skill-authoringlisted

# Skill Authoring Cache over `~/.claude/harness/00-DIAGNOSIS.md` (why), `40-maintenance.md` (edit permissions for THIS environment) and `50-letter-to-future-sessions.md` (degradation modes). Harness wins on conflict (R8). ## Contents §0 who reads · §1 rule format + eligibility gates · §2 ground-truth content · §3 memory architecture, caches, size ceilings, skill anatomy · §4 lessons, probes, growth control · §5 edit safety · §6 the honest limit ## 0 · Who reads rules files — write for that reader The reader is the weakest model that will ever load the file, under time pressure, mid-task, with a polluted context. It executes literally, latches onto whichever sentence pattern-matches the task, and cannot compute current truth from a sequence of updates. Every rule below follows from that. ## 1 · Weak-model-executable rule format - **Current-state imperatives, present tense.** The litmus: a sentence that NARRATES history — what was found, measured, or retracted, with its date and N — belongs in a memory file, not a rules file. A weak reader acts on the retracted claim anyway — it's in context, stated first, stated confidently. Operational metadata is exempt and required where a rule demands it: a version/date pin on a capability-negative, a dated port note, a `SUPERSEDED <date>` marker, a probe-status marker — those are part of the order, not narration. - **Judgment → mechanical checklist.** Criteria with pass/fail boxes and a verdict ("if ≥2 hold, stop re