Update the skills by claude promptings

Author: johnteee

reflective-prompt-library/02-engineering/code-reviewer.md

@@ -16,6 +16,7 @@ Review diffs against spec, tests, and risks. Primary workflow surface: `reflecti
 
 - Spec traceability table completed.
 - Decision is Approve, Request changes, or Reject with non-empty required fixes when not Approve.
+- Every reported defect names a reachable failure scenario or code-supported violated invariant; uncertainty stays explicitly unverified.
 
 ## Falsifiability
 
@@ -42,6 +43,7 @@ Escalate to `reflective-risk` when findings imply trust-boundary or high-blast-r
 {貼上測試結果，如有}
 
 ## Review Checklist
+每項程式缺陷都要指出可到達的失敗情境（輸入、狀態、時序或平台，以及可觀察到的錯誤結果），或以程式碼證據指出被破壞的不變量；否則要標示為未驗證，不能當成已確認的缺陷。
 
 ### 1. Correctness
 - 是否滿足每個 acceptance criterion？

reflective-prompt-library/02-engineering/task-start.md

@@ -17,6 +17,7 @@ Establish a task brief before implementation. Primary workflow surface: `reflect
 - Goal and scope in/out stated.
 - Acceptance criteria and falsifiability each named once.
 - Minimal next step and defer list included.
+- Any clarifying question names the local evidence checked and the material fork it could not resolve.
 
 ## Falsifiability
 
@@ -27,6 +28,7 @@ If the brief cannot name evidence that would prove the framing wrong, stop and c
 
 ## 任務
 {貼上任務}
+若需要提問，請先簡短查閱可用的專案檔案、文件與既有記憶；只詢問查閱後仍無法判定，而且會影響方向的分歧。
 
 ## 請輸出
 

reflective-prompt-library/04-agent/memory-consolidation.md

@@ -16,6 +16,8 @@ Consolidate repeated lessons into maintainable knowledge. Primary workflow surfa
 - Each pattern has recurrence evidence or explicit project decision pointer.
 - Promotion destination classified (principle, lesson, decision index, skill candidate).
 - Authority class separated: project-design judgement vs agent operating rule.
+- Each retained memory is future-useful, durable beyond the current task, self-contained for a reader without the original session, and identifies changeable facts that require fresh-source revalidation.
+- Live task state and lookup-recoverable facts are excluded.
 
 ## Falsifiability
 
@@ -50,5 +52,10 @@ Escalate to `reflective-risk` before promoting executable knowledge (skills, hoo
 - 重複模式
 - 應該制度化的規則
 - 不值得制度化的偶發問題
+
+請另外套用記憶品質檢查：
+- 只保留未來仍可重用、跨任務仍有效，而且脫離原始工作階段也能獨立理解的內容。
+- 排除即時任務狀態，以及可直接從目前原始資料查得的事實。
+- 把記憶視為有日期的線索；會變動的內容在使用前要對照最新權威來源重新驗證。
 ```
 

reflective-prompt-library/PROJECT_KNOWLEDGE.md

@@ -87,6 +87,8 @@ deferred promotions are recurrence-gated — see [panel backlog](plans/multi-age
 > not duplicated here; this is a map, not an archive. Mechanical panel rounds
 > are rolled up; the round-by-round trail lives in the panel record.
 
+- 2026-07-24 Claude Code prompt snapshot survey — seven-lens Parallel Lens Review of Piebald's third-party v2.1.218 corpus decided `AGREE WITH CHANGES` 7/7: retain the nine-core/three-pack architecture; adopt three original in-place repairs (research before clarification, reachable failure scenarios for code findings, memory quality plus revalidation); defer exact proposal-digest binding pending explicit pack approval and enforcing-host proof; reject automatic skill mutation and micro-fragment rearchitecture. Renderer reproduction reached 603/603 published bodies, while a fresh public extractor returned 416, so full extraction and behavioral effectiveness remain unverified → [record](plans/claude-code-system-prompts-survey-2026-07-24.md)
+
 - 2026-07-18 All-skill layer review — seven-lens Parallel Lens Review over all 12 repo skills decided `AGREE WITH CHANGES`: metadata and core/domain registry are sound, the nine-core routing boundary holds, but installed-host portability needed co-installed examples, inline dispatch boundary cues, flow-control human-review parity, evidence-tier caveats in examples/YAML, and field-use/R17 reconciliation. Adopted AS1–AS10; deferred R10 shrink, G9 governance holdouts, full R17 runtime rig, and optional index schema work → [record](plans/all-skills-panel-record-2026-07-18.md)
 
 - 2026-07-18 Agent-governance-scaffold pack panels — initial six-lens Parallel Lens Review decided `AGREE WITH CHANGES` 6/6: admission mechanically sound but wording above evidence tier; adopted runtime-boundary honesty, composite-self-acceptance disclosure, demotion/G9 triggers, concepts tier labels, lint honesty, registry-driven pack guards, with R10 shrink deferred. A second post-remediation six-lens pass also decided `AGREE WITH CHANGES` 6/6 and found the first guard-closure claim overstated: task-minimal vs mandatory artifacts conflicted, the invariant “mapping” was only a name list, parseable templates were called schemas, and token guards preserved receipt/budget/L0/example/experiment inversions. R12–R16 add a conditional Artifact Set, fillable object→invariant→enforcer map, host-owned static security falsifiers, semantic/routing/install guards, G9’s 2026-10-11 review hook, and zh-TW Human Review parity. A third closure-readiness pass again decided `AGREE WITH CHANGES` 6/6 and adopted R18–R24: `deny_write`/`constitutional_paths` parity, Four-Power↔Output reconciliation, interim `context_load: high`, adoption provenance/demotion annotations, Example 2 status-gate parity, wrapper verification qualifier, and stronger section-local guards; formal schema/runtime proof stays R17-deferred and R10 shrink remains checkpoint-bound → [panel record](plans/agent-governance-scaffold-panel-record-2026-07-18.md)