The Masterbrain — Week of 17 August 2026

This week

Agentic Engineering

Prompt, skill, loop, routine: never build at the top rung, and never promote one you have not watched work

Nick Saraev spends six hours building five marketing systems live in Claude Code, and the reusable artefact is none of the builds. Every automation climbs the same four rungs, prompt to skill to loop to routine, and a rung is only attempted once the one below it has run cleanly. He backs the discipline with arithmetic rather than preference: three chained steps at 80% reliability yield 51%, not 80%, so the fix for a flaky pipeline is removing steps from the model's span of control, not better prompting. The last forty minutes carry what most sources skip, which is how these systems break in production. Drag-and-drop platforms handed you execution history and failure alerts for free, agent routines give you none, and the minimum replacement is a dedicated errors channel that every routine posts to.

Apply this

Promote a workflow one rung at a time: freeze a prompt as a skill the moment you say it twice, and schedule it only after the skill has run cleanly several times unattended.
Split the pipeline so a human picks the process and judges the output, and the model executes only the middle. Letting it do all three collapses reliability to roughly half.
Give every box routine an errors channel and a line telling it to post service, timestamp, error and attempted fix. A 5:59am failure discovered at 8am has already cost someone their morning.

Source · Nick Saraev, YouTube ↗

· · ·

Claude Code

Under a prompt cache the cheapest move is to keep talking, and the most expensive is to leave and come back

The 20x headline survives a check against the API reference. Cache reads cost 0.1x base input and cache writes cost 2x at the one-hour TTL that subscription plans run on, so a 500k-token conversation rewritten cold costs about ten dollars against roughly fifty cents warm. That inverts the usual intuition about spend: a long session is cheap to continue, and a one-word message after lunch is not. Time is only one trigger. Switching model, changing effort level, toggling fast mode, connecting an MCP server, denying a tool, upgrading Claude Code and compacting all reset the cache, which means compacting to save tokens is not free. His context-hygiene tip lands too, measured rather than asserted: around 40,000 tokens consumed at conversation start with nothing sent, after already running /doctor.

Apply this

Treat a mid-session model switch, MCP toggle or compact as a pricing event rather than a configuration change. Each one makes the next message pay a full cold write.
Reach for /clear before /compact when working in a codebase. The project files are the state, so most of the history duplicates something already on disk.
Run /doctor and get a number. Our own audit puts always-on context near 30k tokens, 8.8k of it skill descriptions alone, which is the same order as his figure from a different workspace.

Source · Chase AI, YouTube ↗

· · ·

Agent Skills

Delete the procedure, keep the brand: the model out-reasoned the steps and could not guess the palette

Boris Cherny's prescription is a standing ritual rather than a cleanup. Every six months, delete your CLAUDE.md, your skills and your hooks, and watch what the model does unaided, on the reasoning that a new model makes last quarter's corrections obsolete. Nate Herk ran that test on a duplicate repo and produced the partition rule the idea was missing. With the full stack loaded he got correct branding and the format the skill dictated. With nothing loaded he got no branding, messier formatting, and content reorganised into a better structure than the skill had ever produced. So the deletion target is the procedural how-to, and the residue worth keeping is everything the model has no way of knowing: brand palette, asset paths, file topology, tone. Intelligence is not transferable, and preference is not inferable.

Apply this

Before cutting a skill, sort it into two piles, how to do the task and what good looks like. Delete the first, keep the second.
Run the experiment on a duplicate rather than in place, and take a /doctor reading before and after so the verdict is a number instead of an impression.
Weigh where the advice comes from. A harness builder can afford a wasted run, which is a different trade when the output goes to a client under your name.

Source · Nate Herk, YouTube ↗

· · ·

AI Service Delivery

Businesses buy five boring things, and every one of them sells on arithmetic rather than technology

After several hundred client builds, Nate Herk's claim is that the sellable list is short: speed-to-lead, document processing, follow-up sequences, database reactivation, and internal reporting. The list is not the valuable part. Each one arrives with the worked maths that turns a build into a figure the owner can check, an accounting firm at 200 invoices a week carrying about $78k a year of manual handling, a webinar funnel moving from 4% to 10-12% on multi-touch follow-up. The opening question is sharper still: if 500 new clients showed up tomorrow, what breaks first. That makes the owner walk their own operation and name the constraint, instead of presenting a diagnosis they arrived with. His most honest line is that document processing often needs no model at all, because deterministic logic does not hallucinate, drift, or need eval loops.

Apply this

Open discovery with the breaking-point question, then price the constraint the owner names rather than the solution they walked in holding.
Build before-cost, after-cost and delta in the client's own figures. His numbers are self-reported and directional, so they frame the method, never the quote.
Use his framing with our numbers. The AU bands filed to the KB this same week put a first single-workflow build at $7,500 to $15,000, above the few thousand implied here.

Source · Nate Herk, YouTube ↗

· · ·

Design Systems

Anti-slop lists are cheap: the harder gate is making the next page structurally different from the last

Hallmark is a 24.7k-star MIT design skill for Claude Code, Cursor and Codex, and its banned-gradient list is the least interesting thing in it. The differentiator is a gate that fails a page for reusing the generic hero, three features, CTA and footer template, or for repeating the macrostructure of any previous page in the same project. Enforcing that needs memory, so the macrostructure is stamped into the CSS and logged to a project file. The second idea travels further. Before emitting anything, the agent scores its own planned output from one to five on philosophy, hierarchy, execution, specificity, restraint and variety, and anything under three triggers a revision pass. The score is left in the artefact so the next run can see where the last one was weak. Neither trick requires installing the skill.

Apply this

Lift the pre-emit critique into any skill producing a subjective deliverable: name the axes, score before emitting, revise below threshold, and leave the score in the file.
Point its study command at a client's existing site to extract a portable design.md, which turns brand rules into something downstream agents obey rather than read.
Do not install a fourth overlapping design skill. Impeccable, huashu-design and frontend-design are already here, and another one mostly adds trigger ambiguity.

Source · Chase AI, YouTube ↗

· · ·

Game Design

A goal is a pacing mechanism, not a win condition, and boredom lives wherever the player has none

Riot's URF Academy module on goals is a free teaching artefact, and two frameworks inside it travel well past games. The first is the pacing ladder: a player should hold goals at four timescales at once, this possession, this quarter, this game, this season, and any moment where none is in mind is where boredom enters. The second is a five-axis evaluation rubric, innovation, depth, complexity, cohesion and satisfaction, each rated separately with reasoning rather than argued as one verdict. The sharpest teaching is a deliberately broken goal, win by coming exactly second, which fails three ways at once: the optimal line becomes doing nothing, the wording is ambiguous, and the player cannot tell what playing well means. The module also defines objectively bad design as design that serves no audience, which retires the two laziest arguments in any review.

Apply this

State the goal, then name its optimal line. If that line is to do nothing, or you cannot answer without inventing a rule, the goal is broken however interesting it sounds.
Rate a design on innovation, depth, complexity, cohesion and satisfaction separately. The rubric exists to make the disagreement specific, not to produce a score.
Hold the mechanics fixed and swap only the objective, which isolates whether the interest comes from the rules or from the goal sitting on top of them.

Source · Riot Games, URF Academy ↗

Also this week

Own Work The go-forward pricing ladder was ruled on 14 August: free instruments, then a free scoping call, then a $2,500 flat audit credited in full against any build over $5,000 commissioned within 30 days, with the workshop rung held back because it hands the most delivery effort to the least committed prospect.
AI Pricing Three independent methods, Hormozi value doctrine, the KB's own value-based frameworks and AU market benchmarks, converge on the same band for a first SMB workflow build: $7,500 to $15,000, priced as 20 to 25% of audit-measured annual savings rather than off a sticker.
Business Strategy Verified against NAB, CommBank, MYOB, ASBFEO and CPA Australia, the top Australian SMB pain is cash flow at around 43%, but the doors re-rank on fulfilment fit and trust routing, which puts labour first and getting-paid fourth, because owners take a cash flow problem to their bank rather than to a technology firm they just met.
Reports A three-part channel sweep, with every figure labelled measured, vendor-asserted or practitioner, concludes that no organic mix reaches 50 to 100 assessment starts a week inside 90 days, and that Meta paid is the only channel that carries the target on its own at roughly $6 to $14 a start.
Reports Fourteen partner categories tested to one evidence bar after the bookkeeper pick was challenged as an availability artefact: bookkeepers hold as the anchor for a new reason, payroll bureaux join the first wave, and MSPs and accounting-software app advisors are rejected as competitors rather than referrers.
Own Work One hub note now links the whole lead magnet project, the ruled ladder, five research reports and the workspace briefs by path, and it is explicitly marked a candidate source of truth rather than adopted process.
Business Strategy The Offers, Leads and Money Models trilogy rebuilt as a followable seven-layer process, market up to measurement, each layer carrying a done-test that gates the next, on the observation that most operators optimise volume when the market underneath is the actual problem.
AI Service Delivery Freelancer to consultant before any agency, because a scoping mistake made solo costs one project and the same mistake multiplies once a developer or a salesperson sits between you and the work; the portable piece is the client filter, never take on someone using AI to plug a sinking ship. Source · Nate Herk, YouTube ↗
AI Service Delivery Mostly beginner outreach, filed for the 60/30/10 ratio: a working automation is about 60% deterministic logic, 30% AI-assisted and 10% human approval, which makes a 90% model design a sanity-check failure rather than an ambition. Source · Nate Herk, YouTube ↗
Claude Code The 2m34s Short of the token-cost video, filed minutes before the long-form and superseded by it the same day, which is the reminder that a duplicate check on video ID cannot catch a Short and its long-form pair. Source · Chase AI, YouTube ↗
Game Design Module one supplies the vocabulary the goals module is measured against, a five-part game design framework and the eight kinds of fun from MDA, on the argument that fun is too subjective to design for, so you pick which kind you are serving. Source · Riot Games, URF Academy ↗
Game Design Give randomness a memory by drawing without replacement, so a 25% chance becomes two red marbles in a bag of eight and the player never eats four misses in a row that the raw maths quietly permits. Source · candlesan, YouTube ↗
Personal Development Five life domains run as systems rather than in-the-moment calls, filed for the mechanism rather than the tactics: a system exists to remove the emotional decision, not to add willpower, and personal commitments go into the calendar before the business planning because work expands to fill whatever gaps it finds. Source · Ali Abdaal, YouTube ↗

The Back Catalogue

From the Archives

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

Context Engineering

Anthropic deleted over 80% of Claude Code's system prompt and the evals did not move

Anthropic cut over 80% of Claude Code's system prompt for the Claude 5 generation with no measurable eval loss, because guardrails written for weaker models had become conflicting instructions the model burns reasoning to resolve. Context engineering is now a tree of on-demand files, not one exhaustive document.

Delete every absolute rule in your context files that cannot be traced to a failure you actually observed.

Source · Anthropic, claude.com ↗ Originally captured: August 2026

Model Economics

Opus 5 becomes the value pick: better than Fable 5 on agentic work, at roughly half the price

Claude Opus 5 landed at Opus 4.8's price while beating Fable 5 on agentic coding, computer use and business workflows, at roughly half the cost per task. The improvement that matters most is self-verification, which is what carries an agent loop end to end.

Re-test the default executor on verify-heavy loops rather than assuming last quarter's model choice still holds.

Source · Nate Herk, YouTube ↗ Originally captured: July 2026

Harness Engineering

Attention is the new bottleneck, not tokens or compute, so build tools that keep you out of the loop

OpenClaw's Peter Steinberger argues the scarce resource in agentic coding is now operator attention, so every tool is measured by how much of it it buys back. The move is to extend the loop through verification, not just generation.

Extend every agent loop to cover its own verification, and surface only the genuinely ambiguous decisions to a human.

Source · Peter Steinberger (Steipete), YouTube ↗ Originally captured: July 2026