Book · August 10, 2026
S-Tier
Most advice about building with agents stops at prompting. This goes past it: what a skill is mechanically, how an agent decides to load one, how to run several without them colliding, what counts as evidence, and what separates a builder who produces output from one who builds systems that keep producing verified output without them.
Every claim in it points at a file in the public repo, so you can check the ones you doubt. Checking claims is most of what the book argues for, so please do. You do not need the pack installed to read it.
Start reading: Chapter 1 Read it as one Markdown file Download the PDF (90 pages)
Contents
Part I
The Machinery
What a skill is, how it is built, and how an agent decides to load it.
The Competence Gap
A capable model plus a vague prompt produces confident mediocrity. The gap is not intelligence, it is procedure, and procedure is now a file you own.
Chapter 2 · 10 minAnatomy of a Skill
A real SKILL.md taken apart: frontmatter, routing description, body, bundled references and scripts, and the advisory-gate policy that keeps a grader from blocking you.
Chapter 3 · 7 minThe Description Contract
The description field is the router. Trigger words, NOT FOR boundaries, overlapping descriptions, and what a dead route costs inside your agent.
Part II
The Operating System
Outcome-bound work, agent lanes and worker fleets, and the verification discipline underneath both.
Full Send
Outcome-bound work: freeze the mission before you mutate anything, fill useful non-colliding lanes, and bring back a decision instead of a workshop.
Chapter 5 · 8 minLanes, Fleets, and Collisions
Multi-agent work that is not theater: disjoint file ownership, isolated worktrees, atomic leases, and the measured cost of a fan-out that should have been one agent.
Chapter 6 · 7 minEvidence or It Did Not Ship
The verification discipline the rest of the book rests on, built from three real failures: a truncated test log, a page that passed a colour check while rendering unreadable, and a test file serving traffic in production.
Part III
The Craft Lanes
Grading code, treating design and copy as systems, distribution, and the last mile.
Grading Your Own Work
Seven scored lanes, instant-F triggers, grade caps on auth and payment surfaces, and the actual skill: predicting the grade before you run it.
Chapter 8 · 7 minDesign and Copy Are Engineering
Tokens instead of hex codes chosen in the moment, numeric thresholds instead of opinions, and an anti-slop gate for prose that reads fine sentence by sentence and hollow paragraph by paragraph.
Chapter 9 · 8 minGetting Found
Distribution as an engineering surface: the shift from ranking to being cited, nine audit lanes, an A-F visibility grade, and why a page that cannot be quoted will not be cited.
Chapter 10 · 8 minShipping Into the Real World
The last mile: App Store and Play Store gauntlets, release evidence gates, and a procedure that recovered $448.31 by running outside a repo entirely.
Part IV
Becoming S-Tier
The ladder, the judgment layer, authoring your own skills, and ninety days of practice.
The S-Tier Ladder
Five tiers defined by observable behavior, with the week, the transcript fingerprint, the failure that holds you there, and the move that promotes you. Most readers place themselves one rung high.
Chapter 12 · 11 minTaste, Judgment, and Knowing When to Stop
The layer no procedure supplies: which decisions are yours, why a bug fix that becomes a refactor is a judgment failure, and how to tell finished from merely different.
Chapter 13 · 8 minWrite Your Own
Authoring skills: when to write one, what goes in the body versus references, the competent-stranger test, and how to lint an estate before a dead route bites.
Chapter 14 · 9 minThe First Ninety Days
A real practice plan in three phases, a list of anti-goals, two honest measurements, and the close: never end your allocation above zero.
Appendices
Reference
The pack indexed by intent, and every rule the book argues for on one page.
The Skill Index, by Intent
Every public skill in the pack, grouped by what you are trying to do rather than by what it is called.
Appendix B · 3 minThe Rules, on One Page
Everything the book argues for, compressed to 29 rules across evidence, decisions, scope, parallelism, skills, craft, and stopping.
About the prose
The pack ships suede-deslop, a skill that strips AI writing patterns out of text before it goes public: filler openers, manufactured enthusiasm, false agency, formulaic contrast, metronomic rhythm, and em dashes. Publishing a book that tripped every rule in it would have been a poor advertisement.
So the book was written against those constraints, and the prose carries no em dashes. Quoted repo material keeps its own punctuation, because editing a quotation to fit a style rule is the kind of small dishonesty this pack exists to catch.
Source lives in book/. Skill counts written into the prose are guarded by the repo's validator, so a chapter that goes stale fails the same way a stale meta tag does.