36 artifacts at this level
- Architecture · L2
My Actual CLAUDE.md — Annotated and Anonymised
The real rules file, line by line, with the reasoning behind each block. - Architecture · L2
The Folder Structure That Runs a Production AI Agent
How the filesystem is laid out, and why every directory earns its place. - Architecture · L2
The Library Behind The Agent
123 books, a keyword lookup instead of a vector store, and a citation log that measures whether a book changed the method — not whether it got quoted. - Downloads · L2
Brainstorm Engine — Skill Skeleton
A divergence-then-evaluation brainstorming skill with one property most idea-generation tools skip: it forces a decision at the end instead of stopping at a ranked list. - Downloads · L2
Decision Authority Matrix
A table that answers, in advance, which actions the agent takes alone. - Downloads · L2
SKILL.md Template
The shape of a reusable capability: when to fire, when not to, what it does, how it fails. - Failures · L2
The Anonymiser That Passed
A deny-list cleared an article for publication. A second check, written the same afternoon, found the article identified its author anyway. - Failures · L2
The Invitations That Almost Went Out
Writing a past meeting into a calendar is harmless. Writing it with attendees is not. - Failures · L2
The Margin Computed From The Wrong Cost
A clean-looking margin figure, built on a cost base that excluded two real components. - Failures · L2
The Queue That Was Empty For Twenty-One Days
A triage queue reported nothing to do for three weeks. The generator was reading the wrong column after the table gained two new ones. - Failures · L2
The Tool That Outvoted The Screenshot
A human sent evidence of a failure. The agent kept quoting a script that said everything was fine. - Failures · L2
Two Hundred And Forty Dead Links
A convenience folder is cleared on a schedule. Its paths had been written into permanent notes for months. - Log · L2
Dev Log #006 — The Brainstorm Tool That Ran On Itself
Nine runs, 604 ideas, zero decisions. The fix wasn't more ideas — it was a hard gate the tool couldn't skip. - Log · L2
What "Briefly" Should Not Switch Off
A request for brevity suppresses formatting. It must not suppress a warning. - Patterns · L2
"Zero Problems Found" Means Zero Of What You Measure
A freshness monitor reported all clear while the pipeline behind it was failing. The monitor was right about the only thing it could see. - Patterns · L2
Adding A Rule Without Removing One
Every fix adds a rule. No fix removes one. After a year the instructions are longer than anyone reads. - Patterns · L2
Asking Approval For The Reversible
A gate meant for irreversible decisions gets applied to everything, and the person it protects stops reading it. - Patterns · L2
Drawing Conclusions From Truncated Output
A command printed the first twenty lines and the summary described all of them. - Patterns · L2
Empty Is Not Zero
A weekend of missing exports read as a sales collapse in exactly the window a marketing cut was being tested. - Patterns · L2
Maximum Effort By Default
If the model has a hardest setting, the temptation is to leave it there. It costs more and reasons worse. - Patterns · L2
One Source Is A Hypothesis
A single confident source becomes a stated fact somewhere between reading it and writing it down. - Patterns · L2
Revenue As The Metric
The number everyone reaches for is the one that cannot tell a good week from a bad one. - Patterns · L2
The Decorative Citation
A framework quoted after the conclusion was already reached changes nothing except how the answer looks. - Patterns · L2
The Fallback That Lied
A tool call failed, the agent produced a confident answer anyway, and nobody noticed for a week. - Patterns · L2
The Green Light Nobody Owns
A sign-off said "verified". It was produced by the same process that ran one narrow check. - Patterns · L2
The Header That Counts Itself
A file's summary line claimed twenty-five entries. The file contained twelve. Neither number had been wrong when it was written. - Patterns · L2
Treating llms.txt as a Distribution Plan
Shipping the machine-readable layer is cheap and worth doing. Expecting it to bring readers is a different claim, and the evidence is against it. - Patterns · L2
Two Sessions, One Identifier
Two runs allocated the same record number minutes apart, because both read the next number from a header instead of the data. - Patterns · L2
Year To Date Is Not A Year
Two windows, both labelled annual, that disagree by an entire season. - Playbooks · L2
A Decision Authority Matrix In Markdown
Which actions an agent takes alone, which it proposes, and which it never touches — written as a table rather than encoded in prompts. - Playbooks · L2
Degraded Mode: When The Source Of Truth Is Down
What an agent should do when the system holding the authoritative data is unavailable — and what it must never do. - Playbooks · L2
Reviewing The Agent's Own Session
A structured retrospective on the agent's behaviour, run against its own rules, producing a score and a ledger entry. - Playbooks · L2
Scheduled Tasks That Fail Loudly
An unattended job that fails quietly is worse than no job, because it leaves a belief behind. - Playbooks · L2
The Uncertainty Gate: Ask Once, Learn Forever
A rule for when an agent should stop and ask — and what it must do with the answer so it never asks again. - Playbooks · L2
Treating Untrusted Content As Data
Email, web pages and reader submissions reach the agent as text. Here is how to keep them from reaching it as instructions. - Receipts · L2
Receipts #1 — One Month In Ratios
Operating numbers for the agent itself: rules, agents, skills, repeat failures and what share of them got a fix that runs.