# Instruction Files Are the Last Rung Not the First ## Claim A recurring agent failure should be eliminated in the architecture first and enforced by the toolchain second, with a written instruction reached only when neither of those can hold it. <!-- atomicity: 1 idea ✓ --> ## Evidence **Source**: [[3. Resources/Readwise/Articles/Turn Off Claude Code's Memory|Turn off Claude Code's Memory]] — Armin Ronacher, stating the ordering > [!quote] > "skills should be a fallback that you fall into when the codebase can't solve the problem and the lint rules in CI also can't. Skills shouldn't be this thing you reach for all the time to build this crazy set of skills that will solve all your problems. They are a safety net. They are a fallback." The ladder has three rungs and they are ordered by how the fix survives being ignored. Rung one: > [!quote] > "The best thing you can do always is categorically eliminate the problem through better architecture or choice of data structures." An architectural fix makes the mistake unrepresentable — nobody has to remember it, including the agent. Rung two, a CI lint rule, makes the mistake representable but unmergeable; it still fires when everyone forgets. Rung three, an instruction, only works if it is loaded, read, and heeded, which makes it the only rung whose enforcement is probabilistic. This inverts the default practice. Writing an instruction is the cheapest response available in the moment — it takes a minute, it feels like a fix, and it produces a visible artifact. Restructuring the code or adding a lint rule costs hours and produces no document. The ladder says the cheapness is exactly the problem: instructions are reached for first because they cost least, and they are the rung least likely to hold. ## Scope & conditions - **Holds** for recurring, mechanically checkable failures — the class that a type, a schema, a lint rule, or a deleted code path can categorically exclude. - **Breaks** for genuinely judgment-shaped guidance: tone, product positioning, which of two valid designs this team prefers. Ronacher's own carve-out is that an agent should understand the product, and no lint rule encodes that. - **Requires a real CI rung.** Where the middle rung does not exist, the ladder collapses to architecture-or-instruction and the instruction pile grows regardless of intent. - **Cost asymmetry is the whole mechanism.** The rungs are ordered by durability and reached for in the opposite order, by cost. Any practice that does not name this will drift back to instructions. - **Does not follow** that skills and instruction files are bad. It follows that their count is a diagnostic: a growing pile is evidence of problems that were never pushed down to a rung that holds. ## Links - [[3. Resources/MOCs/AI Agent Safety MOC]] — an escalation order that decides which failures are prevented structurally and which are only requested - [[3. Resources/MOCs/Personal Knowledge Management MOC]] — the discipline that keeps an instruction corpus from growing past what anyone reads - [[3. Resources/Permanent Notes/When Review Cannot Scale Correctness Must Move Into the Toolchain]] — extends: that note argues correctness must leave human review for the toolchain; this one gives the order of rungs to move it through, and names instructions as the rung that is not toolchain at all - [[3. Resources/Permanent Notes/Security Intent Belongs in the Initial Prompt Not a Cleanup Pass]] — because contrasts: that note pushes a concern *earlier into* the instruction; this one pushes it *out of* instructions entirely where a structural rung exists, and the two are compatible only because security intent is judgment-shaped - [[3. Resources/Permanent Notes/Agent Memory Is Written Far More Often Than It Is Read]] — evidence-for: an instruction pile that grows unchecked is exactly the store that ends up written three times for every read - [[3. Resources/Permanent Notes/The Privacy Boundary of a Review System Is Its Output Schema]] — evidence-for, cross-domain: Visa's frontier-model review is enforced by output schema "at the architectural level, not just policy level," the same rung-one move this note argues for, applied to enterprise AI review rather than agent coding failures - [[3. Resources/Permanent Notes/Abstract Only After Two Real Implementations Exist]] — because extends: the same restraint applied to a different artifact, where writing the general rule too early is the cheap move that feels like progress - [[3. Resources/Permanent Notes/Interface Surface Area Costs Selection Accuracy Not Just Tokens]] — applies-in: every instruction file added is surface the model pays for and must select among, so the ladder's ordering has a measurable accuracy dividend ## Tests / falsifiers - Take the five oldest rules in an instruction file and ask, for each, whether a type, a schema, or a lint rule could enforce it. A high hit rate means the ladder is not being climbed. - Add a CI rule for a rule currently written as prose and delete the prose. If violations do not reappear, the instruction was never doing the work the rule now does. ## Maintenance - Last tended: 2026-08-28 - Next review: 2026-11-28 ## Source [[3. Resources/Readwise/Articles/Turn Off Claude Code's Memory|Turn off Claude Code's Memory]]