Two tiers, honestly split

The basics are free and open source. The paid edition contains only what we paid to learn, in production.

FREE · OPEN SOURCE

The skill: rules your agent follows by itself

Seven verification rules plus four runnable scripts, installed as a Claude Code skill. Once installed, the agent runs positive controls before trusting a pass, and marks work complete only once the underlying action has actually succeeded.

npx skills add ppcvote/ultralab-skills

Source on GitHub

MIT licensed. The skill is the condensed version of this manual's core rules.

Prefer a book? The starter edition, free

Install and first steps, the five dead ends, CLAUDE.md, remote control. Emailed as a PDF.

PRODUCTION DISCIPLINE

110 pages + runnable toolkit

Five verification disciplines, each with runnable code and the incident that taught it
The prompt/structure boundary: prohibitions take a prompt; obligations take structure
Four safety valves for forced continuation, and what breaks when you skip one
Silent failure: five ways a healthy-looking service lies, and a fixed four-step investigation
One complete bug investigation, from a bug report with no clear symptoms to a verified fix
Self-audit tool you point at your own repo in the first hour

Honest disclosure: the core boundary in 6.1 appeared on our public blog (in Chinese). You pay for 6.2 onward: the full implementation, four complete scenarios, and the toolkit.

7-day no-questions refund

If it is not useful to you, reply to the delivery email within seven days and we refund. No reason needed, and you keep the files.

This book teaches you not to trust anything unverified. We are in no position to ask you to trust us first.

Questions before you buy? Email minyi@ultralab.tw. Real inbox, real person.

Excerpt: read before you decide

Judge the content density, not the copy.

Where prompts work, and where they fail

What you want	Attempts at the prompt layer	What finally worked
Do not fabricate customer data	One rule added	The rule was enough
Never leak internal IDs	Explicitly forbidden; leaked anyway	Server-side string filter
After comparing the data, finish the review	Three prompt rewrites; it still stopped midway	A state flag blocks the wrap-up, a system message forces one more round

A prohibition is a single-point decision. An obligation spans turns, and any turn that decides to report progress first breaks the chain, in a way that reads as perfectly reasonable.

The positive control that saved us once

We ran our own verification script against our own site, and it aborted on the first try. A single-page app returns HTTP 200 with a full HTML page for routes that do not exist, and that page happened to contain the string we were checking for. The check would have passed against any URL, including ones that do not exist.

Without a positive control, we would have trusted a check that could not fail.

Contents of the paid edition

Reliability is layered, not a single point

Incident: two processes stealing messages

Incident: done, but never sent

Discipline 1: verify along the user's path

Discipline 2: every check needs a positive control

Discipline 3: mark complete only after completion

Discipline 4: count outcomes, not effort

Discipline 5: find every consumer before changing

Catching alive-but-producing-nothing

The prompt/structure boundary

Four safety valves of forced continuation

Four full scenarios: report job, quoting, streaming, batch

It fabricates numbers, and forbidding it does not work

When a test fails, suspect the test first

The silent-failure investigation, in fixed order

Walkthrough: one real bug, start to finish

Postscript: the author does it twice anyway

Appendix: checklists, symptom lookup, self-audit

The toolkit

Every code block in the book has a complete runnable counterpart. Change the constants and run.

Script	What it prevents
self-audit.mjs	Finds instances of this book's patterns in your own repo, with file and line
positive-control.sh	Your verification tool silently always passing
number-guard.mjs	Fabricated numbers; 14 built-in self-test cases across English and Chinese
heartbeat-check.sh	The monitor itself dying, taking the alerts with it
walkthrough.mjs	Code that is correct on every line and dead as a whole
mirror-check.py	Done, but never sent (Stop hook)
must-finish-guard.py	The four safety valves as a Stop hook: no wrap-up until the obligation is met
zombie-killer.ps1	Two stale processes stealing your messages
CLAUDE.md / STATUS.md	Re-explaining your project to every new session

Each ships with scheduling instructions and a positive-control procedure, because installing something and having it actually run are two different things. Expect the audit tool to surface one or two real findings on a typical repo, not a long list; that is by design.

Who this is not for

Written last, because we would rather sell fewer copies to the right people.

Skip it if you

Want a one-click AI-builds-my-product shortcut
Never ship agent output to the real world
Expect a comparison of every AI editor
Expect every code block to drop in unchanged (without a backend, the main examples read as theory; a dedicated section maps them to plain Claude Code)

It is for you if

Your agent keeps finishing things that are not finished
Output reaches production: publishing, email, deploys
You run a one-person company or a small team with no dedicated ops
Failures cost you money or trust