2 commits

Author	SHA1	Message	Date
Claude	b99505446a	fix: resolve 11 P0 correctness defects from newgen audit ... - fda-consultant-specialist: QMSR transition (eff. 2026-02-02) correctly framed; legacy QSR sections mapped to ISO 13485:2016 clauses (17-row mapping) - risk-management-specialist: ALARP-with-cost-benefit replaced by AFAP per EU MDR Annex I GSPR 1-4 + EN ISO 14971:2019/A11 - eu-ai-act-specialist: Art. 5(1)(f) sample corrected (workplace, not retail); both dual-published copies kept byte-identical - financial-analyst: 4 scripts now read bundled sample shape (and flat shape), exit non-zero on missing data; verified real outputs (ROE 25%, DCF $13.90/sh) - deal-desk: fixed-COGS margin math unified across script/reference/SKILL.md; margin scoring recalibrated (30% discount on 80% GM now scores 25, not 37) - senior-qa/backend/frontend + email-template-builder: 8 corrupted code literals from past bulk YAML-quoting sweep reconstructed - google-workspace-cli: install coordinates fixed to verified upstream (@googleworkspace/cli); phantom subcommands replaced with verified syntax - project-management: fabricated Atlassian MCP tool names replaced with the 30 real Remote MCP tools; new canonical references/atlassian-mcp-tools.md; unavailable capabilities routed to UI/REST; orphaned scripts wired - command-guide: retired (documented a different repo's ecosystem) - markdown-html: stale v2.10.0 'foundation' text replaced; routing now targets the shipped converters; never hand-render - webinar_funnel_scorer.py: argparse added (--help/--sample/stdin) - handoff redaction linter: missing 17th pattern (private CIDR) added https://claude.ai/code/session_019AJddAL1NADWMXsy1qNPQF	2026-06-10 14:32:43 +00:00
Claude	7bdc98e517	feat(productivity): reflect skill — Path-B light-prompt-flow sibling of capture ... Slice 8: productivity light-prompt-flow sibling. Same shape as capture (11 files, max-1-question intake, fast-to-action), different mode — capture organizes external dumps; reflect re-examines internal conversation state. After this merges: 10 of 13 v2 megaprompts shipped. SOURCE SPEC megaprompts/02-reflect-megaprompt.md (PR #657). WHAT THE SKILL DOES Mid-conversation reflection. Pauses execution, re-reads the FULL conversation from original goal forward (not just recent turns), runs the 5-dimension analysis framework: - Macro Perspective (original goal vs current; drift detection) - Gap Analysis (assumptions / stakeholders / constraints / alternatives / external factors) - Reflective Inquiry (right problem? simpler path? harder valuable path avoided?) - Bias Check (confirmation / sunk cost / anchoring / complexity / recency — each with recognition cues) - Contextual Alignment (does direction serve actual goals + best use of time + external factors) Delivers flowing prose (NO headers, NO bullets). Ends with mandatory directional recommendation: Continue / Pivot to {X} / Pause for {Q}. KEY PATH-B PRESERVED ELEMENTS - Re-read FULL conversation from original goal (not just recent turns) — the discipline that distinguishes real reflection from local summary - Halt-current-thread stop directive (reflection is a pause, not a side-quest) - Honest-output discipline: NO manufactured problems when path is solid; NO vague reassurance ("looks good!") instead of specific reasoning - 5-dimension framework preserved verbatim - 5 biases preserved (confirmation, sunk cost, anchoring, complexity, recency) with recognition cues - Flowing prose enforced (no headers, no bullets in body) - Closing recommendation mandatory (Continue / Pivot to X / Pause for Q) - Low-intake: max 1 optional clarifier (only when context is thin); default to no questions - No name references (generic second-person throughout) - Implicit triggers OFFER reflection, never auto-invoke (10+ detail turns / frustration / dead-ends → ask user if they want to step back, don't unilaterally run) PURE-REASONING SKILL No external APIs. No DOCX generation. No file-system writes beyond audit. Most portable v2 skill — works in Claude Code CLI + Claude.ai web natively, no MCP dependencies, no Node.js, no Consensus account required. REPO STRUCTURE (mirrors capture 1:1) productivity/reflect/ ├── .claude-plugin/plugin.json ├── README.md ├── agents/cs-reflect.md ← reflection persona, honest-output enforcer ├── commands/cs-reflect.md ← /cs:reflect (or auto-triggers on phrases) └── skills/reflect/ ├── SKILL.md ├── references/ │ ├── cognitive_bias_canon.md ← 5 biases + recognition cues │ (7 sources: Tversky/Kahneman, │ Wason, Arkes/Blumer, │ Russo/Schoemaker, Tetlock, │ Karpathy) │ ├── honest_output_discipline.md ← anti-manufactured-problems │ (7 sources: Yegge, Gawande, │ Deming, Russell, Kim Scott, │ Bret Victor, skill spec) │ └── conversation_reflection_practice.md ← Schön reflective practice │ (7 sources: Schön 1983 + 1987, │ Argyris/Schön, Kolb, Polanyi, │ Kahneman/Tversky, Victor) └── scripts/ ├── bias_pattern_detector.py ← stdlib: regex scan for 5-bias │ signal patterns ├── conversation_depth_analyzer.py ← stdlib: turn count + implicit │ trigger signal detection └── directional_recommendation_validator.py ← stdlib: verify output ends with Continue/Pivot/Pause + specific evidence + flowing prose 11 files, 1,554 lines. Comparable to capture (1,560 lines). VERIFIED CLEAN All 3 scripts pass smoke tests: - bias_pattern_detector --sample (notification system + sunk cost + anchoring + complexity scenario): correctly detects 3 biases (sunk_cost via "we've invested", anchoring via "sticking with", complexity via 9 "what about X" hits). Correctly clears confirmation + recency (no strong signals). - conversation_depth_analyzer --sample (19-turn debugging conversation with stuck-ness markers): correctly verdicts OFFER_REFLECT based on frustration (8 hits) + dead-ends (4 hits). Note: "skill should OFFER reflection, not auto-invoke" — honors design intent. - directional_recommendation_validator --sample-pass (honest validation output with 16 specific-evidence references): PASS 6/6. - directional_recommendation_validator --sample-fail (vague reassurance with bullets, no recommendation, no evidence): FAIL with 4 specific issues caught (missing closing recommendation, 2 vague phrases, 3 bullets, 0 specific-evidence refs). All 3 with --output json: valid JSON. plugin.json validates. VERTICAL-SLICE STATUS ✓ Slice 1: capture (PR #659) ✓ Slice 2: pulse (PR #660) ✓ Slice 3: email pair (PR #661) ✓ Slice 4: landing (PR #662) ✓ Slice 5 batch 1: litreview (PR #663) ✓ Slice 5 batch 2: grants + dossier (PR #664) ✓ Slice 5 batch 3: patent + syllabus (PR #666) ✓ Cleanup PR: move pulse + capture (PR #667) ✓ Slice 8: reflect (this PR) ☐ Slice 6: notebooklm (browser-automation, last shape) ☐ Slice 7: 13-research orchestrator + autoresearch-agent reconciliation 10 of 13 v2 megaprompts shipped after this merge. 3 remaining: notebooklm (browser-automation), 13-research (orchestrator), then v2 is complete. NOT DONE IN THIS PR (intentional) - .claude-plugin/marketplace.json not updated (separate concern; done after all 13 ship) - .codex/skills/reflect symlink not added (auto-sync workflow handles on merge per existing pattern) https://claude.ai/code/session_01FEUmeuYhmnxVFq7EZM8ZSw	2026-05-16 05:37:30 +00:00