Appendix C: The Configuration Profile
Your standards, externalized into a document the agents read and you maintain. Companion to Chapter 7.
The configuration file is the onboarding packet for an employee who forgets everything between shifts, and writing it is where the self-knowledge happens. It holds the standards the other artifacts run against. Build it once with real hours, review it on a schedule, and write it only at the layer that does not move.
Version date: ______ Last full review: ______ Next review due: ______
Part 1: The self-interview (answer before you write a single rule)
| Question | Your answer |
|---|---|
| Pick a rule you enforce in your own work without ever having stated it. What is it, and what lesson is it protecting? | |
| Which trade-offs actually matter to you, and which do you pretend to weigh but always resolve the same way? | |
| Which specific failures will you tolerate to move fast, and which will you stop a launch over? | |
| What question do you always ask about user behavior that your peers do not? | |
| Where do you demand certainty, and where do you accept a calculated guess? | |
| What makes you stop trusting someone who is supposed to know your domain? |
Part 2: Encode, and do not encode
The test for every line: does it encode something true about your work, or something true about this model’s current quirks? The first kind is a standard and belongs here. The second kind is a sticky note on a machine you are about to replace.
| Encode (geology) | Do not encode (weather) |
|---|---|
| Standards for what good looks like, specific enough to act on | Phrasing tricks tuned to this month’s model |
| Trade-offs you have learned to weigh a particular way | The magic preamble that stops this version hedging |
| Constraints that exist because something went wrong once | Anything that would be wrong the moment you switched models |
| Which numbers get checked against the source, never against the model’s memory | Workarounds wearing a standard’s clothes |
Part 3: The scars (rules with the story inside them)
| Rule, as the agent reads it | The incident underneath | What it prevents |
|---|---|---|
Part 4: The stranger test (run at every full review)
Hand the file to a model you have never tuned, or a colleague who does not know your work, on a real task. Would you sign what comes back without rewriting it? If yes, the file carries your standards rather than your current model’s habits. If no, the gap is the next thing to learn about your own judgment; name it and encode it.
| Date run | Model or colleague | Signed without rewriting? | The gap it exposed |
|---|---|---|---|
How to read the result
The profile is for you, not the model. A good configuration is never finished, for the same reason self-knowledge is never finished, and on every upgrade day it is the fixed point you coach the new actor back to. If the stranger test keeps passing while the models keep changing, the standards were real. If it fails on upgrade day, what broke was never a standard.