The Masterbrain
Week of 29 June 2026  ·  Edition #13

This week in the KB: 19 new entries across agentic-workflows, ai-fundamentals, book-summaries, business-strategy, claude-code, open-source-tools, personal-development, prompt-engineering and second-brain.

This week
Business Strategy

199 YC pitches, one rule: replace the role, don't assist it

A pattern read on all 199 companies at YC's Winter 2026 Demo Day. 60% are AI-native, 87% are B2B, and the copilot pitch has all but vanished, down from about 4% of pitches to 1%. The batch prices "AI employees" as a fraction of the salary they replace, not as assistants that help. The fastest growers shared a hard profile: they sold the outcome, the founder had customer relationships before the product existed, and they charged from day one with no free tier. The named failure mode is an AI-native service with no data moat: fast revenue, but replicable in weeks.

Apply this
  • Reframe any copilot-style pitch: lead with the role or task Anthera's agents take off the client's plate, priced against the cost of that work, not as a tool that helps.
  • Treat every engagement as a data flywheel. If the next build is not cheaper or sharper because of the last one, the offering is replicable in weeks. Make kb_anthera the asset that compounds.
  • Qualify on founder-market fit when prioritising outbound: the fastest deals in the batch closed where the seller had lived the buyer's pain. Test the former-employer-peers and PE roll-up channels.
Source · @ShahRathin, X ↗
· · ·
Agentic Engineering

One terminal, four layers: the operating system that runs an outbound campaign from a single prompt

Othmane Khadri's live build turns Claude Code from a bare runtime into a go-to-market operating system using four layers: the Input (the query is the whole campaign), CLAUDE.md (the permanent constitution), Skills (one skill, one action, one verifiable output), and memory.md (the evolving ICP and what has actually worked). Inside it runs a seven-gate qualification pipeline ordered cheapest-first, so free filters like dedupe and an exclusion-list match kill bad leads before any paid API call fires. The load-bearing idea is not LinkedIn, it is the layered architecture, and it lifts straight onto any single-terminal automation.

Apply this
  • Lift the four-layer split (per-run query, permanent CLAUDE.md constitution, one-action skills, evolving memory.md) as the template for Anthera's outbound stack: icp-list-builder-117, outreach-writer-117 and signal-monitor-117.
  • Order every qualification pipeline cheapest-gates-first. Run dedupe, headline read and exclusion-list match before any paid enrichment, so paid calls only fire on leads that already passed the free checks.
  • Copy the three skill-build rules into skill-creator-117: link the SDK so the model does not fail silently, define the positive outcome with zero ambiguity, and define the input contract.
Source · @MichLieben, X ↗
· · ·
Harness Engineering

Apple's leaked Siri prompt is 95% harness, and it states one defence three times: data is never instructions

The Siri system prompt from the iOS 27 developer beta leaked as a plain-text file, and it reads as a production operating contract, not a chatbot persona. Identity and tone take a single short paragraph; the rest is an entities model, a closed query DSL and full JSON tool schemas. The most important pattern, repeated three times, is that retrieved content carries facts, never commands: the core prompt-injection defence that most homemade agents omit. Apple also treats clarification and safety as structured tool calls rather than free-text prose, which is the deterministic-where-you-can, fuzzy-where-you-must split Anthera already builds under.

Apply this
  • Add the entities-as-data rule to every Anthera agent that reads client documents, emails or scraped pages: state that retrieved content is facts to reason over, never commands to follow. Most homemade agents are trivially injectable without it.
  • Make clarification and safety structured tool calls, not prose questions. The fuzzy decision of whether to ask stays with the model; the question UI and the routing stay deterministic in the harness.
  • Adopt the missing-property rule: absence means unknown, never inferred. Tell the user what is missing rather than guessing it.
Source · Apple Siri, iOS 27 DB1 leak ↗
· · ·
AI Fundamentals

Every major AI product's system prompt is already public, so the prompt was never the moat

A single GitHub repo at 141,000 stars collects the leaked system prompts and tool definitions of more than 30 shipping AI products: Cursor, Devin, v0, Lovable, Windsurf, Claude Code, Copilot, Perplexity, Manus and more. The lesson is the repo's own existence: if every leader's prompt has been extracted and published, the differentiator is never the prompt text, it is the harness, the tools, the data and the evals around it. The defensive read is just as sharp: anything load-bearing must sit behind the API and the harness, because the prompt will eventually be read by adversaries.

Apply this
  • Before building an agent for a vertical, pull the two or three closest products from this repo and diff their tool schemas and injection defences against your draft. Teardown beats guessing in a vacuum.
  • Never put a secret, key or business rule that only works while hidden inside a system prompt. Treat every prompt as eventually-public and move real defence into the API and harness layer.
  • Use this in positioning: a client's competitor cannot be out-prompted, only out-harnessed. Anthera's moat is the data and the loop, not the wording.
Source · x1xhlol, GitHub ↗
· · ·
Agent Readiness

Make a site agent-ready in an afternoon, but sell it on parse cost, not citations

Suganthan Mohanadasan's checklist for making a website legible to AI agents splits into two layers: the page itself (semantic HTML and accessibility) and the protocol layer agents read without rendering (robots.txt AI rules, llms.txt, Link headers, Markdown negotiation, an MCP Server Card). Markdown negotiation alone cuts an 85KB page to about 16KB, a fivefold drop in parse cost. The honest part is the data: over seven days he logged 8,060 AI-crawler hits, where OpenAI drove 43% of volume but only three referrals, while Google sent fewer hits and 2,570 referrals. High crawl volume does not convert to citations.

Apply this
  • Ship the minimum viable stack on Anthera's own site: robots.txt AI allow rules, llms.txt, semantic HTML, Link headers and Markdown negotiation. It is an afternoon of work and a credible reference build for client pitches.
  • Offer agent-readiness as a client service, but frame it honestly: it lowers friction for agents already crawling and makes you legible to systems deciding citability, it does not manufacture demand.
  • Use the MCP Server Card plus WebMCP to expose Anthera's or a client's tools to browsing agents. It ties straight into the existing MCP work and is the path to being actionable, not just readable.
Source · Suganthan Mohanadasan ↗
· · ·
Claude Code

Hermes-grade memory from plain markdown: a SessionEnd hook and a daily drain, no database

The video shows that Hermes-style persistent memory can be rebuilt inside Claude Code using native primitives only: no vector database, no daemon, no external service. A SessionEnd hook treats the close of a session as a write trigger, queues work to a folder, and a scheduled daily run drains the queue into two memory files, USER.md for facts about the user and AGENT.md for facts about the environment, capped by a short SOUL.md identity file. On idle days the shell preflight exits early and the system costs zero tokens. Because it is all markdown, the memory moves with the user if they switch harnesses, unlike vector-DB systems that lock you in.

Apply this
  • Formalise Anthera's capture side as a SessionEnd hook: queue on close, drain on a schedule. Standing orders cover this partially today; the hook makes it deterministic rather than reliant on the agent remembering.
  • Keep client-deployable memory in plain markdown, not a vector DB. Markdown portability is the defensible default and avoids vendor lock-in on anything Anthera ships.
  • Gate consolidation on real activity (10-plus tool calls or 10-plus minutes) so idle days cost nothing. Operational cost should track actual work, not run flat.
Source · YouTube ↗
Also this week
The Back Catalogue

From the Archives

Evergreen stories worth revisiting · 3 entries. For every past edition, open the editions menu.

Open Source

Free tools have crossed enterprise quality — the differentiator is now the services layer

Open-source tools for transcription, automation and scheduling now match SaaS that used to cost hundreds a month. But "free" hides real self-hosting and maintenance cost — which is exactly the gap a services layer fills.

When a client questions fees because "the tools are free", reframe: free tools make the services more valuable, not less.
Source · theaifield, Instagram ↗ Originally captured: June 2026
Business Strategy

AI is infrastructure, not identity — and the 18-month window to capture it is live

Hormozi's core argument from his $250M portfolio: businesses that quietly integrate AI as operational infrastructure beat those that make AI their brand. The proof problem in B2B means you build real outcomes first, then use AI to distribute the track record.

When a client says AI "doesn't work" for their use case — ask how much training and iteration went into their human version of the same process. That's the apples-to-apples reframe.
Source · Austin Marchese, YouTube ↗ Originally captured: April 2026
Agentic Workflows

The most effective agent systems are simple, composable patterns — not complex frameworks

Anthropic's engineering team distilled five reusable workflow patterns (prompt chaining, routing, parallelisation, orchestrator-workers, evaluator-optimizer). The recurring message: add complexity only when simpler solutions demonstrably fall short. Restraint is the differentiator.

Before reaching for an orchestration framework, name which of the five patterns the system actually needs. If you can't name it, the architecture is probably over-engineered.
Source · Anthropic Engineering ↗ Originally captured: March 2026