evidence: does it appear in the VO, carry its unit, and come from a customer rather than a test harness? Could two figures on one panel read as a claim you did not make? A run that audited the script twice and the panels never shipped a lab number to camera.
10. Measure audio, do not blindly normalize. House rule: measure first, prefer linear gain, and reserve dynamic normalization for genuinely quiet sources.

Step 5 — Publish

Run the executable pre-flight first; it exits non-zero, so do not publish past it. A checklist validates the copy you can read; only the script validates that the runner will read it. Then publish, then verify the published text against the config with a strict equality check rather than by eye. Both, and the 2026-08-12 empty-tweet-to-the-wrong-account incident behind them: references/publish-verify.md → Gate zero. Full checklists and the per-platform traps: references/publish-verify.md. The gate is short and it has already caught a live defect (a literal <YT link> placeholder about to ship inside a YouTube description, 2026-07-28).

Load the destination before you publish, and declare whether conversion is readable. Two lines in the gate, both cheap, both catching failures nothing else catches:

• Open the CTA URL today. Confirm it returns 200 and that what loads pays off the post's promise. A 200 from a page that cannot convert is not a pass — on 2026-07-30 every CTA in a four-platform slate resolved fine and pointed at a homepage that had produced zero trials in eight weeks.
• Declare attribution: readable or not. If utm_campaign capture is not live, write "conversion unreadable for this slate" into posts-log.md at publish time, not at the re-pull. A tag that nothing records is not measurement, and calling it one hides the gap.

The CTA is a direct tagged link, not a comment-gate. Ruling from our own results: comment-gates were tried twice and produced zero gated comments and zero leads both times. ship-posts still describes comment-gating as the lead-gen option — that advice is refuted by our data; do not run the experiment a third time. Direct link, tagged per references/conversion.md.

LinkedIn gets an IMAGE, not video — because images outperform video there (owner ruling 2026-07-28, from personal-post data). This is a performance decision, not a workaround for the connector lacking video upload. Consequence: do not hand-upload a video to LinkedIn "for consistency" either. The image must depict what the copy's opening line describes.

• Publish + measure: recoup-internal-social-ship-posts. That skill owns copy per platform, the connector mechanics, and the ~48h re-pull.
• Log it: one row per post in posts-log.md, including the arc and character it served, the why it was built on (theirs and ours, as written in the plan doc), and which sign-off line was used. An unlogged arc breaks the serial; an unlogged why means the next run cannot tell whether the reason held up.
• Capture what you learned: if the run produced a durable lesson (a new gotcha, a style that worked, a hook that landed), write it into the workspace doc that owns it — HOOKS.md, VIDEO-STYLES.md, or NARRATIVE.md — not just into the chat.

Step 6 — Did it convert?

Engagement is not the deliverable. A post's job is to move a passive viewer to a signup, and until this step runs we do not know whether any post ever has.

1. Every CTA link is tagged at publish time so the visit is attributable. Convention, capture chain and the per-platform caveats (Instagram cannot carry a per-post link in the caption) live in references/conversion.md. Tracked for implementation as row 29 of chat#1889.
2. At the ~48h re-pull, read conversions alongside engagement. Attributed visits, then signups. recoup-internal-sales and recoup-internal-funnel-valuation-pipeline own the account-side reads (Privy signups, Stripe, credits, Attio stage).
3. Log both in posts-log.md — engagement and what it converted, even when the answer is zero. A zero that is written down is a finding; a zero that is never measured is a story we tell ourselves.
4. Feed it back into step 2. Next run ranks by what converted, not by what got likes. Where the two disagree, conversion wins.

Until attribution ships, say so explicitly rather than implying a post worked because it got engagement. As of 2026-07-28 no post on any Recoup account is attributable to a signup: every CTA is a bare recoupable.dev or "link in bio" with no tag.

Guardrails

• No em dashes in published copy. — and – read as AI-written and cost more credibility than the clause is worth. Rewrite the sentence rather than substituting a comma. Applies to captions, tweets, titles, descriptions, first comments and on-screen text; internal docs are exempt.
• No why, no build. Every piece carries a written why for the person in it, not only for us. Permission is not a reason: "they consented" answers whether we may, never whether we should. A piece that cannot say what its subject gets is an ad using their name, however real its numbers.
• Nothing publishes without explicit owner go-ahead. "My goal is to post X" is a draft instruction, not authorization. Build it, then ask.
• Never re-trigger a live customer's task to test a change. Every run emails the customer. Use a test account.
• Real numbers or nothing, including about ourselves. If a cost or result cannot be verified today, publish the counts you can audit and leave the rest blank.
• Our own accounts only. Customer-facing or artist-facing content goes through the recoup-content-* skills, which are credit-metered and analyze-gated.
• A structural distribution collapse STOPS the run. If step 2 shows a platform's numbers down an order of magnitude, do not build an asset for that platform and hope. Report it and get a decision first. Asset quality is not the bottleneck when distribution is broken, and shipping into it wastes the asset and hides the real problem. (On 2026-07-28 this was flagged twice and built past twice: YouTube Shorts had fallen from 129-397 views to 7-10 and the day's post shipped into it anyway.)
• A BROKEN FUNNEL STOPS THE RUN — same rule as a distribution collapse, one layer down. If the destination is dead-ending or trials have been flat at zero, do not build an asset and hope. Report it and get a decision. The failure mode is worse than the distribution one: a broken destination burns first-time visitors, who do not come back, and it hides behind healthy engagement numbers indefinitely. (2026-07-30: /pricing was the #2 marketing page with zero trials since 06-06, and a funded outreach customer's valuation dead-ended on prod. A full slate shipped into it that day.) The highest-value marketing act on such a day is to fix the path, not to post.
• A build that yields two assets queues the second one. Do not orphan it. Name it as tomorrow's episode in posts-log.md or NARRATIVE.md so the next run finds it instead of starting cold.

References

• references/learn-from-socials.md — the scrape, per-platform metric shapes, and reading gotchas
• references/topic-selection.md — choosing the day's beat, and the gates it must clear
• references/video-formats.md — format routing table and handoffs
• references/hooks.md — hook archetypes, retention constraints, the pre-render checklist
• references/conversion.md — tagged CTA links, the capture chain, per-platform caveats
• references/publish-verify.md — the pre-publish gate, post-publish verification, platform traps
• references/video-pipeline.md — the shared build recipe: scaffold, rebuilt UI panels, captions, render, frame QC
• references/voice.md — model vs voice vs delivery, the bake-off, audio tags, loudness, the checks before compositing
• references/seedance.md — Seedance 2.5 (generated live-action plates): hard constraints, locked-vs-unlocked tasks, the fal call, prompt doctrine, which existing skill owns each downstream step, and the run log we grow
• references/seedance-examples.md — every Seedance 2.5 example prompt we have, verbatim and sourced (ByteDance Seed, BytePlus ModelArk, fal, plus our own runs); read it before writing a prompt
• references/structure.md — holding a multi-item piece together after the hook: the count, the ordinals, the step counter, the bookend
• artist music video — split out into its own customer-facing skill, recoup-music-video. Route there; do not rebuild the pipeline here.