How I work
Ten rules I keep, in the order the work happens, each one taken from my own notes, my own products or something that broke.
Start from trust
In January 2019 I saved a passage in my Apres notes about product differentiation. Its point was that competitors can copy a better, faster, cheaper feature almost as soon as you ship it, and the one thing they can’t replicate is the trust customers feel for you and your team. Two years later I put the same idea in one line: we believe that the future of AI begins and ends with trust.
For me trust is a set of commitments a customer can check. At Excellent they are five rules, written on the About page, and each one gives something up: the speed of letting a model approve work, a clean green dashboard, the chance to tidy the record, a copy of your data. A rule that costs nothing doesn’t tell anyone much.
“You can’t trust code that you did not totally create yourself.”
Ken Thompson, “Reflections on Trusting Trust,” 1984
Say it plainly
My 2019 pitch notes open with George Heilmeier’s first question: what are you trying to do, in absolutely no jargon? My answer that day was four words long: “Apres automates AI training.” The questions I couldn’t answer that plainly were the ones I didn’t understand yet.
I write in plain words and short sentences, and I say the number with its source. The writing rules I use ban em dashes, hype words and the sentence that knocks down a straw man before making its point, because they are the easiest tells of a sentence nobody thought about. When the subject is AI, that matters twice.
“Never use a long word where a short one will do.”
George Orwell, “Politics and the English Language,” 1946
Write down what done looks like
Every step I hand to an agent ends with two lines. “Done when” is a condition a stranger could check. “Evidence” says exactly what to keep. “Launch the site” can’t be checked. “The homepage returns 200 on the real domain over HTTPS” can.
I learned this by leaving it out. My 2019 hiring protocol asked what a new hire should have done after 30, 60 and 90 days, and I never wrote an answer. An agent will satisfy exactly what you write down. People deserve the same clarity about what they will be judged on.
“What are the mid-term and final ‘exams’ to check for success?”
George H. Heilmeier, the Heilmeier Catechism, DARPA
The worker doesn’t grade the work
In January 2019 Jorge, an engineer working with Apres, proposed checking our human labelers with a model trained on part of their work. He noted it would help most if it made different mistakes from theirs. At Excellent the first rule is that the agent does not grade itself. Its words are context. A different identity, and a different model family, has to approve.
That sounds like distrust. In practice it is what lets me hand agents real work at all. Between June 3 and November 2, 2026, 12,328 commits landed on main in Excellent’s repository, and I made 1,204 of them. I can only live with that ratio because the author never approves its own change.
“Given enough eyeballs, all bugs are shallow.”
Eric S. Raymond, The Cathedral and the Bazaar, 1999
Show the check that couldn’t run
In 2018 the sample job post in my sales book, for freelancers building lead lists, said that if you can’t find the email address, it is better to move on than to enter a false value. Excellent keeps the same rule for agent work: a check that cannot run stays visible, and it is never counted as a pass.
The failures that cost me most were quiet ones. Eight “completed” tasks that were really workers killed by a timeout. A recovered gate that printed a pass while reading an empty list. Each was a missing check that looked like a passing one.
“The first principle is that you must not fool yourself.”
Richard Feynman, “Cargo Cult Science,” Caltech commencement address, 1974
Only evidence carries a pass
The formula on Excellent’s About page is verified = claim + evidence. The agent’s “done” is the claim. The evidence comes from the repository, the tests, the build and the browser. A model can never carry a pass on its own. Only a check that ran, or a person, can.
Jorge’s 2019 validator worked the same way. The model could raise its hand, and people settled the disagreement. The same goes for numbers in a pitch: a figure without its source is a claim, however good it sounds.
“For a successful technology, reality must take precedence over public relations, for nature cannot be fooled.”
Richard Feynman, appendix to the Rogers Commission report on the Challenger accident, 1986
Keep the work where it was done
Excellent runs beside the agents you already use and keeps its evidence on the machine that produced it. The data is a single SQLite file on your own disk. We run no cloud that holds a copy.
That choice has a price. Evidence doesn’t sync between teammates yet, and a hosted product would have made that easy. I would rather pay it. A receipt you hold yourself is one you can inspect, keep and hand to anyone, years later.
“Data is a toxic asset.”
Bruce Schneier, “Data Is a Toxic Asset,” 2016
Never rewrite history
History is append-only. A re-run is a new receipt, and nothing is edited in place. The failure registry in my repository works the same way: when an entry is retired it moves to the bottom word for word, so its fingerprint stays searchable if the problem ever comes back.
In August 2026 we found ten tasks marked done whose work had never reached main. Nine of those commits were reachable from no branch or tag, one garbage collection from gone. We pinned them with 18 tags and kept every cancelled task’s record. Work you can’t find is work you will do twice.
“Those who cannot remember the past are condemned to repeat it.”
George Santayana, The Life of Reason, 1905
Write failures down, with the numbers
In July 2026 I asked for one central place to record every failure in my agent fleet, so we wouldn’t lose track. The registry now has 59 entries, each with a fingerprint, a root cause, a remedy and the dates it was seen.
The numbers are the point. “Agents are sometimes unreliable” teaches nothing. “Six of eight completions were workers killed at 60 seconds, with zero tokens” tells you what to check tomorrow. I publish these as field notes, because the failures are the most useful thing I have.
“When you can measure what you are speaking about, and express it in numbers, you know something about it.”
Lord Kelvin, lecture on electrical units of measurement, 1883
Give the useful parts away
The small tools I build along the way, screen-map, usertrial, selfdocs, pressroom and usage-cache, are MIT licensed and open source. Excellent itself is Fair Source, and I call it that. The prompts, checklists and rules on this site are free to copy.
My unfinished 2018 sales book offered a free conversation to anyone who didn’t want to read it. The rule I keep now is to give four times for every ask. If something here saves you an afternoon, that is the point.
“Information wants to be free.”
Stewart Brand, first Hackers Conference, 1984
Building something with agents? Say hello.