Your output still waits on you
Every recurring GTM piece still passes through your eyes as the only quality gate. Here is how to hand the check to a self-scoring loop and read the checks, not the drafts.
Save this. 3 AI moves from this week’s Claude Signal, each with my operator take and a move a lean European team can run today, plus the play behind it.
Send it to 1 founder or GTM operator scaling toward 10M in Europe who reads everything and ships none of it.
Applied AI: Write The Loop - Friday July 17, 2026
For a year the pitch was to prompt Claude better. This week the operators pulling ahead stopped prompting and started writing the loops that prompt it for them.
10 signals from this week’s Claude Signal, across 5 tracked creators. 3 problems. 1 playbook to steal.
Loop = an agent that runs Goal, Act, Check, Repeat. It writes down what done looks like, takes a step, scores the result against that, and rewrites until it passes.
Rubric = the short written standard the loop grades against, the 5 lines that separate your best output from your worst.
The 90-second version
The shift: stop prompting, write loops. Boris Cherny, who runs Claude Code, says his job is now to write loops that prompt Claude, not to prompt it himself.
A loop is Goal, Act, Check, Repeat. Charlie Hills built a rubric from 110 past editions, and a fresh agent graded this week’s draft 51, then 95, before he read a word.
Intelligence got metered. Fable 5, the top tier, moved to pay-per-use on July 12. Send strategy to the top model, routine execution to the cheap one, or the bill runs away.
Polished AI output is now the floor. The model drafts, the human still owns the taste pass.
The European read: a lean team spread across markets has no second set of eyes to QA every piece, so the loop is the QA hire you cannot make, and your operator take is the one thing a competitor’s agent cannot copy.
Read time: 7 min
This week’s number 1 move
Stop writing the task. Start writing the check.
For a year the pitch was to prompt Claude better. This week the person who leads Claude Code, Boris Cherny, put the next step plainly: “I don’t prompt Claude anymore. I have loops running that prompt Claude. My job is to write loops.” Charlie Hills shows what that looks like on real GTM work. A loop has 4 beats: Goal, what done looks like, written first; Act, the model takes 1 step; Check, something scores the result against the goal; Repeat until it passes. The beat everyone skips is the check, and it is now the whole job. Charlie built a rubric from his 110 best and worst past editions, and a fresh agent scores every draft cold. This edition went from 51 to 95 before he read a word.
My take: this newsletter runs on a version of this. A loop scores the draft against a rubric before I read it, so my time goes to the judgment, not the first pass. Writing the check instead of the task is the real change in the job, and it is the cheapest edge a lean team has this year.
Do first: take 1 recurring output you already produce and write, in 5 lines, the rubric a stranger would use to score it 0 to 100. That rubric is the check for your first loop.
The rule: if you can write the rubric, the loop can score it. Read the checks, not the drafts.
Sources: Charlie Hills, “How to (actually) use Fable 5” (Jul 5) · Ruben Dominguez, “Claude Code Loops” (Jul 7)
Hi, it is Koen Stam and welcome to GTMcraft: The Future GTM Operator. This newsletter is built from 100,000+ GTM signals collected from 100+ operators and founders, combined with 13+ years of my own lessons and failures from the trenches. I write at the intersection of go-to-market practice and AI-powered systems for founders scaling 0 to 10M+ ARR.
100,000+ GTM relevant signals from LinkedIn, Newsletters and Podcasts indexed. Translated into 100+ GTM plays for you to implement today.
13+ years over 3 GTM operator jobs across 3 GTM motions (SMB, MM, ENT). Scaling from 2-10M+ ARR multiple times. Same recipe. Different motions.
Now all part of GTMcraft OS.
Problem 1: Your recurring GTM output still runs through your own eyeballs
Every week you write the same shaped thing, a follow-up, a one-pager, a weekly digest, and you are the only quality gate it passes. That is the bottleneck. Charlie Hills removed himself from it by writing a check instead. He had an agent export his 110 past editions with their open rates, compared winners to flops, and turned the gap into a rubric out of 100. Now a fresh agent scores every draft, sends anything under 95 back with notes, and only surfaces the pass. It compounds, because every miss he catches by hand becomes a new rubric line the loop never misses again. Ruben Dominguez maps the same idea onto the 4 loop types in Claude Code: turn-based, goal-based, time-based, and proactive. The work you repeat is exactly the work a loop can score.
My take: I would rather write the rubric once than re-read every draft. On a lean team the loop is the QA hire I cannot make, and it gets stricter every week instead of tired.
The European read: a lean multi-market team has no second set of eyes to check every piece, in every market’s language. The loop scores your outbound and your digests against your own rubric, so quality does not drop just because you ran out of hours in one country.
Do first: pick the 1 output you produce most often and write the 5 lines that separate your best version from your worst.
Do this week: in Claude Code, set the loop to draft, score against the rubric, and rewrite anything under the bar, keeping the judge separate from the writer.
Do this month: every miss you still catch by hand becomes a new rubric line, so the loop never misses that one again.
The rule: if you can write the rubric, the loop can score it. Keep the judge separate from the writer.
You know it worked when: you read a score and a pass, not a first draft, and the drafts you do read are already at your bar.
The play: 5 steps to build a measurable loop that scores its own output (2026 series). Turn 1 recurring output into a Goal, Act, Check, Repeat loop that grades itself against your rubric. Get full access below ↓
Sources: Charlie Hills, “How to (actually) use Fable 5” (Jul 5) · Ruben Dominguez, “Claude Code Loops” (Jul 7)
Problem 2: Your team runs on the most expensive model with no routing
Fable 5, Claude’s top tier, moved to pay-per-use on July 12. Ruben Hassid gives the frame a lean team needs. Picture the top model as a senior lawyer at 1,000 dollars an hour and the cheaper Sonnet as a capable intern at 100 who still follows what the senior says. You pay the senior to set the strategy, then the intern drafts against it consistently. The failure mode is a whole team on the top model with zero training and nobody tracking usage, billing the senior rate for mundane work. Hassid’s own habit is to ask the top model 1 or 2 hard turns to set direction, then drop to a cheaper tier for the rest. Charlie Hills runs the same split inside his loop, top model plans and Sonnet executes, and Maja Voje’s design test found Claude Code wins partly because it edits in place instead of regenerating, which burns fewer tokens.
My take: now that Fable moved to pay-per-use, paying the senior rate for routine drafting is a leak a lean budget feels fast. I ask the top model 1 hard turn to set direction, then route the rest to the cheap tier, and the bill stays sane.
The European read: a lean European budget cannot subsidize the whole team running every task on the top model. Route strategy to the top tier and execution to the cheap one, and you get the same output at a fraction of the spend, which matters more as intelligence gets metered.
Do first: list your 3 highest-volume Claude tasks this week and mark each one strategy or execution, then route the execution work to a cheaper model.
Do this week: set a default in your team’s workflow: top model for the plan, cheap model for the drafting, and track usage so nobody is quietly billing the senior rate for intern work.
Do this month: review your model routing monthly, so a price change is a decision you make, not a surprise on the invoice.
The rule: top model sets the strategy, cheap model does the routine execution. Never pay the senior rate for intern work.
You know it worked when: your Claude bill drops and no output got worse, because only the strategy turns hit the top tier.
The play: 5 steps to cut your Claude bill 50 to 90% (AI-First series). Route strategy to the top model and execution to the cheap one, and track usage so the bill stops running away. Get full access below ↓
Sources: Ruben Hassid, “Fable 5.” (Jul 8) · Maja Voje and Anze Voje, “AI Design Tools vs Human Designers” (Jul 10)
Problem 3: Your AI content is polished, and nobody could tell it is yours
Every AI has read the same internet, so everyone’s first draft sounds the same. Kieran Flanagan gave away a Claude Skill this week that acts as a pre-publish coach rather than a rewriter. It runs 2 gates. The Share Test, would your ideal reader forward this. The Onlyness Test, could only you have written it. It spars with you until your differentiated take shows up, instead of smoothing the draft into generic prose. Maja Voje’s team reached the same place from the design side: across 6 AI tools, Claude Code won for production because it carries your brand context, but a human still owns the final taste pass on anything meant to last. Ruben Dominguez names why it compounds, you cannot improve the model, so the moat is the system you build around it, your data, your memory, your standards.
My take: the taste pass is the one thing I will not hand over. Polished is the floor now, everyone has it. The operator take, the lived European read a competitor’s agent has never had, is the moat, and no model has my week.
The European read: across markets, the thing a competitor’s agent cannot copy is your lived operator take, in your voice, on the market you actually sell in. The taste gate protects the one part of the content that is genuinely yours as you scale it.
Do first: take your last AI-assisted post and ask 1 question of it, could only your company have written this, or could any competitor have shipped the same thing.
Do this week: add a taste gate to your content loop, the Share Test and the Onlyness Test, that spars with the draft until your differentiated take shows up.
Do this month: save your standards and your best work as the memory the system builds around, so the taste gate gets sharper every run.
The rule: if a competitor’s agent could have written it, it is not done. Add a taste gate that asks, could only we have said this.
You know it worked when: your published content carries a take only you could have, and readers forward it because of it.
The play: 5 steps to ship AI content only you could have written (2026 series). Add a Share Test and an Onlyness Test to your content loop, so polished stops being the finish line. Get full access below ↓
Sources: Kieran Flanagan, “I Built an AI Skill to Sharpen My Taste” (Jul 10) · Maja Voje and Anze Voje, “AI Design Tools vs Human Designers” (Jul 10)
Steal this move: wrap your highest-volume GTM output in a self-scoring loop
Play: 5 steps to build a measurable loop that scores its own output (2026 series)
Fixes: recurring GTM output that still passes through your own eyes as the only quality gate, so your time goes to re-reading instead of judgment.
Best for: a founder or operator at 0 to 10M+ ARR producing recurring output like outbound, competitor digests, or content by hand and checking it yourself.
Week 1, Define: pick 1 recurring output and write the 5-line rubric for a 90-plus version. Collect your 3 to 5 best past examples as the answer key.
Week 2, Build: in Claude Code, set the loop with a goal and a check so it drafts, scores against the rubric, and rewrites under the bar. Route drafting to a cheaper model and keep the top tier for the rubric and strategy.
Week 3, Validate: run it for a week against your manual version and tighten the rubric until you would ship without edits. Add a taste gate, could only we have said this.
Week 4, Operate: add a memory file the loop appends lessons to, and share the rubric with the team. Keep a human approval step on anything that goes external.
Template: a 5-line rubric scored 0 to 100, plus a folder of 3 to 5 best-in-class past examples as the answer key the loop grades against.
Paste this into your AI:
Here is 1 recurring GTM output I produce and 3 to 5 of my best past examples [paste]. Draft the 5-line rubric a stranger would use to score it 0 to 100, then set up a loop that drafts, scores against the rubric, and rewrites anything under 95. Keep the judge separate from the writer, and end with a taste-gate question that asks whether only we could have written it.
Full play, template and workbook inside GTMcraft OS. Reply to this email or DM me to get access.
Also on the radar
Ruben Dominguez, “Self-Evolving Agents” (Jul 10): the reason agents compound is a memory layer, which cuts tokens about 90% and latency about 91%. Give your top GTM agent a notes file it appends 1 lesson to after every run.
Maja Voje and Anze Voje, “AI Design Tools vs Human Designers” (Jul 10): their 6-tool design bake-off landed on a clean split, Claude Code for production, humans for the final pass. Run your next on-brand asset in Claude Code with your brand guide loaded.
Why this matters now
For a year the whole game was better prompts. This week 3 separate creators described the same next step: you stop being the one who prompts and become the one who writes the check. This builds on The GTM Engine from July 5, where the move was to put recurring GTM work on a schedule. The loop is the next layer: the scheduled work now scores itself and improves without you. And it lands the same week intelligence got metered, so routing the cheap execution away from the top model is no longer optional.
For a lean European team scaling toward 10M: write the check, route the model, and keep the taste pass. The edge is not access to the model, it is the loop, the routing, and the take you build around it.
What is your number 1 takeaway this week? Reply or send me a DM.
Go deeper with GTMcraft OS
Scale on plays, not prompts.
You do not need another newsletter, podcast or LinkedIn post. You need the system under it.
Read the plays. Every play, live and current, inside GTMcraft OS.
Download for your archive. Take the PDF or workbook offline.
Get the markdown. Drop the raw play into your own stack or AI tool.
Build your own. Use the GTMcraft Notion databases (100,000+ signals) to spin up a play for your exact motion.
The full Plays database is over 100 plays and growing every week.
Curious to see a sneak peek of GTMcraft OS, DM me or reply on this email.
See you inside GTMcraft OS, your GTM co-operator. Not another community.
Koen
Your (human) GTM Agent
The loop that scores its own output is exactly where most founder-led teams stop. They stand up the AI that drafts the outreach, then never build the scoring step, so quality drifts within a month and nobody notices until reply rates fall. The output was never the hard part. The loop that checks it is the actual build, and it's the part that gets skipped because it doesn't demo well.