Social Media Production Engine

A cold lead list is usually used once, for outreach, and then sits idle. This concept build puts it back to work: it scrapes every lead's site, turns what those companies actually say about themselves into five distinct content angles, and — once a human approves an angle — writes a LinkedIn post, a Facebook post, and an Instagram caption from it in parallel.

At a modeled 25 minutes of manual research and drafting per lead (range 20–30 minutes) across a batch of 50 leads, this replaces roughly 20 hours of manual work per batch. That figure is an illustrative estimate based on a reasonable time range, not a confirmed client result.

Turning a lead list into social content usually means a person opening each company's website, figuring out what they actually do and care about, and drafting a post around it by hand. That research step is the easiest thing to skip under time pressure, and skipping it leaves a generic template repeated across every lead, which reads exactly like what it is.

• Cold lead lists get used once for outreach, then sit idle
• The research step — reading each site to find what a company actually cares about — is the first thing dropped under time pressure
• Without it, the fallback is one generic template repeated across every name on the list
• Even when the research gets done, writing for three platforms triples the work, so one platform gets the good version and the rest get copy-paste

Separate idea generation from content writing. The system runs in two stages instead of one pass. Part 1 scrapes every lead's site and turns the combined content into five distinct angles per batch: cheap to generate, nothing committed yet. Part 2 only runs once a human picks a specific angle worth using, and that is when the system spends effort writing three platform-specific versions. Content generation that nobody will use never happens.

Ground every angle in the lead's actual site, not a prompt template. Each batch's five angles come from scraping and combining the real text of every site in that batch, so the audience appeal and messaging points are specific to what those companies say about themselves rather than a generic industry angle applied to everyone on the list.

Catch API shape mismatches before they silently break downstream steps. Two response-shape gotchas would otherwise have failed quietly. The OpenAI Responses API (v2.3) returns generated content at output[0].content[0].text rather than the older message.content shape, so the angle array lived at output.0.content.0.text.angles. Get that path wrong and the “split into five rows” step just returns nothing, with no obvious error. Separately, Airtable's get-record operation nests all fields under fields["Field Name"] while create-record does not. Mixing the two up in Part 2's angle-lookup step would have silently read undefined into every generated post instead of the actual angle text.