SUBJECT PROFILE 01
Recruiting & staffing agencies
Constant posting need, public competitor content, results tied to a measurable pipeline.
Case File 04 · Workflow Blueprint · Declassified
A bespoke workflow for LinkedIn-first B2B companies: agents research your competitors, trends, and traffic, draft written, image, and video content behind hard brand gates, and stage everything in an approval queue with a scheduler — and nothing posts without a human click.
BLUEPRINT — AN EXAMPLE OF THE BESPOKE SYSTEMS I DESIGN FOR CLIENTS · ARCHITECTURE PROVEN IN MY OWN PRODUCTION PIPELINE · STATUS: NOW BEING IMPLEMENTED FOR MY OWN BRAND — RECEIPTS TO FOLLOW
Who This Is For
This workflow is designed for B2B companies of roughly 5–50 people whose pipeline runs on LinkedIn visibility — where consistent, on-brand content is the #1 marketing lever and nobody has time to be a content team.
Constant posting need, public competitor content, results tied to a measurable pipeline.
Thought leadership literally is the funnel — the partner's voice can't go dark for a month.
Compliance requires human review before anything posts — this architecture makes that a feature, not a bottleneck.
Already sold on AI content — what's missing is the system that keeps it on-brand at volume.
The Problem
The voice that wins deals belongs to one or two busy people. Content happens in bursts when they have a free evening, then goes silent for weeks — and the algorithm punishes silence.
Outsourced content is expensive and generic — ghostwriters who don't know the business produce posts that sound like everyone else's, and every revision cycle costs days.
Raw AI tools generate volume without judgment: off-brand phrasing, invented claims, wrong audience. Ungated automation posting in your name is a reputation incident waiting to happen.
The Workflow
The same architecture as my production agent pipeline — cheap models for volume work, strong models where quality is the bottleneck, hard gates instead of vibes, and a human holding every exit — aimed at content operations.
Public posts · industry news
Site analytics · engagement
Fast, cheap models condense both lanes into one brief.
Drafted, not decided.
Minutes, not hours. Strategy stays human.
Voice guide · visual kit · banned-topics list
Written — posts · articles · carousels | Image — branded graphics on templates | Video — scripts · short-form cuts
Brand-voice rubric · source check · banned topics · platform specs. Fail = back to draft, never forward.
Approve, edit, or reject. Nothing auto-posts.
Missed slot = skip + notify.
FEEDBACK LOOP — ENGAGEMENT DATA FLOWS BACK INTO NEXT WEEK'S INTEL BRIEF · WEEKLY DIGEST VIA TELEGRAM OR EMAIL · API BUDGET CAPS ENFORCED ON ALL KEYS
Stage by Stage
Scheduled agent runs scan competitors' public posts and industry sources, while a second lane reads your site traffic and past post performance. Both condense into a weekly intel brief: what's landing, where the gaps are, which of your topics actually drive visits. High-volume, low-stakes work — routed to fast, cheap models.
From the brief, the agent proposes the week: themes, angles, formats, and posting slots. The owner reviews in minutes — keep, kill, or redirect. Strategy stays human; the agent just makes the decision cheap to arrive at.
Written posts and articles go to a strong writing model grounded in your sources and voice guide — quality is the bottleneck there. Branded graphics render onto locked visual templates. Video means scripts and short-form cut plans a phone can execute. Which model handles which task class is decided by discriminator-task evaluation, per client — never by brand loyalty.
Every draft must clear deterministic gates before a human ever sees it: a brand-voice rubric score with anchored scales, a source check that flags any claim without grounding, the banned-topics list, and platform-spec checks (length, dimensions, hashtags). Failing drafts go back, never forward — the same soft-heuristics-fail lesson my production pipeline learned the hard way.
Everything that survives the gates stages to an approval queue the owner can clear from a phone: approve, edit, or reject, with the draft, its sources, and its gate scores in one view. A post published in your name is a reputation-carrying action — it gets the same human finish as my own pipeline's applications. For regulated firms, this queue is the compliance record.
Approved content flows into per-platform slots and posts via API or scheduling tool — or stays in click-to-post mode if you want full manual control. A missed slot skips and notifies; the system never improvises. Engagement data feeds next week's intel brief, and a weekly Telegram or email digest shows what ran and what it did.
Model Routing
| Task class | Model class | Why |
|---|---|---|
| Flagship written content | Frontier writing model | This is the output that carries the founder's name. Quality is the bottleneck — it gets the strongest model, grounded in sources and the voice guide. |
| Trend & competitor scanning | Fast extraction model | Runs constantly across many sources. Volume is the bottleneck; cheap per call matters more than eloquence. |
| Rubric scoring & gate checks | Mid-tier model | Judging against an anchored rubric is easier than writing — a mid-tier model clears the bar at a fraction of the cost. |
| Branded graphics | Image model + locked templates | Generation happens inside brand templates — fonts, colors, and layout are constraints, not suggestions. |
| Scheduling & digests | No model — deterministic code | Posting at 9:15 Tuesday is not a judgment call. Anything that can be plain code, is. |
Specific models are chosen per client through discriminator-task evaluation on your actual content, with API budget caps on every key — so the stack is an evidence-based decision that gets re-tested as models improve, not a subscription that quietly renews.
Deployment Options
The entire pipeline runs as scheduled Claude Code sessions (Routines) in the cloud: an orchestrator model coordinates cheaper worker models, research runs on built-in web tools, state lives in a private repo, and drafts stage to your review queue.
For higher volume, non-Claude models in the routing mix, or pipelines that must run inside your own environment: an agent framework on a small server with per-call API access — the same architecture as my production pipeline.
Which option fits is a cost question with real numbers — posting cadence × content volume × model mix — estimated in discovery, not guessed.
Positioning
The founder's judgment stays in the two places it matters — the calendar and the approval queue — and stops being spent on blank pages, resizing graphics, and remembering to post.
The system is built once, documented, and owned by you — grounded in your sources and voice rather than a ghostwriter's best guess, with revision cycles measured in minutes in a queue.
The difference is everything around the model: grounding, brand gates, an audit trail, budget caps, and a human between the draft and your reputation. Tools generate; systems are accountable.
How It Gets Built
Build the brand inputs once: voice guide from your best existing content, visual kit, banned-topics list, competitor set, and the quality rubric your gates will enforce.
The pipeline runs with every gate at maximum and a human reviewing everything. Early rejections aren't failures — they're the training data that tunes the rubric and the templates.
Your team owns the queue, the calendar gate, and the documented system. Oversight eases where the gates have earned it — the human click on publishing never goes away.
What gets measured is instrumented from day one: hours spent per post before and after, drafts produced per week, gate rejection rate, and engagement per post — so the system's value is a number you read, not a claim you're asked to believe.
This blueprint is the starting point — discovery tunes it to your brand, your compliance needs, and the platforms your buyers actually read.