What it is: A tactics video on avoiding Claude's session/usage limits (Pro, Max, API) so you can keep shipping without the cooldown wall.

How it works (typical Nate playbook for this topic):

• Diagnose which limit you're actually hitting — 5-hour rolling session window vs. weekly cap vs. per-model (Opus burns budget ~5x faster than Sonnet)
• Downshift models intentionally: default to Sonnet for routine work, reserve Opus for hard reasoning, use Haiku for cheap parallel tasks
• Split workloads across surfaces — Claude Desktop/Code on the subscription, route bulk/agent traffic to the API (or via OpenRouter) so they don't share the same quota bucket
• Stack accounts/plans where ToS allows (e.g., separate workspace plan for team automation vs. personal Pro)
• Use sub-agents and `/compact` aggressively in Claude Code to keep context small and avoid burning tokens re-reading state
• Offload non-reasoning steps (scraping, file I/O, formatting) to cheaper models or deterministic code instead of letting Claude do it
• Monitor usage proactively via the Anthropic console / `ccusage`-style tooling rather than discovering the limit mid-task

Tools & links:

• AI Automation Society Plus (paid) — Nate's paid community with full courses
• AI Automation Society (free) — free resources tier
• Podcast application
• Anthropic Claude (Pro/Max/API) — the thing being rate-limited
• Likely referenced: OpenRouter, Cursor, Claude Code sub-agents (`/agents`, `/compact`)

Caveat: Transcript wasn't available, so specifics above are inferred from Nate's usual format and the title's framing — treat the tactic list as the likely shape, not verbatim.

Why it matters for you: You run Claude Code heavily; knowing which limit is biting (session vs. weekly vs. per-model) and how to route work across Sonnet/Opus/API is the difference between shipping on Friday afternoon and staring at a cooldown timer.

What it is: A walkthrough of ditching Claude Code's ever-growing built-in system prompt and context in favor of a custom, minimal coding agent ("Pi") orchestrated by Archon, so you fully own your workflow, context, and tools.

The thesis:

• Claude Code started simple and moldable — now it's huge, opaque, and the system prompt shifts every release
• "Your context is not really your context anymore" — hidden tools, hidden prompts, unpredictable behavior
• The fix: stop renting someone else's harness, build a thin one you control and plug it into Archon for knowledge + task orchestration

How it works:

• Pi is a lightweight, BYO coding agent — minimal system prompt, only the tools you explicitly give it, deterministic context window
• Archon acts as the orchestration + knowledge layer: RAG over your docs/repos, task/plan state, MCP server that any agent (Pi, Claude Code, Cursor, etc.) can hit
• Pi connects to Archon over MCP, so the agent stays dumb/thin and Archon supplies the brain (retrieved docs, task queue, project memory)
• You swap models freely (Claude, GPT, local) because the harness isn't married to a vendor
• Custom subagents, hooks, and slash commands live in your repo, not inside a vendor's opaque system prompt

Tools & links:

• Archon — Cole's open-source agent orchestration + knowledge base (MCP server, RAG, task mgmt)
• Dynamous — his paid community; the Archon workflow masterclass is hosted there
• Pi coding agent — framed as a minimal, open harness alternative to Claude Code (check Cole's GitHub / Archon repo for the reference implementation)
• MCP (Model Context Protocol) — the glue between Pi and Archon

What's different vs. his prior videos:

• Earlier Cole content leaned into Claude Code + Archon together; this one explicitly pushes *away* from Claude Code as the default harness
• Emphasis shifts from "prompt Claude Code better" to "own the harness, keep the context small, let Archon be the smart layer"
• Strong take that harness bloat is now a liability, not a feature

Why it matters for mx-workflow:

• mx-workflow is essentially a harness layer on top of Claude Code (agents, slash commands, hooks) — same category Cole is critiquing
• Actionable takeaways:
• Keep mx-workflow's injected context *auditable* — document exactly what each agent/skill adds to the prompt so you don't recreate the Claude Code opacity problem one layer up
• Consider an MCP-backed knowledge layer (Archon-style) so project-specific docs, PRDs, and decisions are retrieved, not stuffed into every prompt
• Decouple commands from Claude Code specifics where possible, so the same workflow survives a harness swap (Pi, Cursor, future tools)
• Watch for Anthropic system-prompt drift breaking your agents — pin behavior with explicit instructions in your skills rather than relying on CC defaults

What it is: A reaction/overview video on Anthropic's Claude Opus 4.7 release, framed around whether it's actually the new king of coding vs. GPT-5.4 and other frontier models.

How it works:

• Matt walks through Anthropic's published benchmarks for Opus 4.7 (SWE-bench, coding tasks, agentic eval scores)
• Compares head-to-head with GPT-5.4 and prior Claude models
• Shares his own take on which model he's defaulting to for coding going forward
• Light on hands-on demos — mostly benchmark commentary and vibes

Tools & links:

• Claude Opus 4.7 — Anthropic's latest flagship, pitched as best-in-class for coding
• FutureTools — Matt's AI tool directory

Why it matters for you: If you're building web apps with Claude Code or any Claude-backed coding agent, Opus 4.7 is the default model now — this is a quick sanity check on whether the benchmark hype is worth switching your daily driver for, without having to dig through Anthropic's model card yourself.