How I work

1. 01

   Map the system

   Understand the operating environment before touching it: who does what, in what order, and what the output actually has to be.

   Seen in

   • A Spanish visa is nine procedures, one per consulate, and jurisdiction follows where the applicant lives rather than where they are going.
   • The broker's day drawn before software, so the system answers a real day and not a category.
   • Several divisions and levels of an account worked at once until the firm made the preferred-vendor list.
2. 02

   Find the constraint

   Locate where time, information, decisions or capacity are being lost. It is rarely where the complaint is.

   Seen in

   • A solo operation's binding constraint is the operator's attention, so the valuable work is the work that removes the operator from the loop.
   • The appraiser, not the buyer, sets the ceiling on price per square foot.
   • Supporting documents form a chain that takes weeks, so the checklist models dependencies rather than steps.
3. 03

   Separate fact from inference

   Know which numbers are measured, which are derived, and which are guesses, and never let the three share a column.

   Seen in

   • Every visa fact has one home with its source and verification date. A claimed quote must exist in the source or the alert is rejected.
   • Walls that could not be inspected before purchase: the termite risk was priced blind, then the walls were opened.
   • Confirmed negatives outrank criteria, because the dangerous claim is the adjacent-true one.
4. 04

   Decide what should be deterministic

   Anything that has to be right every time, explainable in a sentence, or auditable later is computed, never generated.

   Seen in

   • Scoring uses no language model. Intent, urgency, engagement, quality and recency each return their own score and reasons.
   • Guidance was AI and was rebuilt to derive from the knowledge hub. Removing the model was the improvement.
   • After-remodel value from comps and the budget as a range, best case to worst. Those two numbers set the maximum purchase price.
5. 05

   Delegate the right work to models

   Models get the work they are good at: prose, extraction, drafting and review, each with a fixed role and a boundary drawn around it.

   Seen in

   • The model touches the two ends of the pipeline, never the middle.
   • Eight agents, each pinned to a model. The cheap model files, the middle one drafts, the expensive one only reads contracts.
   • The evaluator stays on a small model on purpose, because that is the tier doing the real screening.
6. 06

   Keep humans where judgment matters

   Irreversible actions and ambiguous calls stay with a person, and the system is built to route to them rather than around them.

   Seen in

   • Ambiguous matches go to a human queue. False positives corrupt memory permanently; missing links are cheap to fix.
   • Nothing sends and nothing is marked paid by an agent.
   • Generated forms are never flattened, because the user's review pass is the final accuracy checkpoint.
7. 07

   Build the smallest useful system

   The first version is the least machinery that produces the output, and everything it does not need is left out on purpose.

   Seen in

   • No service, no database, no runtime: markdown files and Claude Code.
   • Two records stored. Every view is generated on request, so no list can drift.
   • One application, no separate services, as a deliberate solo-operator constraint.
8. 08

   Measure what changed

   The result is a number the next decision can use, and where the number is noisy the noise is reported with it.

   Seen in

   • Days to contract and price per square foot from entry to exit, on every project.
   • The chained preparation ran in roughly half the time the same work took by hand.
   • Nondeterminism is measured, not smoothed. Verdicts that move across runs on unchanged text are labeled unstable.
9. 09

   Document weaknesses instead of hiding them

   Every project records what went wrong, what is still unknown and what would be tested next.

   Seen in

   • No ground truth, small outcome data, and stubs that halt rather than improvise.
   • Seven inherited contradictions parked in one file for the operator to arbitrate, not papered over.
   • The rental was managed remotely and should not have been.

The decision structure

Every project on this site records its decisions the same way.

What was chosen, what drove it, what improved, and what is still imperfect. The limit says where the next decision starts. One from a construction project and two from the software, in the same four parts.

The five systems, each with its decisions and limits