🤖 AI Weekly Recap (Week 36)

Covers the whole org chart: Content creation, marketing, design, sales, ops, and productivity, including things like email triage, daily briefings, SEO fixes, churn warnings, and influencer outreach.

Agent-agnostic: The prompts are plain markdown, so they work with Grok Bot, Rakazo, or basically any agent that can connect tools and run on a schedule.

Fully open and forkable: The repo is public, there is a JSON API and a raw feed for mirroring the catalog, and adding your own bot is a pull request.

Explicit meshes, not splats: Output is textured, instance-level assets rather than video, Gaussian splats, or a panoramic illusion, so scenes drop straight into Blender, Unity, or Unreal without conversion.

A pipeline, not a model: Three stages (intent analysis, global terrain generation, regional object placement) orchestrated by Claude Opus 4.8, using GPT-Image-2, SAM3D, and Hunyuan3D as specialist tools inside Blender.

Extras for real pipelines: Each scene ships with depth maps, normal maps, and instance maps, plus a render-inspect-refine loop that catches scale and object-terrain contact errors.

14 expert personas: Not "you are a developer" but "you are DHH." The roster models Bezos for strategy, Munger for pre-mortems, Werner Vogels for architecture, Seth Godin for marketing, and Ben Thompson for research, plus 30+ loadable skills.

One markdown file as memory: No vector database. Cross-cycle state lives in a single consensus.md that gets read at the start of every cycle and rewritten at the end, passed like a relay baton.

Forced convergence and guardrails: Cycle 1 brainstorms, cycle 2 does a GO/NO-GO, cycle 3 must ship. Discussion-only loops are banned, and hard constraints block repo deletion, force pushes, and credential leaks.

Your whole Ollama library, in Claude: Qwen, DeepSeek, Kimi, GLM, and the rest show up as selectable options, running either locally on your machine or through Ollama Cloud.

Nothing leaves your laptop: Telemetry is off by default and Ollama enforces a zero data retention policy across local and cloud, so a local model means your prompts never touch a server.

One switch, fully reversible: Toggle "Use Ollama models" on in the Ollama menu bar, toggle it off to restore your normal Anthropic setup. Existing features like web search keep working.

Efficient at scale: 770B total parameters with only about 49B activated per token across 78 layers, using Gated DeepSeek Sparse Attention and a built-in MTP layer for speculative decoding.

Terminal Bench leader: Scores 85.4 on Terminal Bench 2.1, surpassing DeepSeek V4 Pro, and jumps from 28.0 to 64.3 on DeepSWE. In a blind eval of 203 engineering tasks judged by 163 experts it hit 2.99/4, ahead of Kimi K3 (2.94) and GLM-5.3 (2.92).

Cheap to run hosted: Weights on Hugging Face, ModelScope, GitCode, and CNB, with FP8 variants, plus API access via Tencent Cloud TokenHub and OpenRouter at $0.834 per million input tokens.

25+ integrations plus your machine: GitHub, Slack, Jira, Notion, Linear, HubSpot, Outlook, Gmail, Google Calendar, monday.com, and any MCP server you point it at, alongside your local files and terminal.

Any model, including free local ones: Bring your own key for OpenAI, Anthropic, or Google, or run fully local and free with Ollama. The agent loop, conversations, and keys stay in a local secret store.

Approval gates on anything risky: Every tool call is sorted into read, write_local, exec, or external tiers. You approve a tier once and the agent works inside those guardrails, with full logs. It can also run on a schedule or fire on a Slack mention or incoming email.