JBS Weekly
This week I'm walking through a project I built using Claude Code's /goal command to generate a full week of marketing content autonomously.
Not 'autonomously' in the vague way people throw that word around. I mean you set the goal, walk away, and come back to a Claude Artifact with seven ready-to-post pieces that each passed an objective quality threshold before you ever saw them.
The key is something know in the AI space as “loop engineering”: instead of prompting once and hoping, you build a system where the AI checks its own work, rejects what doesn't pass, and keeps iterating until it hits your standard. If you're using Blotato for content distribution and you're paying someone to write and edit posts manually, this is worth your full attention.
🛠️ This Week’s Build
The business problem is straightforward. Most small-to-mid companies either hire someone to manage social content or the owner does it themselves between other things. Both are expensive in time or money. The goal here was to replace that manual loop with an autonomous one.
The stack is Claude Code, Blotato's Content Pack skills, and Anthropic's API with access to their Fable model. Before you touch anything else, you install all 7 skills from the Blotato Content Pack into Claude Code. Instructions are at help.blotato.com/claude-skills/claude-skills. Unzip the file, copy the 7 skill folders into ~/.claude/skills/, restart Claude Code, and you're ready.
Once the skills are loaded, you write a /goal prompt using five components:
Task (what you want)
Motivation (why it matters, which gives the model context)
Outcome (the specific finished result)
Constraints (guardrails like which tools to use and when to stop)
Verification (the part most people skip).
Verification is what makes this different from a normal prompt. The /goal command runs a secondary model, in this case the PostGrader skill from Blotato, which is pre-trained on viral content metrics. Every post the Writer skill generates gets scored by the PostGrader against objective criteria: hook quality, platform-specific formatting, CTA effectiveness. If a post scores below your threshold, say 7/10, the loop rejects it and sends it back for revision. The loop keeps running until every post clears the bar.
When it finishes, you get a Claude Artifact with a full week of content ready for review. You're not reading drafts. You're approving posts that already passed a quality check you designed.
A few things that matter in practice. Use Fable for initial setup when the agent is learning a new environment, then switch to a cheaper model like Haiku for the recurring loop. Fable costs more and the maintenance loop doesn't need that level of reasoning. Also, when the agent first starts requesting permissions, use the 'Always allow' option after reviewing the first several requests. Otherwise you'll be babysitting approvals for the full run. The Claude mobile app lets you handle any blockers remotely if the agent stalls.
The takeaway: the quality of your output is only as good as how objectively you define 'good.' If your verification step says 'make sure it sounds good,' the model will lie to you and say it does. If your verification step says 'score must be 8/10 or higher across these four criteria,' the model has no room to fudge it.
📰 AI News This Week
Slackbot Is Now an AI Agent Inside Your Existing Slack Workspace
Slack's updated Slackbot pulls in prior team conversations, past decisions, and accumulated context so it can act more like a teammate and less like a search bar. It can read, write, and interact across connected apps, while respecting existing permission structures so it only sees what the user is allowed to see. This is a meaningful upgrade from the trigger-response bot most teams have been ignoring for years.
Joe's Read: This affects ops teams at companies already running Slack, specifically anyone who has been manually briefing new hires or repeating context across channels because institutional knowledge lives in people's heads instead of a system.
OpenAI Releases GPT-5.6 and a New Real-Time Conversation Model
OpenAI dropped GPT-5.6 alongside GPT-Live, a model built for smoother, more natural back-and-forth conversation inside ChatGPT. This is part of a dense week of releases across OpenAI, Meta, Google, and Grok. The practical question is whether GPT-Live changes anything for teams already using voice or chat interfaces in customer-facing workflows.
Joe's Read: This affects small teams using ChatGPT for customer support scripting or internal Q&A, where more natural conversation handling could reduce the editing pass required before output is usable.
Ava 2.0 Automates the Full Outbound Sales Workflow Without a Human in the Loop
Artisan's Ava 2.0 handles prospecting, personalized outreach, objection handling, and meeting booking from a database of over 300 million contacts, with no human required at each step. It runs multivariate testing on its own campaigns and supports cold, warm, upsell, and signal-triggered outreach types. Companies like Quora and SumUp are already using it, and there's a $300 free credit to start.
Joe's Read: This affects small sales teams at B2B companies where one or two people are managing outbound manually, specifically anyone spending more than a few hours a week on prospecting and first-touch sequencing.
Read more: https://dashboard.artisan.co/sign-in
🧰 Tool Worth Trying This Week
Ava 2.0 by Artisan
Ava 2.0 is a fully autonomous AI sales rep that handles outbound prospecting, personalized outreach, and meeting booking without requiring a human to manage each step. It pulls from a 300 million contact database, runs its own multivariate testing, and handles objection responses automatically. If your outbound process currently lives in one person's head and calendar, this is worth a close look.
Caveat: This is built specifically for outbound B2B sales workflows. If your business runs primarily on inbound leads, referrals, or relationship-based sales cycles, Ava won't map cleanly onto how you actually close deals.
🤔 Joe’s Take
The part of this build that surprised me most was how much the verification step changes the management burden.
Most people think the hard work is writing a good prompt. It's not. A good prompt gets you a decent first draft. A good verification step gets you a finished product without you in the loop for every revision cycle.
The PostGrader acts as a QA function, not a writing function. That's an important distinction. You're not asking the model to be creative and rigorous at the same time. You're separating those roles the same way a real team would: one person writes, one person edits against a standard. The model is just doing both, in sequence, autonomously.
If you apply that same thinking to any repeatable process in your business, the question becomes: what is the objective standard for 'done'? If you can define that clearly enough for a secondary model to check it, you can probably automate the loop.
⚒️ Tools I Use
n8n — The automation tool I use to connect apps, trigger workflows, and stop doing things manually. If there's a repetitive process in your business, this is where you start fixing it.
VoiceInk — A local AI dictation tool for Mac that transcribes your voice with near-perfect accuracy and runs entirely on your device, meaning nothing you say ever touches a cloud server.
Blotato — This week's build runs directly through Blotato's Content Pack skills inside Claude Code. The PostGrader skill is what handles the verification loop, scoring every post before it ever reaches you, and the Writer skill handles generation. If you're distributing content across platforms after the loop runs, Blotato handles that side too, publishing natively to 9 platforms with no per-post fees.
Beehiiv — What you're reading right now is published on Beehiiv. If you're thinking about starting a newsletter or moving off a clunky platform, this is the one I'd recommend. 20% off your first 3 months with my link.
Google Workspace — Beyond email and Docs, a Business Standard plan includes Gemini Pro built into every app, NotebookLM Plus, and access to the enterprise versions of the whole suite. Better value than a standalone Gemini subscription when you're already paying for Google anyway. 14-day trial and 10% off your first year.
Descript — Video and podcast editing that works like a text document. You edit the transcript and the media follows. Cuts filler words, cleans up audio, and handles captions automatically. 50% off your first two months on the Creator Plan.
💭 Final Thoughts
Most automation projects stall at the prompt. People write a good instruction, get a decent result, and still end up doing cleanup by hand. The fix is designing verification before you design the task. If you know exactly what 'done' looks like, you can hand that definition to a secondary model and let the loop run without you. That's not a Claude-specific insight. It's how any reliable process should work.
PS: If you want a week of ready-to-post content running on a loop without manually reviewing every draft, book a discovery call and we'll build the solution together in a live session.
Cheers,
Joe