WS 01Pitched
Pitched
Angles filed by the strategy & scan runs, waiting on approval.
A 7B model with a task graph just matched GPT-4 on agent tasks. My writing desk's quality jumps all came from architecture, not model swaps.
HeavyAI slop is an inputs problem
This kernel explainer earned 1,300 bookmarks with zero hype: every claim ends in a check you can run, profile it and count the rows.
HeavyAI slop is an inputs problem
levelsio built a bot with exactly one user, him, and that is the right first product: the single-user version finds the real failures.
HeavyBuild the personal tool first
GPT-5.6 deleted files in full-access mode. The fix isn't a safer setting, it's a guard outside the agent it can't override.
HeavyA refusal gate beats an editing pass
Same origami skill, same prompt, Fable 5 vs GPT-5.6, 32 min each. Only one felt like a film. That gap is the taste no prompt reaches.
HeavyName the residue you cannot reach
Slop being a human taste problem is right, but 'have taste' isn't executable. A gate that won't draft without real inputs kills more of it.
HeavyA refusal gate beats an editing pass
I built a 141-test evaluator to gate my client deliverables, and it has failed drafts written by my own AI ghostwriter.
Heavy▚ evaluator-gated-ai-deliveryA refusal gate beats an editing pass
Volume is not the constraint. My desk ran daily and still repeated one thesis in 4 of 8 posts, so now a 30-day gate blocks the repeat.
HeavyAI slop is an inputs problem
Your feed shapes your taste only if you write it down. My desk keeps a file of things I refuse to say, and drafts fail QA against it.
HeavyAI slop is an inputs problem
Run computer-use agents on your own accounts first: a 2FA migration teaches you where the agent breaks while the stakes are only yours.
HeavyBuild the personal tool first
Top Stories just moved inside AI Overviews. Google is rebuilding the whole SERP inside the answer, and that's where the leverage went.
StandardTechnical correctness is table stakestests H-2026-07-20-1
AI Mode now adds groceries to a cart mid-search. The store's next fight isn't the click, it's being the catalog the AI picks from.
StandardNO-STANCE
Want AI to write like you? Skip the writing samples. My desk learns from the emails I actually sent, pulled daily from my outbox.
HeavyAI slop is an inputs problemtests H-2026-07-05-1
Google now lets you opt out of AI answers without ranking risk. For most brands the safe-looking switch is self-erasure.
StandardTechnical correctness is table stakes
He's right, from the non-engineer side: I aimed agents at SEO I already know and got a working autoposting desk. The domain was the input.
HeavyBuild the personal tool first
My desk kept shipping em dashes after being told not to, so a grep now blocks the post; every named AI tic needs its executable check.
HeavyA refusal gate beats an editing pass
A 296KB enterprise page passed every technical audit, but its key content sat behind a ZIP prompt a crawler never clicks.
Heavy▚ the-296kb-page-that-passed-every-auditTechnical correctness is table stakes
The usual advice for showing up in ChatGPT (listicles, Reddit) is cargo cult; the bucket it files your source under is the real lever.
HeavyTechnical correctness is table stakes
AI didn't make non-engineers obsolete, it made us builders. I directed a writing desk I couldn't have coded, and it now runs unattended.
HeavyBuild the personal tool first
Notion as code is right about the shape and dangerous with live rows: a git stash once reverted my desk's running queue file mid-job.
HeavyBuild the personal tool first
Lightreel scores whether a video's shape travels. It can't tell you if you had a reason to post, which is where my drafts actually die.
StandardAI slop is an inputs problem
The desk runs unattended on a schedule, so it stamps a heartbeat after every run and the alarm fires when one goes stale.
Heavy▚ automation-that-survives-unattendedBuild the personal tool first
Anthropic cut 80% of Claude Code's prompt to trust a better model. The rules you can't lose go the other way: out of the prompt, into code.
HeavyA refusal gate beats an editing pass
Graph the agents however you like, but one critic node must be code: LLM judges passed em dashes, the grep that blocks the post never does.
HeavyA refusal gate beats an editing pass
Paste the reference, don't describe it: my mockups got good when I handed the agent real brand colors and type instead of adjectives.
HeavyAI slop is an inputs problem
ChatGPT drops 29% of its referrals on a site's own search box, not the page that answers. Fix the landing, not the ranking.
StandardTechnical correctness is table stakestests H-2026-07-05-1
An unattended agent's most likely outage is a bill, not a bug. Mine lost two weeks to an X API spend cap that threw no error.
Heavy▚ Automation that survives unattendedBuild the personal tool firsttests H-2026-07-05-1
I approve my agent's posts by replying 'post 5043' to its daily email. Unattended does not mean unsupervised; the inbox is the interface.
Heavy▚ Evaluator-gated AI deliveryName the residue you cannot reachtests H-2026-07-05-1
My AI-slop gate caught 14 multi-word phrases and missed the 19 single words that actually give a draft away. Those shipped for months.
Standard▚ Agentic content systemsAI slop is an inputs problemtests H-2026-07-05-1
A second model trimming the first one's 300 extra lines is a humanizer pass for code: cutting is easy, the missing brief is the actual bug.
HeavyA refusal gate beats an editing pass
Buzz is where my agents post events, not where the project lives: MEMORY.md and git stay canonical, so a lost bus never loses the work.
StandardName the residue you cannot reach
Asking for plain writing failed, so a gate now counts sentence length: 20 words warns, 25 fails. Take STE's rules, never its dictionary.
HeavyA refusal gate beats an editing pass
Every section on that $5K audit list is table stakes now. The part a buyer pays for is whether AI answers name them.
StandardTechnical correctness is table stakes
A design skill pulls a record shop's colors and type off its live site first, so the mockup it builds already looks like that shop.
Heavy▚ ai-design-for-niche-commerceAI slop is an inputs problem
Scrape every competitor hook from 90 days and feed it to Claude, and you get the average of your rivals' feed. Slop starts at the inputs.
HeavyAI slop is an inputs problem
last30days started as one builder's own research tool and just passed 55k stars. The single-user version is the seed of the product.
HeavyBuild the personal tool first
Peter's right that nobody reviews agent output forever. My desk shipped em dashes after being told no, so now a grep blocks the post.
HeavyA refusal gate beats an editing pass
Reverse prompting isn't a trick. The model has the smarts; it lacks your inputs. My writing desk interviews me weekly for the same reason.
HeavyAI slop is an inputs problem
Anthropic published its own model breaking into real systems during evals. Naming your failures is what makes safety claims believable.
HeavyName the residue you cannot reach
My no-em-dash rule sat in the agent's instructions and five still shipped on one page, so now a grep checks every draft before it posts.
Heavy▚ five-em-dashes-shipped-anywayA refusal gate beats an editing pass
TrustMRR's July: 210K human visitors, ~1M AI bot crawls. The machines that decide who gets cited read the site 5x more than people do.
StandardTechnical correctness is table stakes