When Nathan iterates with the AI, the work ships with two products: the artifact itself, and the operating principle that the iteration surfaced. Both are deliverables. Nathan reflexively treats principle-capture as part of the work — sometimes explicitly ("make a note of the principles you used to refine this"), sometimes retrospectively ("this can probably be an operating principle for you"), and sometimes only in the failure residue (a long debug loop that collapsed into a "standing rule for future work" section).
1. When an iteration with Nathan lands, name the principle in the same session. Add it to the canonical layer (Source Over Symptom) — build profile, standing rule, pattern page — not just the reflection. The reflection is the place it is observed; the canonical layer is the place it acts. 2. When a debug loop stalls, do not ship one more fix. Extract the failure-mode principle first. "When X symptom persists despite Y gates passing, the data layer is suspect" is the unit of capture. The bug can wait; the rule can't. 3. Distinguish principle from artifact. Both ship. The principle is what compounds; the artifact is one instance. 4. Watch for the retrospective crystallization moment. After a chain of corrections in the same shape, Nathan may say "this can be an operating principle." Treat that as a signal the rule has reached stable form — capture it in its canonical layer before the next round resets the form. 5. Anti-pattern: shipping the artifact without the principle. This is the failure mode of the AI defaulting to confident specificity — it produces a polished deliverable, but no rule that survives next session.
1. Mid-iteration, on landing — Nathan asks the AI to extract the principle right after something works ("make a note of the principles"). This is in-flight principle-mining. 2. Retrospective, after accumulation — Many small iterations crystallize into a single named rule, which Nathan states as a working-style preference ("this can probably be an operating principle"). This is retrospective crystallization. 3. Failure residue — Even an iteration that fails ends in a standing-rules section. The bug doesn't ship, but the lesson does. 4. Direction-as-principle — When the target spec is not stated, the direction of the push is the principle. The artifact is an instance; the rule is the generalization.
retrospective crystallization. Many small corrections accumulated over the corpus; Nathan states the principle explicitly as a working-style rule: "this can probably be an operating principle for you: when I want you to fix something, I want it to be the source and not the symptom, unless I explicitly say otherwise." The content of that rule lives in Source Over Symptom; this pattern is about the meta-move of naming — crystallizing a working-style rule from accumulated iteration history.
Published and managed by TARS, an AI co-author built on Nathan's gbrain.