You will be able to say what an engine is, what the loop does, and whether your work has one in it.
For owners, managers, and experts who have never built an ai system. About 45 minutes.
LESSON 01 · 8 MIN
What an expert knows that the manual does not
Every trade has a manual and a veteran. The manual has the tables; the veteran knows when the table is wrong. A cementer knows that a certain well below a certain depth wants a retarder even though the table does not say so. A dispatcher hears 'water at the meter' and sends two people. That knowledge is tacit: it is real, it is consistent, and nobody has written it down.
Tacit knowledge is expensive in three ways. It walks out the door when the person retires. It does not scale past one person's day. And it cannot be checked, so a bad habit looks the same as a good one.
The point of a Tacit Engine is not to replace the veteran. It is to let the veteran teach a system the same way they would teach a trainee: by correcting its work. The corrections are what we are after.
✓Tacit knowledge is consistent judgment nobody has written down.
✓The goal is to capture corrections, not to replace the expert.
✓If your veteran corrects a trainee's work, you have the raw material for an engine.
LAB · SORT THE KNOWLEDGE
Drag each statement to 'in the table' or 'in the veteran's head'. There is no trick: the point is to notice how much of a real job is the second kind.
5½" 17# casing has an ID of 4.892 in.
Below 9,000 ft on Class H, always add a retarder.
A plug must be at least 100 ft long in this state.
That operator's wells always wash out; run 50% excess.
Class A neat weighs 15.6 ppg.
If the company man sounds nervous, send a second truck.
Hydrostatic pressure is 0.052 × ppg × depth.
Friday afternoon jobs get the senior cementer.
LESSON 02 · 7 MIN
What we mean by an engine
An engine is a function: inputs go in, a complete proposal comes out, every time, the same way. For a plug job the inputs are casing, depths, and cement class; the proposal is volumes, sacks, and a ticket. For dispatch the input is a request and the proposal is a crew, a priority, and a window.
Most of an engine is not AI at all. It is the tables and formulas the trade already uses, encoded carefully and tested. That part is deterministic: the same inputs always give the same answer, and you can print the arithmetic.
The AI shows up at the edges: reading a messy description into clean inputs, drafting the explanation, finding the most similar past job. We keep the AI out of the arithmetic on purpose, because arithmetic should not have opinions.
✓An engine is inputs → complete proposal, deterministic and testable.
✓The tables you already use are the engine's core.
✓Language models read and write; they do not do the math.
HANDS-ON · LONGLEAF PLUG DESIGNER
Open the plug designer, change only the bottom depth, and watch which numbers move. Every one of them is arithmetic from the capacity table.
The engine proposes. The expert reviews, exactly as they would review a trainee's work: accept it or change a line. Each change is captured with the conditions it applied under. The next proposal applies what was captured.
The word that matters is capture. A correction that is not written down with its conditions is just a one-off edit. A correction that is written down as 'when depth is below 9,000 ft and class is H, add retarder' is a rule, and rules compound.
The loop has a number attached: the agreement rate, which is how often the expert accepts a proposal unchanged. It starts wherever your tables start and climbs as rules accumulate. When it is high enough for a job type, that job type can go out without review.
✓Propose, review, capture, improve. Capture is the step people skip.
✓A rule is a correction plus the conditions it applied under.
✓The agreement rate tells you when to trust the engine for a job type.
LAB · RUN THE LOOP
Press Review to accept or correct proposals. Watch the rule count and the agreement rate. Then try reviewing without capturing and see what happens to the curve.
Reviews
0
Rules
0
Held-out agreement
70%
LESSON 04 · 8 MIN
Does your work have an engine in it?
Three questions. Does one person make the same kind of decision many times a month? Is there a table, a spreadsheet, or a price list they start from? Would they recognize a wrong answer in under a minute? Three yeses and there is an engine in it.
Good fits: job design from capacity tables, quotes from rate sheets, routing and triage, selecting a product from conditions, first-pass compliance checks. Poor fits: decisions made once a year, decisions with no starting table, and decisions where nobody can say what wrong looks like.
The first engine should be the decision with the highest volume and the clearest wrong answer. Not the hardest one. The hardest one comes third.
The scarce input is the expert's time, so the design goal is to spend it on corrections, not on meetings or data cleanup. We take the tables as they are and ten to twenty past jobs with their answers. That is the whole onboarding.
Returns come in three forms. Faster proposals: the engine drafts in seconds what took an hour. Fewer misses: rules do not forget. And a first pass for customers: once the agreement rate is high, a bounded version can quote for your customers directly, within the rules your expert approved.
The lab below lets you put your own numbers in. It is deliberately conservative.
✓Onboarding is tables as-is plus a handful of past jobs.
✓Returns: speed, fewer misses, a customer-facing first pass.
✓Run your own numbers; do not trust ours.
LAB · YOUR NUMBERS
Enter how many proposals a month, how long each takes today, and what an hour of the expert costs. The lab shows the hours returned at three agreement rates.
AGREEMENT
EXPERT HOURS RETURNED / MO
VALUE / MO
70%
23.3
$2,217
85%
25.3
$2,407
95%
26.7
$2,533
Assumes a review takes 4 minutes plus 20 minutes times the disagreement rate. Deliberately conservative; put your own numbers in.
CHECK YOUR UNDERSTANDING
Quiz
1. Which of these is tacit knowledge?
2. What is the step most teams skip?
3. Where should a language model never be used in an engine?