Foreword
This handbook was private for a long time—an internal document we used to train key staff, one person at a time. I’m now opening it up to everyone.
The principles here were written between 2009 and 2017, while we grew OnDemandWorld organically with a small team building mobile apps. Some of it is a product of that particular time and team. But a lot has held up, and many of these ideas live on in the company’s current agentic form, now odw.ai.
Sharing it is part of a simple belief: in an era of agentic AI, the advantage was never the document—it’s the people and the practice. So take what’s relevant, adapt what isn’t, and ignore the rest. That’s exactly what we did.
Recruitment Principles
We hire people who have the right attitude and the right talent to join our team. Recruitment is a mutual agreement: an interview is as much about the candidate evaluating us as it is about us evaluating them. Our job is to determine whether a candidate fits what the team needs and to be honest about whether we can offer what the candidate is looking for. In a market where strong engineers and operators have many options, treating the process as a two-way evaluation isn’t just polite—it’s how you actually close the people you want.
In an AI-driven environment, this principle takes on added weight. The baseline expectations for what a single person can produce have shifted dramatically. A capable individual augmented by GenAI tools can now ship work that previously required a small team. This raises the bar on judgment and lowers the premium on raw output, which changes what “fit” means. We are no longer primarily hiring for the ability to produce code, copy, or analysis. We are hiring for the ability to direct, evaluate, and take responsibility for work that is increasingly produced in collaboration with AI systems.
What We Screen For: Attitude
The single trait I look for above all others is a genuine orientation toward recursive self-improvement—someone who reflexively gets better at getting better. They don’t just learn a tool; they build a system for learning tools. They don’t just fix a mistake; they adjust the process that produced it.
This trait has become the most important predictor of success in an AI-driven workplace, because the tools, workflows, and best practices are changing on a quarterly basis. The specific GenAI model someone mastered six months ago may be obsolete today. What endures is the meta-skill of rapid, self-directed adaptation.
Concrete signals of this attitude:
- AI fluency through curiosity, not credentials. Look for candidates who already use AI tools in their own workflow and can speak specifically about where the tools help, where they fail, and how they’ve adjusted. Someone who says “I use ChatGPT sometimes” is different from someone who can tell you which tasks they’ve stopped doing manually and why.
- Comfort with being wrong fast. Recursive self-improvers treat errors as information. In interviews, ask about a recent failure and listen for whether they describe a changed process afterward, not just a lesson “learned.”
- Taste and judgment over execution speed. When AI can generate ten plausible options in seconds, the differentiating skill is knowing which one is actually good. Probe for the ability to critique and discriminate, not just produce.
- Ownership of AI-assisted output. A red flag is a candidate who treats AI output as something they merely passed along. We want people who take full accountability for anything they ship, regardless of how much of it an AI drafted.
What We Screen For: Talent
Talent remains genuinely hard to articulate precisely, and I won’t pretend otherwise. But in the current environment it has become more legible by reframing it around what AI cannot yet do reliably:
- Problem framing. AI is extraordinary at solving well-specified problems and poor at deciding which problem is worth solving. Talented candidates instinctively reframe a vague ask into the right question.
- Systems thinking. The ability to hold a whole system in their head—dependencies, second-order effects, edge cases—is exactly what AI tools struggle to maintain across long contexts. This is a durable human advantage.
- Domain depth that anchors AI use. AI accelerates experts and dangerously misleads novices, because experts can spot when the output is subtly wrong. Deep knowledge in some domain is now more valuable, not less, because it’s what makes someone a safe operator of powerful tools.
A practical way to assess this: assume the candidate has AI tools available and design the evaluation accordingly (see below).
Modern Interview Practices
Take structured notes
Take notes during every interview—not optionally, but as a discipline. Use a shared digital tool (Notion, a structured form in your ATS, or a collaborative doc) rather than paper or scattered files, so that notes are searchable, comparable across candidates, and attached to the candidate record automatically.
A modern enhancement worth adopting: with the candidate’s explicit consent, use an AI notetaker or transcription tool so the interviewer can stay fully present in the conversation rather than splitting attention between listening and writing. Always disclose this and get agreement first. The interviewer still writes the evaluation—the AI only captures the record. Never outsource your judgment to a transcript summary.
Run a consistent interview structure
Have a prepared deck or outline for each team. The classic three-part structure still works well:
- About Us — A quick introduction to what we do. If the candidate has clearly done their preparation, compress this to a high-level overview and spend the saved time elsewhere.
- About the Candidate — The core of the interview, where we learn how they think and work.
- About the Role — If the first two parts went well, go into genuine detail here. If not, keep it to a brief overview.
If we’re likely to advance the candidate, we may have them meet other team members—but keep this tight and time-boxed.
Add an AI-era evaluation component
The most important update to our process: evaluate candidates the way they will actually work—with AI tools available, not banned.
The old “whiteboard the algorithm from memory” exercise tests a skill that AI now handles trivially. Replace or supplement it with assessments that reveal judgment in an AI-augmented workflow:
- The collaborative work session. Give the candidate a realistic, ambiguous task and explicitly allow them to use whatever AI tools they prefer. Watch how they use them: Do they prompt well? Do they catch the model’s mistakes? Do they know when to stop trusting it and think for themselves?
- The critique exercise. Hand the candidate a piece of AI-generated work (code, a strategy memo, a marketing draft) that contains subtle but real flaws. Ask them to evaluate it. This directly tests the taste and judgment that now matter most.
- The scoping conversation. Present a loosely defined problem and ask them to talk through how they’d break it down. This surfaces problem-framing ability—the scarcest and most valuable skill.
A useful guiding principle: if AI can ace your interview question, the question is testing the wrong thing.
Write up the summary same-day
After the interview, write your summary and recommendation before the end of the workday, while the conversation is fresh. You can use an AI assistant to help organize your notes into a first draft, but the assessment, the recommendation, and the accountability for it are yours. We file both the interview notes and the summary as part of the employment record for successful candidates.
About This Handbook
These notes come from OnDemandWorld’s early years, 2009 to 2017, when we were a team of around 25 working on mobile apps and growing organically. They were written as practical, in-house guidance—not theory—so the voice is direct and occasionally specific to who we were at the time.
The company has since evolved into an agentic AI business and rebranded as odw.ai. Where the original principles still applied, we’ve adapted them; where they didn’t, we’ve let them go. We’ve kept the original text largely intact rather than rewriting history, so you’re reading what we actually used, with light modernization where it helps.
If any of it proves useful to you, that’s the whole point. And if you’re curious where these ideas led, that’s what we’re building at odw.ai.
0 thoughts on “Our People”