Section 8 · 37 entries · principles · corrections · cross-cutting
Learnings library.
Principles authored across the team — per-agent learnings, cross-cutting principles (R-numbered governance + SHARED-LEARNINGS), and the session-corrections cluster (S58 → S75 codification log). The team's accumulated discipline.
Per-agent principles · 15 of 15
Anti-pattern #1 — cookie-cutter AI aesthetic
8 specific defaults to reject. Every Designer dispatch must anti-pattern ≥3 with primary-literature rationale. Cookie-cutter pastel-on-white frequently fails WCAG 1.4.3 AA.
Anti-pattern #2 — lead-with-user-summary-as-target
Designer derives IA shape from user verbal phrasing rather than primary research. R34 violation; sibling C24 root.
Anti-pattern #8 — WCAG-as-post-hoc-patch
Accessibility checked at handoff time only. WCAG 2.2 AA is non-optional standing gate per Cluster 4.
Anti-pattern #10 — single-direction-proposal-skipping-divergent-half
Phase 5 task decomposition outputs single design direction without enumerating 2-3 alternatives considered + rejected (Buxton 2007 Double Diamond).
H-1 first-time-student simulation runs FIRST on every screen
Before any structural / heuristic check. If section is not obvious from on-screen alone, section FAILS. Krug 2014 first-law floor.
Anti-pattern #11 — soft-sign-off-after-R30-codification
Phrase-level PASS verdict instead of formal sign-off-block (reviewer + timestamp + scope + verdict). R30 codified after S75 C15 fired.
Anti-pattern #13 — decorative cross-class invocation
Citing Researcher / Designer corpus literature decoratively without primary-literature anchor in cross-class claim = explicit failure.
Researcher anti-pattern — vendor-survey-instead-of-first-principles
Researching agent landscape via LangChain / CrewAI marketing instead of grounding in agent-architecture primary literature. Symptom: brief reads as feature comparison.
Progressive elicitation (S43 origin)
When writing multi-section arcs, each section should add one visible element to the running artifact. The 3 C's pattern (Card → Conversation → Confirmation).
Session 2 retention (S43 origin)
When designing Session 2 of any course, the aha must escalate from S1 (not repeat it). The student's artifact is the commitment device.
Four infrastructure patterns for tutor_active sections
Every tutor_active section needs context_rule, narration_rule, named generation directives, and length-constrained teaching_point. The #1 structural cause of quality gaps.
Re-verify after fixing
After fixing code based on agent feedback, re-invoke the reviewing agent to verify the fix. Don't assume the fix is correct without independent verification.
Read seed data as the student
Seed data must be read as the target student would experience it, not as the author.
QA runtime verification
QA narrative journey must include runtime evidence (Playwright, screenshots, console logs). Seed data analysis alone insufficient.
Exhaustive testing not representative
Test plans must click every link, tab, and button — not a sample. QA's 17-item test missed broken links Jon found in 30 seconds.