Home · Go-to-market systems · Case study
// a system I built for a brand's organic content
RECON · the research engine
RECON is a research engine that turns a brand's own data into tested organic content ideas. It reads two data streams, hands the best ideas to the marketing manager, and goes back and forth until you land the one worth making. Then it tests the idea, tracks the winners, and loops.
Scroll to run. Every crate is one task automated.
The two engines
Inside the company these are two separate engines that share the same data. One decides what to make. The other makes it. This page is about the first.
RECON This engine
Reads the data, finds what's working, and turns it into content ideas you test. It ends at a proven winner, then loops to find the next one.
reads dataideatestests and measuresranks winners
FORGE
A separate, specialized engine. It generates the actual creative, on-brand images for ads and organic posts, from the same company data. Every output is approved by a person and editable by the marketing team.
gpt-imageImagenFluxIdeogramFireflyHiggsfield
◂ Shared data source ▸ The same proven-performance data feeds both. RECON uses it to decide what to make. FORGE uses it to make it.
Where the ideas come from
RECON goes into the platform and finds what's trending in your world, scoped by the industries, hashtags, topics, and signals you target. Fresh outside ideas, filtered down to what fits you.
RECON pulls from your own track record: your best-performing organic posts and your best-performing ads. New ideas start from what your audience has already rewarded.
Both streams become a ranked short list of content ideas. The marketing manager reviews them, and RECON goes back and forth with them (accept, edit, combine, reject) until you land the final idea. Only then does it get tested. The engine suggests. A person decides.
The dashboard
RECON reports what it has learned on a live board: how many ideas are in flight, which cleared the bar, and which angles tend to win. The figures below are sample data.
Tier distribution
Top angles · reach-weighted
The loadout
RECON is a set of jobs. Each has one role and a shortlist of tools that fill it, so you pick what you already run. Tap a chip to build your own loadout. The engine's code never changes, only the adapter behind each slot.
◂ Your loadout ▸
The coding agent that builds and maintains the engine, docs-first and test-gated.
5 options · swap freely, the engine stays the same
Where ideas get tested (published) and where performance is read back, plus the net-new signal scan. The one live dependency. Everything else is built and tested before it connects.
6 options · swap freely, the engine stays the same
The shared company data: your proven organic posts and ad results, normalized so the engine can tell what works from what doesn't. Direct from the platform, an aggregator, or an in-house dashboard on top of several.
7 options · swap freely, the engine stays the same
Turns a net-new trending video into frames and a spoken transcript, so the engine can read what makes it work, then propose a version for you.
7 options · swap freely, the engine stays the same
Where the engine files the final, approved idea as the next creative request for a person or editor to action.
10 options · swap freely, the engine stays the same
Where source media and finished assets are stored, versioned, and pulled from.
10 options · swap freely, the engine stays the same
Runs the daily cycle on a schedule, unattended. Nobody has to kick it off.
7 options · swap freely, the engine stays the same
Holds API tokens and keys securely, out of the code and out of the repo.
8 options · swap freely, the engine stays the same
How I build
These sit on top of the same 8 steps I use for every build.
Living docs: a current-state handoff, an append-only decisions log, an open-questions list. The chat is disposable. The repo is the source of truth.
Isolate the one live dependency behind an adapter and a sample-fed stub. Build and test everything else first.
Every module ships a plain test. One command runs them all on sample data. Green is the bar, and no account access is needed.
Thresholds, routing, rate limits, and every human gate live in versioned config. The engine proposes. A person decides.
Publishing is the only outward write, gated with preview, approval, idempotency, and a kill switch. Every call is paced like a person would pace it.
Every human touchpoint is listed with its owner and its path to automation, removed later by a config change, never a rebuild. Nothing goes live unapproved.
Commit only what the session touched, keep the gate green, and write every decision as Gap, Recommendation, Why.
# one idea, one measurement window:
WINNER concept C-014 · variant V-03 · +4,120 shares · $0 spend (organic)
→ filed next-round request as ticket TASK-1042Build order
Replicate
The whole method, as one brief. Paste it to your coding agent, fill in the brackets, and it builds RECON around the loadout you picked.
# Build a research engine that turns a brand's data into organic
# content ideas, tests them, and tracks the winners.
TWO DATA STREAMS:
A) net-new signal: scan the platform for what's trending in our
industries / hashtags / target signals.
B) proven winners: our own best organic posts + best-performing ads.
Merge both into a ranked short list of content ideas.
HUMAN-IN-THE-LOOP: the marketing manager reviews the short list; the engine
goes back and forth (accept / edit / combine / reject) to a FINAL idea.
Only an approved idea gets tested. The engine suggests; a human decides.
WORK DOCS-FIRST, repo as memory: HANDOFF.md (state) · DECISIONS.md
(append-only) · OPEN-QUESTIONS.md · ADRs. Read to start, update to end.
BUILD AHEAD OF THE BLOCKER: isolate the live dependency (the social platform
API) behind an adapter + a sample-fed stub. Build + test everything offline.
MY LOADOUT (swap the adapter, never the engine):
ai build agent = <Claude Code>
social platform = <Meta Graph (IG/FB)>
data source = <Platform insights (direct)>
content teardown = <Your own Claude (vision)>
task tracker = <ClickUp>
digital asset manager = <Iconik>
scheduled runner = <GitHub Actions>
secrets manager = <Google Secret Manager>
SHIP A TEST GATE: every module gets a plain, dependency-free test; one
command runs them all on sample data. Green is the bar.
CONFIG OVER HARDCODE + SAFETY: thresholds, routing, rate limits and human
gates in versioned config. Publishing is the ONLY outward write: preview,
approval, idempotency, kill switch, and human-paced API calls.
DISCIPLINE: commit only this session's files, keep the gate green, surface
every decision as Gap / Recommendation / Why.RECON reads what already works, proposes the next idea, and only ships what a person signs off. Then it proves it and loops. It is built behind swappable adapters and verified on sample data, so the one live account is the last switch you flip.
Data-backed ideasA person decidesYou own itBuild offlineConnect last
Questions
It reads two data streams, what is trending in your space and your own best organic posts and ads, and turns them into a ranked short list of content ideas. After a person picks the final idea, it tests it, tracks the winners on a dashboard, and feeds them into the next round.
The marketing manager reviews the short list and goes back and forth with the engine until there is one final idea. Only an approved idea gets tested. Publishing is the only outward action, and it needs a preview, an approval, and a kill switch.
Each layer is a slot with a shortlist of tools, such as Meta, TikTok, or YouTube for the social platform, ClickUp, Asana, or Notion for tasks, and GitHub Actions or Cloudflare Workers to run it on a schedule. You pick what you already use. The engine stays the same.
RECON is the researcher. It decides what to make. FORGE is its sibling engine, the brand creative system. It generates the actual creative, images for ads and organic posts, from the same company data, with every output approved by a person.
Yes. Consulting starts at USD 50 an hour, and builds start at USD 1,500. The engine is built and tested on sample data first, and your live account is connected last.
Related systems
The sibling engine. It turns the ideas Recon finds into on-brand ads and organic posts.
See the breakdown ▸System 01From keyword and AI-answer research to a content plan, briefs, published pages, and lead capture.
System 03Strategy, audiences, ads, landing pages, pixels, and UTMs for Google, Meta, and X, with every lead in your pipeline.
System 04Search, ads, email, CRM, and accounting data joined in one dashboard, so you see which campaigns make money.
Request a quote
Tell me the problem that eats your team's week. Answer the 8 steps below as best you can and skip the rest. I'll send you a plan for how I would fix it, with the tools you already use, and a price. Builds start at USD 1,500. Consulting starts at USD 50 an hour.
Workflows, automations, and code I build for clients are their company property, so I don't share them.
Contact
Email me or send me a message on LinkedIn. I reply to every message.