Two Fish LabsLow stakes, clear answer
AI handles it on its own. Formatting, renaming, filing.
How I work.
There's no fixed playbook here. Every engagement starts the same way: understand how work actually happens before proposing a fix. Sometimes that means changing team structure or leadership. Sometimes it means putting AI directly into a workflow, not as a headline feature, but as leverage that frees people from repetitive work.
01 / Principles
Three things I hold to
01
Find the real constraint
It might be the team. It might be the tech. It might be a gap in the product. Before adding process or headcount, I find out which.
02
Fix what's actually broken
Leadership, team structure, technical architecture, delivery process, or the product itself: whatever's creating the constraint gets addressed directly, not patched around.
03
Design for independence
I build so things can keep running without me, not so they depend on me. Less key-person risk, without adding overhead that slows things down.
02 / An engagement
What an engagement looks like
There's no fixed playbook, but most engagements move through the same six steps. Some companies need all six. Some need two.
01
Diagnose the bottlenecks
A fast technical and process audit. I talk to the team, read the code, and watch how work moves, to find where it's actually getting stuck.
02
Build a repeatable delivery system
Lightweight engineering process that makes delivery predictable, without ceremony that slows people down.
03
Stabilize the team
Mentor, upskill, reset expectations, and put the right roles in place.
04
Reduce technical risk
Clean up unstable architecture and remove tech debt that could hurt you.
05
Give founders clarity and control
Reporting and a roadmap that show what's happening, what's blocked, and what's next, so you're not relying on someone else's summary.
06
Scale with confidence
Hire, reorganize, or expand the team to match the growth curve.
03 / Where AI fits
Where AI fits
AI is a tool in the work, not the product. Most teams lose a large share of their day to work about work: status updates, handoffs, hunting for information. That's the robot work, and it's where AI helps first. A model can draft something, summarize something, or take a repetitive task off someone's plate, but deciding whether the result is right stays with a responsible person. AI might draft, but I decide. I approach it in four steps.
01
Find the robot work
Before automating anything, I find out what's worth automating. I ask the team what's repetitive and frustrating, look at where work sits waiting in your tools, and see where the time actually goes. The result is a ranked list of opportunities by impact and effort.
02
Design the workflow
For the best opportunities, I define what the AI does, what tools and information it can reach, how work hands off to a person, and where it has to stop. Small, focused helpers with clear jobs are easier to trust than one assistant that does everything.
03
Keep a person in the loop
Every output passes an approval point owned by someone on your team. The AI stops and asks when something is ambiguous or high stakes. People correct it, and those corrections make the next result better.
04
Roll it out
Start with one workflow and prove it works. Then extend it, keep an eye on how it performs and what it costs, and help the team adapt to working this way.
04 / Oversight
Deciding what AI does on its own
Not every task deserves the same oversight. I sort tasks by how much is at stake and how uncertain the answer is.
High stakes, clear answer
Automated checks, with a person watching. Running the test suite before a release.
Low stakes, unclear answer
A person steers. Drafting a first version of a client update or an article.
High stakes, unclear answer
A person leads and AI assists. Architecture decisions, hiring, anything hard to undo.
05 / Rules
Three rules I hold AI to
I can see what it did
Every step is visible and logged, not just the final result.
Someone checks what matters
High-stakes actions are verified by a person or an independent check before they take effect.
It stops when it's unsure
When the AI hits a case it doesn't understand, it asks instead of guessing.
The biggest risk with AI isn't that it fails. It's that it fails quietly and nobody checks. When something goes wrong, accountability stays with a person, which is why I design for a person to be in the loop from the start.
06 / Beyond engineering
For teams outside engineering
Marketing, sales, HR, and operations carry much of the same work about work. The approach is the same.
01
Map the process
Write down how work actually moves, including the parts that live in email threads and people's heads. Separate the routine decisions from the ones that need judgment.
02
Add focused helpers
Research, drafting, and coordination are common starting points: gathering background, producing a first draft in your voice, keeping schedules and trackers current.
03
Keep people in charge
People review, decide, and say why something was changed, so the helpers learn how you work.