Pinned Tweet
The value frontier
undominated.ai/frontier/
AEO is two different jobs hiding in one acronym.
In customs it means Authorized Economic Operator. On the web it means answer engine optimization: pages written so an AI answer can quote them.
Show one AEO slide to a colleague. If they cannot say which job you mean, split the label before the next meeting.
lenvanderhof.com/en/blog/wha…
An AI agent's limits need enforcement outside the prompt.
Test: cap a sandbox run at two tool calls, then give it a task needing three. Check the log: the third call must be blocked. If it runs, the cap is only a suggestion.
lenvanderhof.com/en/blog/usi…
An AI handoff needs a brief, not a chat archive.
In CHORUS, that brief includes pass/fail criteria.
Test: give a colleague one brief and a sample output. If they cannot accept or reject it without asking you, rewrite the brief.
lenvanderhof.com/en/blog/the…
AI instructions need tests, just like code.
Test: pick one rule in AGENTS.md, the agent's instruction file. Run your checks on an output that breaks it. If they pass, that rule has no working test.
lenvanderhof.com/en/blog/the…
An AI demo shows capability. An agent charter sets authority.
Test: open one charter. Name an action that needs human approval and its approver. If either is missing, withhold production access.
lenvanderhof.com/en/blog/wha…
Every copied AI rule is another file that can miss the next update.
Test: count the files you must edit to change one rule for your coding agents. If the count is more than one, keep a single source and check that each agent loads it.
lenvanderhof.com/en/blog/sta…
Training a model on your policy does not keep its answers current.
Test: ask for today's refund window, then open the passage it cites. If that passage is old or missing, fix the lookup before you retrain.
lenvanderhof.com/en/blog/rag…
The first forecast said aloud pulls every later estimate toward it.
Test: pick one deadline. Each person writes the odds of hitting it before anyone speaks. Talk, estimate again, and name the new evidence behind each change.
lenvanderhof.com/en/blog/cal…
A stronger model does not get a pass on your data rules.
Test: take one prompt that includes customer or policy text. If that text must stay on machines you control, run it locally or block the call.
lenvanderhof.com/en/blog/fro…
Fluent AI prose can hide an assumption the conclusion needs.
CALIBRATE: name the assumption that would collapse the last answer you forwarded, then point to the evidence. No evidence: mark it unverified.
lenvanderhof.com/en/blog/eva…
AI Overviews can only use a page Google is allowed to quote.
Test: view source. If the robots meta tag contains nosnippet, Google may not quote the page, so AI Overviews cannot use it.
lenvanderhof.com/en/blog/how…
A model can win on score and price and still fail your job.
Capability envelope test: send your longest real input and the output length you need. If the other model cannot take either, it cannot replace yours for that job.
lenvanderhof.com/en/blog/wha…
A model on a fixed path is still automation.
Test: draw the path in five boxes or fewer. Circle any box that can pick a tool the drawing does not name. No circle: keep calling it automation.
lenvanderhof.com/en/blog/age…
Calibration is whether your stated odds match what actually happens.
Test: pull forecasts you called 80% likely. Did about 80% come true? No log yet? Write one forecast, its probability, and a review date today.
lenvanderhof.com/en/blog/wha…
A task you can decide with fixed rules does not need an AI agent.
Test: write one recurring task and every exception. If each decision is a rule you can state before it runs, start with a script.
lenvanderhof.com/en/blog/whe…
Len P. van der Hof - DEEP TRACE // 180+ BPM Subterranean Industrial Hard... youtu.be/RuddV8vePwA?si=CbaA… via @YouTube
A link to your page is not the same as being named as the source.
Test: ask one question your page answers in three fresh ChatGPT Search chats. For each reply, score two things apart: was the page linked, and were you named?
lenvanderhof.com/en/blog/cit…
An AI agent drafts a pitch. A literary agent is hired to place your book with publishers.
Test: write the job in one sentence. If it needs a publisher relationship, hire a literary agent. If it is a draft, use software and leave sending off.
lenvanderhof.com/en/blog/lit…
