Ship 30 for 30: week 2 drafts (Days 10+, from Mon Oct 19)
Ray drafts, Ben edits and posts. Each piece is an Atomic Essay, 300 words or fewer. Word counts include the headline and exclude the [BEN] and [CHECK] notes. No hashtags. Week 1 (Days 1-9) lives in [[2026-10-07-week1-drafts]]. Days 10-12 continue The Fifth Job spine in the topic-bank order.
| Day | Date | Headline | Ship 30 rule | Words |
|---|---|---|---|---|
| 10 | Mon Oct 19 | How creative agency owners can stop selling hours and sell judgment once agents carry production | Headline + audience + outcome | 285 |
| 11 | Tue Oct 20 | How physical-product owners can give agents the back office, cut admin hours, and protect quality and repeat customers | Headline + Twist the Knife (stacked outcomes) | 274 |
| 12 | Wed Oct 21 | Most people think a one-person business can't run on agents. The data says some could. | 5 headline styles, ship one (Opinion) | 295 |
| 13 | Thu Oct 22 | A model grading itself is not a measurement | X hook: a 3-second preview, not the headline repeated | 281 |
| 14 | Fri Oct 23 | I passed two AI certs by studying my wrong answers | LinkedIn zero-to-hero hook | 297 |
| 15 | Sat Oct 24 | The 4 checks a business must pass before I name it | Prep the Page | 266 |
| 16 | Sun Oct 25 | The docs changed while I was studying | Single-sentence openers (moment in time) | 234 |
Day 10, Mon Oct 19
Rule: Headline + audience + outcome (live session: LinkedIn Carousels; 2026 rules not yet published, so this stays an Atomic Essay). Source: [[agentify-3000-series]] spine #4 (scoped to marketing, design and film agencies per its CHECK note; no advice to consultancies; agents lift people up kept in frame); [[foils]] §Agency and productized service; [[archetypes]] §Per archetype. "Agency" in the headline narrowed to "creative agency" to honor the spine's scope.
How creative agency owners can stop selling hours and sell judgment once agents carry production
If you run a marketing, design or film agency, your biggest cost is people. As agents carry more of the production, the hours you bill shrink. What's left to sell is the judgment clients were paying for all along. [BEN: confirm this framing is yours]
I had an artificial intelligence (AI) model score 2,500 businesses from Starter Story (starterstory.com), a site of founder interviews, on five jobs: find customers, sell, deliver, support, and admin. [BEN: confirm you ran this scoring] Each job got a score from 0 to 3 for how much of it an agent can do today. The model scored each business from a short public blurb, so treat this as a rough first sort.
Agencies were the odd result:
- The model judged that agents could carry about half their work (48%).
- Delivery scored 1.5 of 3.
- Median revenue per person was the lowest of any business-to-business type, about $5,300 a month.
Two small agencies in the data show where the score stops. A two-person creative agency that already runs on its own AI software scored lowest on taste and the client relationship. A five-person AI film studio scored lowest on direction. Its directors are the product.
Here's the split I'd try:
- List what clients pay you for: the brief, the taste, the call on what ships.
- List where your team's hours go: campaign drafts, content, reports.
- Hand the second list to agents first. Put the hours you get back into the first.
One caveat: some clients pay for the account manager, not the output. For them, cheaper production may not change what they buy.
If you run a creative agency, what do your clients actually pay you for?
Day 11, Tue Oct 20
Rule: Headline + "Twist the Knife": stack the extra outcomes (cut admin hours, protect quality, keep repeat customers). Source: [[agentify-3000-series]] spine #5; [[archetypes]] §Per archetype (figures re-derived from data/businesses-scored.csv on 2026-10-20); [[foils]] §DTC physical product brand. Headline shortened from the topic-bank row. Critic: ITERATE x1, all seven fixes applied (triplet "three wins" line cut, "business around it" narrowed to "back office", one term for repeat customers, ops archetype relabelled, revenue marked self-reported).
How physical-product owners can give agents the back office, cut admin hours, and protect quality and repeat customers
If you sell a physical product, let artificial intelligence (AI) agents run the back office around it. Keep your people on the product. [BEN: confirm this framing is yours]
I had an AI model score 2,500 businesses from Starter Story (starterstory.com), a site of founder interviews, on five jobs: find customers, sell, deliver, support, and admin. [BEN: confirm you ran this scoring] Each job got a score from 0 to 3 for how much of it an agent can do today. The model scored each business from a short public blurb, so treat this as a rough first sort.
For physical goods, the gap was wide:
- Admin scored 2 or more, meaning the agent does most of it, for 417 of 425 brands that sell their own goods online. [CHECK: dtc_brand job_admin_score >= 2, 417 of 425 clean rows, businesses-scored.csv]
- Delivery, which here means making and shipping the product, averaged 0.8 for those brands. Manufacturing, logistics and supply businesses averaged 0.4. Handmade goods averaged 0.1. [CHECK: re-derived 2026-10-20 from businesses-scored.csv, 2,500 clean rows: dtc_brand deliver 0.78 (n=425), ops 0.38 (n=85), maker 0.13 (n=79)]
In one two-person skincare brand in the data, the founder is the chemist. No agent mixes the next batch.
Here's the split I'd try:
- Give agents the back office: invoices, inventory reports, order emails.
- Count the hours you get back each week.
- Spend them on the product: quality checks, the next design, the repeat customers. [BEN: does this match what you've seen?]
The product isn't always where the money is. Median self-reported monthly revenue per team member was about $13,000 for those online brands and about $3,000 for handmade goods. [CHECK: archetypes.md per-archetype table, median $/mo per person: dtc_brand 13.0k, maker 3.0k; Starter Story header figures] Protecting delivery doesn't make a small workshop pay.
If you sell something physical, how many hours a week does the back office take from the product?
Day 12, Wed Oct 21
Rule: 5 headline styles (List / Story / Opinion / Question / Framework), ship one; this ships the Opinion style from the topic-bank row, spares in the Sources comment. Source: [[agentify-3000-series]] spine #6 (next in topic-bank order after Day 11); [[foils]] §Courses, info products and communities and §Content, media and audience; every figure re-derived from data/businesses-scored.csv on 2026-10-21. Built on the five-of-five count, not on admin: per [[2026-10-21-agent-shaped-job-taxonomy-grounding]], the admin score may be an aggregation artifact and there is no build job, so the essay says the build gap out loud. Critic: ITERATE x1, all four fixes applied (headline overclaim, % bases in-text, one-meaning 'job', unsupported setup-gap line).
Most people think a one-person business can't run on agents. The data says some could.
You don't need a team before artificial intelligence (AI) agents can carry most of a business. Some one- and two-person businesses already score that way. [BEN: confirm this framing is yours]
I had an AI model score 2,500 businesses from Starter Story (starterstory.com), a site of founder interviews, on five jobs: find customers, sell, deliver, support, and admin. [BEN: confirm you ran this scoring] Each job got a score from 0 to 3 for how much of it an agent can do today. The model scored each business from a short public blurb, so treat this as a rough first sort.
What the sort found:
- 279 businesses with one or two people scored 2 or more, meaning the agent does most of it, on all five jobs. [CHECK: clean rows (revenue_suspect False) = 2,500; all five job scores >= 2 AND team_size <= 2 = 279 (136 of them solo). Drop admin and require only the other four: 281, so the count does not rest on the admin score]
- That was 36% of the 783 small teams. Bigger teams scored that way 35% of the time (547 of 1,560). Size didn't decide it. [CHECK: team_size known for 2,343 clean rows; 279 / 783 teams of 1-2 = 35.6%; 547 / 1,560 teams of 3+ = 35.1%]
Three two-person examples in the data:
- An online piano course. The recorded method runs with little upkeep.
- An exam-prep site. What's left is the founders' exam know-how.
- A recipe blog. What's left is recipes that taste good.
Each still has one thing only the founder brings.
The scores don't say whether owners can set agents up. [BEN: does this match what you've seen?]
Here's the test I'd try:
- Score your five jobs from 0 to 3.
- Circle every job at 2 or more.
- Hand one circled job to an agent this month.
Two caveats. Starter Story publishes winners. And the sort has no job for building the product. For the 165 software businesses among the 279, that job was never scored. [CHECK: archetype saas 114 + consumer_app 51 = 165 of the 279]
If you run a business with one or two people, which of your five jobs already scores a 2?
Day 13, Thu Oct 22
Rule: X hook, a 3-second preview, not the headline repeated (live session: X Articles; 2026 rules not yet published, so this stays an Atomic Essay and the hook is the two lines above the LinkedIn "see more" cut). The hook teases the hand check. Source: topic-bank-v1 Day 13 row; [[agentify-3000-series]] spine #1 counterpoint ("the scores are a model grading itself from a blurb"); agentify-3000/IMPLEMENTATION-NOTES.md §Spot-check; every figure re-derived from data/businesses-scored.csv on 2026-10-22. The spot check was run by the AI setup that built the scores, so the essay says so instead of claiming an independent read. "About 2,500": 2,566 scored, 66 revenue-suspect rows dropped. Critic: ITERATE x1, fixes applied (hook corrected to 28 of 30, "scores" vs "businesses", hand-check bar set above the always-2 baseline, signpost cut, self-grading line narrowed to the second read, "near 300 characters"); round 2 PASS (optional "20 of 25 exact" cutoff for "clearly" declined: unsourced threshold).
A model grading itself is not a measurement
All 30 businesses passed the spot check. A reader who answered "2" to every job would have passed 28. [CHECK: spotcheck_ids.json, 30 businesses; 28 have no 0 score, so an always-2 read is within one point on every job; Honeydew and Photobooth Supply Co. have deliver 0 (two points off)] [BEN: confirm this framing is yours]
I had an artificial intelligence (AI) model score about 2,500 businesses from Starter Story (starterstory.com), a site of founder interviews, on five jobs: find customers, sell, deliver, support, and admin. [BEN: confirm you ran this scoring] Each job got a score from 0 to 3 for how much of it an agent can do today. The model scored each business from a short public blurb.
Then the same AI setup re-read 30 random businesses. Every job score landed within one point of the second read. [CHECK: IMPLEMENTATION-NOTES.md §Spot-check: 30 random, seed 2026, "30/30 within one point of my own read on every job", judged by the producer, not an independent reader; all 30 ids in spotcheck_ids.json present in businesses-scored.csv]
On a 0-3 scale, one point is wide:
- Of the 12,500 job scores, 69% are a 2. [CHECK: clean rows (revenue_suspect False) = 2,500 x 5 jobs = 12,500 scores; score 2 = 8,604; 8,604 / 12,500 = 68.8%]
- Answer 2 for every job and you land within one point 96% of the time. [CHECK: scores of 1, 2 or 3 = 12,020 of 12,500 = 96.2%; zeros = 480]
- Most blurbs (2,405 of 2,500) stop mid-sentence near 300 characters. One was mostly a founder calling himself driven. The model gave four of its five jobs a 3. [CHECK: description ending "..." = 2,405 of 2,500 clean rows; max length 299, median 290; the example is a 13-person comparison site, scores 3/3/3/2/3, 290-character blurb]
So "within one point" mostly measures how often the model says 2. And the second read came from the same setup as the first.
Here's the hand check I'd run:
- Pick five businesses at random.
- Read the full story, not the blurb.
- Score each job yourself before you look.
- Count exact matches, not near misses.
Answering 2 every time already gets about 7 in 10 exact. Beat that clearly and the scores measure something. If not, they're a rough first sort. [CHECK: always-2 exact = 8,604 / 12,500 = 68.8%] [BEN: have you run this hand check yourself? If yes, add your match count]
One caveat: your read is one person's read. Two people scoring blind beats one.
If you use AI scores at work, who checks them, and against what?
Day 14, Fri Oct 23
Rule: LinkedIn hook in zero-to-hero form ("A few years ago, I {problem}. Today, I am {accomplishment}. How? By doing {something unexpected}."). Source: topic-bank-v1 Day 14 row; [[cert-progress|01-projects/phdata/cert-progress.md]] (study log, attempt log, 2026-09-23..28 quiz entries); [[founder-exam-experience-2026-09-28|certifications/snowpro-genai-c02/founder-exam-experience-2026-09-28.md]]; memory feedback_study_quiz_style.md. Study method only, no exam content (certification agreements): no exam questions, answers or topics, and the Snowflake format detail stays out. The bank's angle ("he used to cram") is not in the vault and runs against it: he crammed on exam day for both certs, so the cram stays in as a targeted cram and the "a few years ago" problem is a [BEN] slot. No phData internals, pay or bonus. Critic: fresh sw-critic r1 ITERATE (hook contradiction "instead of"→"more than"; miss sort credited to Ben → agent; "answer key" scoped to quiz; 2 [BEN] tags added; headline reframed), 8 fixes applied by parent, 297 words.
I passed two AI certs by studying my wrong answers
A few years ago, I [BEN: your real old study problem, one line. Suggested: "studied for a cert by rereading the course until it felt familiar"]. Today, I hold two artificial intelligence (AI) certifications, from Anthropic and Snowflake, passed six weeks apart. How? By studying my wrong answers more than the material. [BEN: confirm this framing is yours]
For Claude Certified Architect (Foundations), I took a practice exam cold, before the prep course: 50 of 60. My second dropped to 42. [CHECK: [[cert-progress]] attempt log, 2026-06-14 50/60 cold pre-course, 2026-06-27 42/60]
I had my AI agent sort all 18 misses. About six weren't gaps in what I knew. They were a habit: I picked the clean answer when the question asked for the most complete one. [CHECK: cert-progress.md "Miss pattern (practice exam 2)", thread 1, test-craft ~6 Qs of 18]
On exam day I crammed, but only the ranked list of misses. I skipped what the exam guide marks out of scope. I passed with 850 of 1000. [BEN: confirm the exam-day block is how you remember it]
For SnowPro Specialty: Gen AI, my AI agent quizzed me by text, one question at a time. [BEN: confirm "my AI agent" is how you want to name it] Three rules came out of it:
- Rotate the answer letter. After five straight B's, I told it, "Let's make sure not all the answers are B."
- Put a doc link under every explanation.
- Check each quiz answer against the current docs. Some were out of date.
I also finished a video course. The quiz helped more. I passed with 854. [BEN: confirm "the quiz helped more" in your words; cert-progress.md 2026-09-28 13:32 is Ray's paraphrase]
Here's the loop I'd try: [BEN: confirm this is the loop you'd recommend; step 1 is Ray's 2026-06-13 diagnostic-first advice that you followed, not your own stated method]
- Take a practice test before you study.
- Sort every miss: knowledge gap or habit.
- Study the misses, and check each answer against the current docs.
One caveat: the Snowflake exam was "much harder and different than what I studied for." The quizzes found my gaps. They didn't predict the test.
I'm running the same loop for a third cert, Claude Certified Developer (Foundations), by November 10. [BEN: confirm you're using this loop for CCDV-F]
The last time you studied for something, did you study the material or your misses?
Day 15, Sat Oct 24
Rule: Prep the Page: headline, credibility line, main points and takeaway decided before drafting. Headline: the 4 checks. Credibility line: the 2,566-business scoring. Main points: the four gate steps, each shown by a real conflict. Takeaway: describe until it passes, plus the same-page caveat. Source: topic-bank-v1 Day 15 row; [[agentify-3000-series]] "Gate for every foil"; [[foils]] gate line and conflict list; agentify-3000/IMPLEMENTATION-NOTES.md (2,566 scored, header fields). Headline changed from "The 4-step check I run" because the gate has not been run on a named foil yet (Days 10-13 dropped names for that reason). Conflict-list businesses stay unnamed in the body, per the series rule. Critic: fresh sw-critic r1 ITERATE (11 fixes: 2,566 not 2,500 scored, header lists employees, "owns" not "runs", headline term match, signposts cut); r2 ITERATE (6 fixes: [BEN]-free shortlist line, check 1 got its own example (Pep AI), headline vs story title disambiguated, what was scored, suspect-revenue paragraph cut as off-thesis); all applied by parent, 2-round cap reached. One r2 fix narrowed: the "opening claim" gloss on the blurb was dropped (foils.md uses it for the post, not the blurb).
The 4 checks a business must pass before I name it
A revenue figure in a founder interview is the founder's own claim. Before I name a business in a post, it has to pass four checks against its full public story. [BEN: confirm this framing is yours]
I had an artificial intelligence (AI) model score how much of each business's work an AI agent could carry. It scored 2,566 businesses from Starter Story (starterstory.com), a site of founder interviews. [BEN: confirm you ran this scoring] It read only a public blurb of under 300 characters. Each story page also opens with a header box that lists revenue, founders and employees.
My AI agent shortlisted 30 businesses to name. Five of their pages disagree with themselves. [CHECK: foils.md, 30 foils; conflict list, 5 entries] Each check caught at least one.
- Read the full story, past the header box. One header said $30,000 a month. The story body said the app was five weeks old. [CHECK: foils.md Pep AI, header $30k/mo, five weeks old at publication]
- Confirm the founder still owns it. A second business was sold to another company in April 2026. [CHECK: foils.md StandOut CV, acquired April 2026]
- Check that the revenue agrees. That sold business's header said $360,000 a month. Its story body said $40,000 a month at the time of sale. One story title said $6,000 a month. Its header said $23,000. [CHECK: foils.md StandOut CV header $360k/mo vs $40k MRR at sale; Excel Exercises headline $6k/mo vs header $23k/mo]
- Check the team size in the header. One story title said "one-man." The header said two people. [CHECK: foils.md BTNG]
Until a business passes all four, I describe it and leave out its name. [BEN: confirm this is your rule; Days 10-13 dropped names because the gate had not been run]
All four checks trust the same page. If the founder's numbers are wrong everywhere, the story agrees with itself and still misleads.
Before you cite an example at work, what do you check first?
Day 16, Sun Oct 25
Rule: Single-sentence openers (6 types), one at the start of every section; opener type "moment in time" (each section opens on a date or a point in time). Source: topic-bank-v1 Day 16 row; [[2026-10-02-post3-snowpro-genai-cortex-v1]] "What changed since I studied" and its Verified product claims table (rows 11, 12); [[cert-progress|01-projects/phdata/cert-progress.md]] (study start 2026-09-16, study guide re-sent 2026-09-23 byte-identical to the 05-18 file, exam 2026-09-28). Study method only, no exam content (certification agreements): the two doc changes cited are public release notes, and the post says nothing about whether the exam covers them. Quiz items, weak areas and the "exam key vs current docs" notes in cert-progress.md stay out because they map to tested topics. Release-note URLs re-checked 2026-10-25 (HTTP 200). Critic: sw-critic ITERATE → 4 exact-text fixes applied by parent (start-date [BEN], Gateway timing, gap-list opener, caveat pair).
The docs changed while I was studying
Five days into studying for a Snowflake certification, Snowflake shipped a feature my study guide couldn't know about. [BEN: confirm this framing is yours]
On September 16, I started studying for SnowPro Specialty: Gen AI. [BEN: vault logs Sep 16 as a materials question and Sep 23 as "back to studying"; confirm the start date, or move it and recheck "Five days"] I used the official study guide, a video course, and my artificial intelligence (AI) agent quizzing me by text. [CHECK: cert-progress.md 2026-09-16 15:10 first study action; 2026-09-28 13:32 video course finished; quiz rows 2026-09-23..28]
On September 21, Snowflake put Cortex AI Function Evaluation into public preview. It measures the output quality, cost and token use of AI functions against labeled datasets. The day before I started, Cortex AI Gateway had entered preview: one endpoint for model traffic, with per-user quotas. [CHECK: docs.snowflake.com/en/release-notes/2026/other/2026-09-21-ai-function-evaluation-preview and .../2026-09-15-cortex-ai-gateway; post3 Verified table rows 11-12]
Back in May, I had downloaded the official study guide. The copy I studied from in September was the same file. [CHECK: cert-progress.md 2026-09-23 13:46, re-sent PDF byte-identical (sha1 0ef1aea7) to the 05-18 copy] [BEN: confirm you studied from that guide, not a newer one] Study material is a snapshot. The product keeps shipping.
Next time, I'll open the release notes before the study guide. [BEN: confirm this is what you'd do next time] Here's the loop:
- Note the date on every piece of study material.
- Read the current docs for each section of the guide, plus the release notes since that date.
- Keep a gap list: what the material says, what the docs say today, and when it changed.
After the exam, the gap list is the part I'd keep. [BEN: confirm] At work, I support the product as it ships today. [BEN: did your Anthropic certification prep in August have the same gap? Add one line if yes, or cut]
On exam day, a gap list says nothing about what the exam covers.
When was your study material last updated, and what has shipped since?