The Marketing Skill: Turning Saved Inspiration Into Grounded Content Ideas

Ask a general-purpose AI chatbot for video content ideas and you'll get something confident, well-formatted, and almost useless: post consistently, hook viewers in the first three seconds, ride trending audio. Every point true, none of it specific to the business asking. That's the ceiling of ungrounded brainstorming, and it's the ceiling we built our Marketing Skill specifically to break through, by making it structurally incapable of answering without looking at real evidence first.

The problem: generic advice dressed up as a suggestion

The actual failure mode of most AI content brainstorming tools isn't that they're wrong, it's that they're untethered. A model with no grounding will happily generate ten video ideas that read like a content marketing 101 checklist, because that's exactly what its training data is full of: generic, widely repeated advice about what tends to work in general. What it can't do on its own is know what's actually working for one specific business, in one specific niche, with one specific audience's taste, because that information was never in its training data to begin with. It lives somewhere else: in what that business's own team has already noticed resonating.

What we built: a Skill that has to look before it speaks

The pattern we've used across every product on our harness applies directly here: a Skill is a system prompt plus a scoped set of tools, and this one gets exactly one tool, a reference lookup. The idea is straightforward: as someone comes across Instagram, YouTube or TikTok videos they find genuinely interesting, on brand, or worth learning from, they save them. The Skill's entire job is to turn that growing pile of saved taste into concrete new ideas, and its instructions are explicit that it has to call the lookup tool and actually look at what's saved before it's allowed to brainstorm anything, never falling back on generic content-creator advice as a substitute.

What the lookup actually returns is more than a list of links. For directly uploaded or downloaded reference videos, it includes a description and the on-screen text pulled from genuinely analyzing the video content, not just its caption. And for every reference, it surfaces real platform data: who posted it, how many likes, comments and views it has, and which audio track it uses when the platform exposes one. That last detail matters more than it might seem. Recurring music or audio choices across several saved references are themselves a signal, a pattern in what's resonating that a human skimming a folder of saved links might not consciously notice, but that becomes visible once it's laid out data point by data point.

Why grounding beats a generic list

The instructions push the Skill to look for what its saved references actually have in common, format, hook style, pacing, topic, tone, and use those specific patterns to shape its suggestions, rather than reaching for a generic content playbook. That's a meaningfully different task than "give me ten ideas." It's closer to "look at what this business has already told you it likes, and its audience has already told the algorithm it engages with, and extend that pattern forward," which produces suggestions that feel like a natural next step for a specific brand rather than something that could have been generated for literally anyone in any niche.

What the output actually looks like

The Skill's instructions require a specific shape for every idea it proposes: a hook, the first one to two seconds that has to earn the rest of the video's watch time, a short beat-by-beat outline of how it plays out, and, critically, which saved reference or references it drew the idea from and why. That last part is what keeps the tool honest and useful rather than a black box. If a suggestion doesn't trace back to something real, a genuine pattern across saved references, it doesn't get generated in the first place, and if it does, the person reading it can immediately see the actual evidence behind the suggestion rather than having to trust the AI's judgment on faith.

Read-only, and deliberately so

The reference lookup tool needs no approval before it runs, the same reasoning we apply to every structurally read-only tool across the harness: it can only look at what's already saved, it can't change or delete anything, so requiring a human to approve every single lookup would add friction without adding any actual safety. That's what lets the Skill's instructions demand it call the lookup before every single brainstorming request without that becoming annoying to actually use.

How this connects to the rest of a real content workflow

This Skill isn't meant to be the entire marketing operation, it's the ideation layer feeding into a workflow we've written about separately: testing new content safely before committing it to a main channel. Once this Skill produces a grounded idea, tied to a real hook style or pacing pattern that's already resonating, that idea becomes a candidate for the kind of structured testing we describe in our piece on using Instagram Trial Reels to test content variations without risking a main feed. The loop closes naturally: save what performs, ground new ideas in it, test the variations privately, and whatever wins becomes tomorrow's new reference to save and learn from. Ideation without a testing step is a guess with extra confidence. Testing without grounded ideation is throwing content at a wall. Neither is complete alone, which is part of why we treat them as one continuous pipeline rather than two separate tools.

What it deliberately doesn't do yet

It's worth being precise about scope rather than overstating what's built. This Skill generates ideas for a human to act on. It doesn't draft captions on its own initiative, schedule posts, or touch any social platform's publishing API, and that boundary is deliberate, not a gap we haven't gotten to. Every other Skill on our harness that can take a consequential action, sending a message, writing to a database, has that action gated behind human approval by default; a Skill whose only tool is a read-only lookup simply hasn't been extended into that territory yet. The natural next step, if a client wanted it, would be adding a tool that drafts a caption or a shot list for a chosen idea, still stopping short of anything that publishes without a person reviewing it first, consistent with how every other write-capable tool on the harness already works.

Why this runs on the harness instead of as a standalone tool

We could have built this as a small, standalone script bolted onto a chat interface, and it would have worked for a demo. We built it as a Skill on our existing harness instead, which means it inherited a working memory system, a database its future write-capable tools could safely use, and a self-monitoring pass, all proven elsewhere, on day one rather than needing to be built or debugged again from scratch. If this Skill grows into something that drafts captions or schedules posts, the approval gating and reliability guarantees that would need to sit around those actions already exist and are already tested. That's the entire point of building one harness instead of one-off tools: the marketing Skill got that infrastructure for free, the same way the tutoring Skill and every future product will.

A concrete example of grounded versus generic

Say a business has saved a handful of reference videos, several of them short, fast-cut clips using a specific trending audio track, high engagement, comments mostly asking a follow-up question rather than just reacting. A generic brainstorming request produces the usual list: post more, use trending sounds, ask questions in captions. This Skill, grounded in the actual saved data, produces something narrower and more useful: an idea explicitly built around that same fast-cut pacing and that specific audio track, with a hook styled to prompt the same kind of follow-up-question comment its references are already getting, and a note explaining that's exactly why, pointing at the specific saved videos the pattern came from.

The difference isn't that the second version is more creative. It's that it's falsifiable, someone can check whether the pattern it's pointing at is actually real by looking at the same saved references, and disagree with the read if it doesn't hold up. A generic suggestion can't be checked against anything, because it wasn't built from anything specific in the first place.

Why this matters for how we do content work for SMEs

The actual value for a business isn't that an AI generates ideas faster than a person could, plenty of tools do that. It's that the ideas it generates are explainable and specific rather than generic and unfalsifiable. A business owner reviewing this Skill's output can see exactly why an idea was suggested, which of their own saved references it's built on, and judge for themselves whether that reasoning holds up, the same way they'd judge a suggestion from a human strategist who could point to specific evidence rather than gut feeling. That's the bar we hold AI-assisted content work to across every client engagement: useful because it's grounded, not just because it's fast.