smm-factory

Context

Noble is a booking SaaS for barbershops and independent salons — the kind of business that lives on Instagram but rarely has time for it. Solo owners already juggle staff, chairs, and no-shows. Marketing consistently falls off the table.

I built smm-factory as Noble's Instagram team: a system that plans a week of content, generates it, posts it, engages with the community, and reports back — end to end — with human approval only where it actually matters.

Problem

Noble had a chicken-and-egg: acquire barbershops on Instagram, but need Instagram credibility first. Doing it manually meant either a full-time SMM specialist ($2–3K/month), an agency ($5K+ and generic B2B templates), or founder-posting (a pattern that collapses in week two).

I wanted something that runs itself but where I stay the taste-maker. Not "AI writes everything" — "AI does everything except the taste calls."

Division of Labor

Model stratification wasn't premature optimization. First month running Opus for everything ≈ $8/day. After stratification: ~$0.30–0.50/client/week. Same output quality where it matters, 20× cheaper.

Prompt Architecture

The system's superpower is the knowledge/ directory — 12 markdown files (00_overview through 11_orchestrator) covering the 8 agents, 3 shared systems (content pipeline, approval manager, database schema), and the orchestrator. One rule I set for myself: Rules = Code. If a rule lives only in markdown, it doesn't exist. Every markdown rule must be in code.

That flipped every Claude session from "here's a codebase, please read it all" into "here's the spec and the one file we're touching." Sessions got sharp.

Every folder in agents/ has a matching file in knowledge/. When the two drift, the knowledge file is the bug — code is ground truth, markdown is the spec that must agree with it.

The engagement agent — where the system actually touches humans

Content generation is the showy part. The engagement agent is the one that moves the needle on a booking SaaS, because it runs while the owner is cutting hair. Two behaviors, both on cron, both bypassing me entirely:

• Comments on target feeds. A daily batch picks in-niche posts (own-niche hashtags from trends, competitor accounts flagged in strategy). Haiku writes one conversational line per post — matched to context, no CTA, no emoji spam. The orchestrator rate-limits by account so the pattern reads organic, not scripted.
• Replies to DMs. Incoming DM triggers the agent inline: Haiku reads it, classifies against response templates in 06_engagement_agent.md, drafts, sends. No Slack-approval on this hop — the whole point of the engagement loop is closing the conversation before the lead cools.

Comment landed · heart + thank-you from author

Started a back-and-forth with the owner

The path I wanted and couldn't take. The clean version is Instagram Graph API through Meta Business Suite — official endpoints for instagram_manage_comments and instagram_manage_messages. App Review for those write-scopes requires passing Business Verification first: a registered legal entity, business documents on file, DUNS or local equivalent, a reviewed privacy policy on an active production domain. I was running this as a solo pilot — no incorporated company behind it, no fast path to one that didn't bulldoze the point of the pilot. Meta's gate doesn't bend for individuals, by design, and there's no sandbox-mode escape hatch for write scopes.

So the fallback is a Playwright session driven by a long-lived IG cookie, same mechanic the publisher uses to post. It's fragile — cookies rotate, flows break on IG UI updates (the login_required in the log below is exactly this failure mode in the trends path) — it's not TOS-clean, and it won't scale past a handful of clients before IG's spam heuristics catch on. For a two-month pilot it was the only way to ship. The 0 IG violations in Results below reads differently once you know this: it's a function of staying slow and under rate-limit, not of the architecture being clean.

00:21:01 [INFO] Generating content for 1 clients
00:21:01 [INFO] Starting content pipeline for Noble
00:21:04 [INFO] Generating week 2026-06-29 for 'Noble'
00:21:11 [INFO] Trends fetched for niche='barbershop booking SaaS' geo='US': 0 trending, 3 rising
00:21:16 [ERROR] Hashtag top medias failed for #barberlife: login_required
00:22:28 [INFO] Instagram trends collected: 0 hashtags, 0 audio tracks, 0 patterns
00:22:58 [INFO] POST https://api.anthropic.com/v1/messages → 200 OK
00:22:59 [INFO] POST https://cloud.leonardo.ai/api/rest/v1/generations → 200 OK
00:23:08 [INFO] Leonardo background generated: bg.jpg (324KB)
00:23:13 [INFO] Uploaded feed_post.png → r2.cloudflarestorage.com/clients/.../2026-06-29/day01/
00:23:14 [INFO]   Day 1 (Monday) [feed_post] saved
00:29:18 [INFO] Day 2: bg_video=found (Pexels, portrait 2160×4096)
00:31:46 [INFO] Day 2: animation.mov ready
00:32:01 [INFO] Day 2: composite OK (2690 KB)
00:32:01 [INFO] Day 2: music → hustle.mp3

Iteration Moment — a decision I unwound in 21 minutes

Ship, test, kill in 21 minutes.

What this taught me: with Claude, the useful skill isn't defending design-pass decisions — it's the speed to unwind them once reality argues back.

Trade-Offs I'd Explain to a Client

• No Meta Business API — cookie-session instead. The clean path for comments + DMs is Instagram Graph API. App Review for the write scopes requires passing Meta's Business Verification, which requires a registered legal entity — I was a solo pilot without a company. No fast way around the gate, by design. Fell back to a Playwright-driven IG session cookie. Fragile, outside TOS, shippable the same evening. The single biggest reason the system isn't production-ready without a corporate wrapper around it first.
• GitHub Actions as render farm instead of a GPU box. 5–15 min latency, offloaded to free tier. Would break past ~50 clients — good problem.
• Slack as the only UI. Every approval, status, insight — Slack. Custom dashboard would've cost weeks; Slack costs zero. Rate limits push me toward batching, which turned out to be a UX win.
• Self-Improvement never auto-applies. Monthly it writes a new strategy version, archives the old, posts summary in Slack. I approve or ignore. Costs me weekly attention. Prevents silent drift into whatever Claude currently thinks is right.

Results — 2-month pilot (2026-07 → 2026-08)

Account · 34 posts, zero manual posting

Reel · 159 views, Haiku copy, FFmpeg render

The honest read: 13 followers is small. This is a genuinely narrow audience (specific service tier, specific behavior). The pilot proved the system can run at cost. The open question it didn't resolve — and the reason this stayed a pilot rather than scaling into Noble's live marketing stack — was whether the creative ceiling of autonomous content was high enough for the same targeting to actually convert.

Reflection