Field Report 01 · Saturday 2026-11-21Filed 13:10 EST
Over-engineering an Etsy shop with a crew of agents
A friend launched an Etsy shop by texting a consumer AI app. J answered with a server, a crew of agents, a board charter, and a blind replication—and still has zero listings.
±The scoreboard, so far
Adam
consumer agent app on his phone
0public Etsy sales
- 1live shop
- 66listings
- ~$30spent
J
custom multi-agent setup on his own server
0public Etsy sales
- 0listings
- 1operating charter
- 2independent research passes
- 65scored niches
- 12primary-source fact-checks
- ·this website
Neither of us is turning into an e-commerce mogul. That's kind of the point.
§0The day, 11:57 → 13:10 EST
IThe short version
- Adam let an agent app called Instinct spin up an Etsy shop by text and floated a friendly contest. J dumped the screenshots into his own agent setup.
- The agent's first pitch leaned on J's taste in punk tees. J killed it: his edge isn't shirt design, it's managing AI.
- The agent wrote a board-style operating charter: contribution profit as the north star, staged budget gates, decision rights, and a shadow cost for the agent's own compute.
- J spotted the obvious hole—no marketing. Distribution became seven workstreams and a scored research criterion.
- Pass one (one Claude Sonnet 5.5 subagent, ~17 minutes) picked digital party-game printables and surfaced an Etsy rule requiring individual human sign-off on every listing.
- Pass two (a blind 15-minute sprint with three subagents) independently landed on digital wedding and party printables—despite a silent config swap that ran the subagents on a cheaper model, thin samples (10 listings per search), and about three-quarters judgment calls.
- Current status: Still no shop. Round 2 research, contest rules, and actually opening the store are all still open.
This is a field report from one Saturday—November 21, 2026, roughly 11:57 to 13:10 EST—comparing bleeding-edge agent orchestration against normie consumer AI, and showing what happens when you run agents like a small team. Start with the story, skim the Top 10 and ledger, or inspect the raw receipts in the file room.
IITop 10 niches · est. net per salefull table
- 01Wedding invitation templates (editable Canva/Corjl)Digital · confidence Med-High~$15.2
- 02Bachelorette party games bundleDigital · confidence Med-High~$6.2
- 03Wedding seating chart templateDigital · confidence Medium~$5.6
- 04Football/"big game" party games printable (no NFL marks)Digital · confidence Medium~$3.8
- 05Resume template (Canva/Word, ATS)Digital · confidence Medium~$4.3
- 062027 digital planner (GoodNotes)Digital · confidence Medium~$6.3
- 07Valentine's printable cards (classroom/kids)Digital · confidence Medium~$1.6
- 08Wedding budget spreadsheet / planner bundleDigital · confidence Low~$10.4 (UNVERIFIED price)
- 09Baby shower gamesDigital · confidence Medium~$1.6
- 10Personalized dog-mom mug (first physical/POD alternate)POD · confidence LowUNVERIFIED
IIIContents
Part one — the day
- §1The betA friend launches an Etsy shop by texting a phone app, floats a friendly contest, and J hands the screenshots to his own server.11:57 EST
- §2The wrong first answerThe agent's first pitch was smart, well-sourced, and built around the exact thing J didn't want to sell: his own taste.12:00 EST
- §3The redirectJ steps back. The edge isn't his taste—it's managing AI. The agent rewrites the job as a board-level operating charter.12:03 EST
- §4The blind spotThe charter had governance, gates, and a cockpit—and zero plan for getting a single customer. J caught it, and the agent found nine more holes.12:11 EST
- §5Pass oneOne Sonnet subagent spends seventeen minutes scraping Etsy search pages and returns with a pick, a list of blind spots, and a policy quote that kills the hands-off dream.12:05–12:23 EST
- §6Start over, on purposeWith a solid answer in hand, J orders a blind replication—and a brief engineered not to hint at the answer.12:56 EST
- §7Fifteen minutesA fresh lead writes its rubric, dispatches three children, catches a margin error, and ships a Top 10 eight minutes early—while its subagents silently run on the wrong model.12:58:33–13:05:37 EST
- §8Where two passes agree, and what's still unknownTwo blind passes, different models, different rubrics, one broad answer—and a clear-eyed look at how little that actually proves.13:09 EST
- §9Why the work is the outputStill no shop. What came out of one Saturday hour is a repeatable playbook for steering agents—and a reminder that none of us are becoming Etsy moguls.19:25 EST
Part two — the research
- AThe methodHow a niche gets found and scored/method/
- BThe ledgerAll fifty candidates, sortable, re-weightable/ledger/
- CThe landscapeResearch question one/landscape/
- DThe channelsResearch question two/channels/
- EThe Top 10The deliverable/top10/
- FTwo passesPass 1 and pass 2, side by side/two-passes/
- GUnverifiedWhat we couldn't check, and what we corrected/unverified/
- HThe file roomEvery raw artifact, downloadable/artifacts/
- IColophonHow this was made/colophon/