## Sources

1. [Stop writing data governance policies](https://andrewrjones.substack.com/p/stop-writing-data-governance-policies)
2. [I Used Claude Code And Codex Together, Here’s What Surprised Me](https://aimaker.substack.com/p/claude-code-vs-codex)
3. [[AINews] Fable and Mythos officially too dangerous to release](https://www.latent.space/p/ainews-fable-and-mythos-officially)
4. [Enrollment Is Open for Agentic Academy](https://aimaker.substack.com/p/claude-code-codex-course)
5. [5 Steps to Use AI in Sales Without Losing the Human Touch](https://aimaker.substack.com/p/ai-sales-workflow-trust)
6. [[AINews] not much happened today](https://www.latent.space/p/ainews-not-much-happened-today-7a8)

---

### **AI Sales Workflow: What to Automate and What to Protect** by **David Roy**

*   **Main Arguments**
    *   AI in sales should not replace human interaction but should be used to **remove administrative "drag"** so that sales professionals can show up more human in critical moments [1, 2].
    *   A fundamental rule for AI implementation is: **"AI drafts. Humans decide. Trust is the asset"** [3, 4].
    *   AI cannot fix a broken sales system; users must first establish a repeatable, successful manual process before attempting to automate it [3, 5].
*   **Key Takeaways**
    *   Sales processes should be audited to distinguish between **"Busywork"** (admin and pattern work safe for AI) and **"Trust moments"** (anything that changes the relationship and must stay human-led) [5, 6].
    *   AI acts as a "cheat code" for rapidly understanding unfamiliar industries, surfacing common pain points, and learning the specific language buyers use [6, 7].
    *   AI-generated communication must always be treated as a **first draft**, as relying on it for final versions can lead to generic messaging or factual errors that burn customer trust [8].
*   **Important Details**
    *   An effective AI sales stack includes tools for **call notes/summaries**, research and preparation, writing support to tighten drafts, and process hygiene [2].
    *   **Trust moments** include discovery calls, objection handling (especially regarding price or timing), negotiations, and handling emotional or high-stakes pushback from buyers [9].
    *   For call notes, a recommended prompt asks AI to list what the customer cares about most, objections raised, decision criteria, next steps, and unresolved items [10].
    *   A high-leverage research sequence involves asking AI to find company challenges based on website data and industry benchmarks before a meeting [7, 11].

### **Claude Code Course for Knowledge Workers: Cohort Open** by **Wyndo and Michael Simmons**

*   **Main Arguments**
    *   While AI chat is useful, it has a "ceiling" where every project requires manual setup, repeated instructions, and constant re-pasting of context [12, 13].
    *   The next productivity jump comes from **agentic AI work**, which involves building a structured environment (folders, rules, and skills) around your work so agents can function autonomously [14, 15].
*   **Key Takeaways**
    *   **Agentic Academy** is a 10-week live cohort designed to help professionals transition from simple chat to building working agentic setups [14, 16].
    *   The cohort focuses on learning the underlying patterns of agentic work—such as source libraries and correction systems—so the skills remain applicable as tools change [17].
    *   The program is specifically tailored for **creators, consultants, operators, and entrepreneurs** who feel the manual "drag" of setting up AI tasks repeatedly [18].
*   **Important Details**
    *   Enrollment includes an optional **setup session on June 15, 2026**, to help participants select projects and orient themselves around tools like Claude Code Desktop [16, 19, 20].
    *   Participants build a project "home base" that defines what good work looks like, a skill library for their specific tasks, and a **trust layer** with quality checks [21].
    *   A tool called **"Chat X-Ray"** is used to turn existing chat history into a starting point for agentic projects [20, 22].
    *   The course teaches patterns in Claude Code that are designed to be transferable to other tools like **Codex** [17].

### **Claude Code vs Codex: What I Learned Testing Both** by **Wyndo and Dheeraj Sharma**

*   **Main Arguments**
    *   Rather than treating Claude Code and Codex as rivals, users should treat them as **specialized agents** that can be coordinated through clean handoffs [23].
    *   Using two different agents creates a natural **review loop**, where one agent can inspect and improve upon the work of the other [24, 25].
*   **Key Takeaways**
    *   **Codex** is often the preferred tool for writing and interface design due to its "super-app" feel and built-in browser [26-28].
    *   **Claude Code** is typically favored for technical coding, brainstorming, and more complex agentic tasks [26].
    *   Successful multi-agent coordination requires a **shared handoff system**, such as a `changes.log` file, so both tools stay in sync [29, 30].
*   **Important Details**
    *   Codex features an **in-app browser** for live verification and a built-in image generator (GPT Image 2), whereas Claude requires Chrome to browse the internet [31, 32].
    *   A effective handoff note should specify what changed, why it changed, which files matter, what is unfinished, and what the next agent should avoid undoing [33].
    *   In a live demo, Claude Code added functional year-by-year breakdowns to a calculator, while Codex improved the UI and verified the growth chart in its browser [34].
    *   There is significant naming confusion between the tools regarding integrations, which are variously called connectors, plugins, skills, or apps [28].

### **Stop writing data governance policies** by **Andrew Jones**

*   **Main Arguments**
    *   Traditional data governance, which relies on central authorities publishing document-heavy policies, fails in **federated architectures** like data mesh [35, 36].
    *   To scale, organizations must stop writing document-based policies for humans and instead **codify and automate** them through the data platform [37].
*   **Key Takeaways**
    *   Data owners in decentralized systems lack the time and expertise to read and manually implement complex management standards [36].
    *   Automation provides the governance authority with higher confidence that standards are followed consistently across the entire organization [37, 38].
    *   **Data contracts** are essential for providing the automation with the necessary context (such as data categorization) to apply policies correctly [39].
*   **Important Details**
    *   An example of a codifiable policy is the requirement to delete or anonymize customer data after two years of inactivity [39].
    *   The source mentions the use of **ODCS** as a way for data contracts to provide context for anonymization strategies [39].
    *   The newsletter also references emerging roles like **Data Product Managers** and technical developments like autonomous web scraping with Claude Code [38, 40].

### **[AINews] Fable and Mythos officially too dangerous to release** by **Latent.Space**

*   **Main Arguments**
    *   In a major platform event, the US government abruptly forced Anthropic to suspend access to **Claude Fable 5 and Mythos 5** for all customers due to national security and cybersecurity risks [41, 42].
    *   This event has triggered a debate over **"model sovereignty,"** highlighting the risk that closed frontier APIs can disappear overnight due to geopolitical factors [42].
*   **Key Takeaways**
    *   Anthropic disputes the government's claim, suggesting the order was based on "verbal evidence" of a potential narrow jailbreak that Anthropic believes is a misunderstanding [43].
    *   Evaluation benchmarks are shifting; **Artificial Analysis** swapped SWE-Bench Pro for **DeepSWE** to reduce benchmark gaming, with Claude Code + Fable 5 currently ranking top [44].
    *   **Sandboxing** is becoming a core part of agent infrastructure to safely run untrusted, AI-generated code on Kubernetes clusters [45].
*   **Important Details**
    *   Moonshot AI released **Kimi-K2.7-Code**, a 1-trillion parameter MoE model that reportedly uses 30% fewer reasoning tokens [46, 47].
    *   NVIDIA released **DiffusionGemma**, a block-diffusion architecture that is roughly 4x faster than autoregressive models but prone to significantly more factual mistakes [48].
    *   A developer used Fable 5 to **reverse-engineer a 1989 DOS game** overnight, completing work that previously took six months [49].
    *   A new plugin called **Ponytail** adds a "lazy senior dev" mode to Claude Code, reducing generated code volume by approximately 84% [50].

### **[AINews] not much happened today** by **Latent.Space**

*   **Main Arguments**
    *   Anthropic reports observing early signs of **Recursive Self-Improvement (RSI)**, where AI is significantly accelerating its own development [51].
    *   Cloudflare’s acquisition of the **VoidZero** team (creators of Vite) signals a move toward a more "agent-friendly" application stack for developers [52].
*   **Key Takeaways**
    *   At Anthropic, **80% of merged code** is now authored by Claude, and engineers are shipping 8x more code per quarter than in previous years [51].
    *   **ChatGPT** has crossed **1 billion Monthly Active Users (MAU)**, approximately five months behind its original schedule [53, 54].
    *   A Stanford-linked study found that AI (specifically Gemini 2.5 Pro) outperformed law professors at answering contract law tutoring questions, winning **75.33%** of blinded comparisons [55].
*   **Important Details**
    *   NVIDIA released **Nemotron 3 Ultra**, a 550B parameter MoE model designed for long-running agent workloads, claiming it is up to 5x faster for such tasks [56].
    *   Google DeepMind released the **Gemma 4** family, which features an **encoder-free architecture** and supports multimodal input [57].
    *   **Agent Arena** launched to measure agentic performance from millions of live sessions; **GPT-5.5** currently holds the top ranking [58].
    *   Ideogram 4.0, a 9.3B text-to-image model, was open-sourced but received criticism for being heavily "safetymaxxed" and censored [59].