R&D Intake Log
Source: departments/executive/rd-analyst-riley/scans/intake-log.md
R&D Intake Log
Owner: Riley (R&D Analyst) Purpose: Normalized entry point for all R&D discoveries (TFD-0012 Stage 0)
| Date | Topic | Source | Triage | Rationale |
|---|---|---|---|---|
| 2026-03-28 | Paperclip (paperclipai/paperclip) — Multi-agent orchestration platform | ceo-request | Evaluate | 37K stars, MIT, TypeScript. Org charts + budgets + governance for agent teams. Directly relevant as potential delivery platform for digital talents or competitive signal. CEO-flagged. |
| 2026-03-31 | Codex Plugin for Claude Code — adversarial review as multi-model QA gate | video-analysis | Evaluate | Released today. GPT-5x4 beats Opus on most coding benchmarks at lower cost. Adversarial review detects bugs Opus generates but can't self-detect. Free with ChatGPT subscription. RD-0019 scored 4.4 — fast-tracked. |
| 2026-03-29 | Claude Agent SDK + Harness Design Patterns — Multi-agent orchestration for digital talents | ceo-request | Evaluate | Anthropic's harness design article describes planner-generator-evaluator pattern. Agent SDK enables multi-agent products (quality jump, complexity ceiling raised, cost optimization via model routing). Could unlock premium product tier. Ref: TFD-0014, anthropic.com/engineering/harness-design-long-running-apps |
| 2026-05-14 | Skills-as-package.json: installeur toolkit-catalog + convention CONTEXT.md + mode grill-workshop | video-analysis | Evaluate | Source RD-0006 (Pocock skills repo). Distribution overlaps RD-0024 (.zip/--plugin-url) — extend that. Net-new: toolkit-catalog-specific bundle (synergy maya-rag-wiki), CONTEXT.md token-discipline convention for delivered talents, /grill-me applied to workshop re-runs with customer. Strategic for foundry model. |
| 2026-05-19 | Claude Design (Anthropic Labs) — visual design/prototyping tool, research preview | ceo-request | Evaluate | Opus 4.7-powered. Generates polished designs, prototypes, slides, one-pagers from text. Exports standalone HTML + Canva/PDF/PPTX. Reads codebase for brand consistency. Fills gap our code-first skills don't cover: client one-pagers, pitch decks, onboarding collateral. Available now at claude.ai/design. |
| 2026-05-26 | Karpathy LLM Wiki — pattern technique canonique (raw/wiki/CLAUDE.md + lint) | video-analysis | Evaluate | Source RD-0017 (Teacher's Tech). Brief technique manquant de WO-0008 Maya. Architecture 3 couches + schéma CLAUDE.md ingestion/lint = produit livrable factory. Confirme direction maya-rag-wiki-pattern. Score 15/15 🔴. Action: greenlight WO-0008 Stage 1, ajouter /lint-wiki au toolkit-catalog. |
| 2026-05-25 | Talent AI-Layer starter kit — harnais standard embarqué par chaque digital talent (CLAUDE.md en couches, stop-hook auto-améliorant, skills scopés par paths, subagent explorer read-only, MCP codebase-search) | video-analysis | Evaluate | Convergence du lot RD-0007→RD-0016 (synthèse 2026-05-25). Thèse « harness > model » = position foundry. Cole Medin/RD-0016 expose 5 gaps factory + impl. réf. coleam00/helpline. Adjacent mais distinct de RD-0021 (Managed Agents), RD-0031 (toolkit-catalog bundle = 1 des 5 gaps), RD-0029/030 (skills). Net-new : packager le harnais COMME produit livré. Reframe produit fort → scorecard justifiée. |