Sponsored by
From Our Sponsor:
Everything is coming into focus.
Join beehiiv live on July 16th at 1PM ET for a first look at the future of audience-led business.
This isn’t just another feature launch (though there will be plenty of those). It’s a look at a more connected future for creators and brands that are tired of juggling disconnected tools, platforms, and data.
If you care about building an audience online, this is worth your time.
Guide: How to Get Claude to Revise Its Own Work Until It's Actually Good
Most people use AI the same way they'd manage a new hire they don't trust: give one small task, check the output, give the next small task, check again. You're the engine. The AI only moves when you push it.
Loop engineering flips that dynamic. Instead of walking Claude through every step, you define what "done" looks like, give it criteria to check against, and tell it not to come back until every box is checked. The AI drafts, scores its own work, finds the weak spot, rewrites, and keeps going until it actually clears the bar — all within a single conversation turn. You get the finished output, not the first draft.
The term was coined in June 2026 by Addy Osmani at Google, building on statements from Boris Cherny (head of Claude Code at Anthropic) and Peter Steinberger. But while nearly everything written about it targets developers using terminal-based tools, the core technique works for anyone with a Claude Pro plan and a Project. No code, no installation, nothing to configure.
Every loop needs three pieces: a goal (what "done" looks like), a check (how the AI knows it's not done yet), and a stop condition (when to quit trying). Without the check, the AI just agrees with itself. Without the stop condition, it runs forever.
What You'll Need
Claude Pro plan (for Projects with custom instructions)/ The principle also works with ChatGPT and Gemini
A repeating task where "done" is measurable, not a matter of taste
Success criteria you could explain to a new hire - if you can't describe how to tell whether the work is finished, you can't describe it to Claude either
For the advanced section: Claude Desktop with Cowork (optional, covered in the full SOP)
Step 1: Know When a Loop Is Worth Building
A loop earns its setup time only when all four conditions are true:
The task repeats at least weekly. Anything less frequent and the setup cost never pays back. A one-off is still better served by a single well-crafted prompt.
Something can objectively reject bad output. A character count, a keyword checklist, a structural requirement. If nothing can fail the work automatically, the loop just spins.
The AI can complete the work end-to-end without handing half of it back to you for judgment calls at every step.
"Done" is objective. A loop can produce a listing draft, but it can't decide your brand positioning. If quality is purely a matter of taste, a human still wins.
Miss one of these, keep it as a manual prompt.
Step 2: Build Your First Project-Level Loop
This is the version any seller can use today. You paste a loop protocol into the custom instructions of a Claude Project, and every conversation in that project follows the protocol automatically.
Anthropic's own prompting documentation explicitly recommends appending "Before you finish, verify your answer against [test criteria]" to system prompts. What you're doing here is making that more rigorous.
The Loop Protocol Template
Paste this into your Project's custom instructions, then customize the success criteria section for your specific task:
You work in a loop until the task meets the bar. Do not present work to me until it passes.
LOOP PROTOCOL (run every turn):
1. PLAN — state the single next step
2. DO — produce or improve the work
3. VERIFY — score the result against the success criteria below. Be brutally honest. List exactly what is still weak.
4. DECIDE — if every criterion scores 8+ out of 10, present the final work. Otherwise, fix the weakest score first and go again.
RULES:
- Never present work until every criterion is 8 or higher
- Each pass must fix the weakest score from the last VERIFY step
- Do not ask me questions unless something is genuinely ambiguous. Make a sensible assumption, note it, and keep going.
- Show me only the final output unless I ask to see your process.
SUCCESS CRITERIA:
[This is where you customize for your specific task]The success criteria section is everything. Vague criteria like "make it good" produce vague loops. Measurable criteria like "title under 75 characters, first bullet starts with a benefit, all five keywords from the provided list appear naturally" produce real improvement.
Step 3: Use Ecommerce-Ready Loop Templates
Drop any of these into a Project's custom instructions to get a self-checking loop for that specific workflow.
Amazon Listing Optimization Loop
Add these success criteria to the loop protocol template above:
SUCCESS CRITERIA:
- All keywords from the provided keyword list appear naturally in the listing
- Every bullet point leads with a benefit, not a feature
- Title is under 75 characters (Amazon's enforced limit as of July 27, 2026)
- All five bullet points are present and complete
- Item Highlights field uses its 125-character allowance to carry keywords and details that no longer fit in the shortened title
- Language is optimized for Alexa for Shopping (Amazon's AI shopping assistant, formerly Rufus) — answers likely customer questions directly with clear use-case and comparison language
- Each criterion scores 8+ out of 10Email/Newsletter Draft Loop
SUCCESS CRITERIA:
- Hook appears in the first line
- One clear CTA, not multiple competing asks
- Total word count is under [your target]
- Tone matches brand voice guidelines defined above
- Subject line is under 50 characters with a curiosity gap
- Each criterion scores 8+ out of 10Competitor Analysis Loop
SUCCESS CRITERIA:
- At least 5 specific data points cited with sources
- Every recommendation is actionable, not just an observation
- Each recommendation is tied directly to supporting evidence
- Recommendations are prioritized by effort vs. impact
- Each criterion scores 8+ out of 10Step 4: Understand What Changes with Cowork
The Project-level loop from the steps above still requires you to open a chat, paste in your brief, and sit there while Claude iterates. You're still the trigger. Close the tab and it's gone.
Cowork changes three things. First, Claude acts on real files instead of generating text in a chat window — it writes directly into documents in your folder. Second, Skills make the loop reusable without re-explaining the context every time. Third, Scheduled Tasks put the loop on autopilot: you define a cadence, and it runs while you're doing something else.
For the rest of the SOP — including three ready-to-use Cowork loop configurations (weekly competitive price monitoring, content production pipeline via Dispatch, and weekly PPC health checks), plus the honest cost assessment of when Project-level loops are enough and when Cowork loops genuinely warrant the extra compute — a free gift from me :)
Do You Love The AI For Ecommerce Sellers Newsletter?
You can help us!
Spread the word to your colleagues or friends who you think would benefit from our weekly insights 🙂 Simply forward this issue.
In addition, we are open to sponsorships. We have more than 66,000 subscribers with 75% of our readers based in the US. To get our rate card and more info, email us at [email protected]
The Quick Read:
AI vendors sell replacement fear to grab attention, but it's a trap: Yale finds no AI job displacement over 33 months, NY layoffs cite zero AI reasons, Klarna rehired after cuts hurt quality. Position AI as enhancement, not replacement.
LLMs answer differently each time because they sample probabilistically by temperature, stating confident guesses as fact. As stakes rise in cyber and biotech, expect verification layers like Lean 4 that answer only when proven.
Cannes' polished AI talk hides a quieter admission over dinner: throwing a full codebase at models gets astronomical, and with rising token costs it's now vastly pricier than just hiring engineers, says Converge Digital's CEO.
Ads Engine pulls creatives from Google, Meta, and LinkedIn, localizes text, image, and dubbed video into 50+ languages in one session, then flags ad fatigue to refresh before performance drops. Weeks of production become one pass.
Taste is easy now that influencers spread it in real time, so curation stops winning. Bandit, Liquid Death, and Chanel break out by creating new worlds, not picking from existing ones. The AI flood rewards whoever refuses to blend in.
Gemini Spark runs as a 24/7 agent under your direction: auditing project trackers against meeting notes, aggregating Gmail newsletters, organizing Drive, prepping client meetings, and coaching reps to counter discount requests.
In under 12 months AI code generation went mainstream in marketing teams, who now build their own dashboards, automations, and tools with no engineers. They replace vendors and rethink hiring, once they push past the hard early hump.
OpenAI will limit GPT 5.6 to vetted partners approved customer by customer at the Trump administration's request, mirroring Anthropic's Project Glasswing. Frontier cyber models find and exploit software bugs faster than any analyst.
Google launches Ask Ad Manager, an AI assistant helping publishers troubleshoot underdelivering campaigns and benchmark bidder performance. It suggests but never decides, for now, as Google builds toward agentic ad negotiation.
The Tools List:
🪄 Magic - An optimization tool that upgrades your AI workflow with integrations with Bard, Claude, Midjourney, & more.
🎨 Basedash AI - Visualize, edit, and explore your data with AI.
⚙️ The calibration game - Get better at identifying hallucinations in LLMs.
📹 VEED Captions App - The simplest way to create engaging short-form videos.
🤗 Dorik AI - Generate beautiful websites from a single prompt.
About The Writer:
Jo Lambadjieva is an entrepreneur and AI expert in the e-commerce industry. She is the founder and CEO of Amazing Wave, an agency specializing in AI-driven solutions for e-commerce businesses. With over 13 years of experience in digital marketing, agency work, and e-commerce, Joanna has established herself as a thought leader in integrating AI technologies for business growth.
For Team and Agency AI training book an intro call here.
