Prompt caching changed our unit economics.
When a 140 page handbook is generated against the same 20k token context forty times, the cache pays for the whole engagement. The math, with numbers.
W1 Writing
Essays on what happens when you run a
consulting firm with a model inside it.
Long-form writing on the actual mechanics of operating with Claude, not on whether AI will eat the world. We publish one piece a week in four themes, and we do not publish the ones that do not teach. You can subscribe by email below, or read it all here.
W2 Featured
The handbook thesis, defended.
Why we ship a long-form document at the end of every engagement, and why we believe the slide deck is about to die as a consulting artifact. Twelve thousand words. Citations. A companion sample you can read over lunch.
Essay
- 38 min read
- 2026-04-10
W3 Pillars
Four threads. One voice.
We publish in four themes. You can read them all or filter to the one you care about. The pillars correspond to how we think about service delivery: what the model can do, what the operator must do, how you sell that combination, and how you keep it safe.
Why we stopped writing one-shot prompts.
Every deliverable runs through four stages now: draft, self-critique, human edit, eval. How we broke the pipeline apart and why one-shot prompts always underperform at engagement scale.
The pentest report is a literary form, and most of them are bad.
Sixty percent of pentest reports we've inherited are unreadable to the executive and unactionable to the engineer. A taxonomy of failure modes, and what Claude specifically helps with.
We measured share of answer for 400 queries across 4 LLMs. Here's what moves the needle.
A structural audit of what actually gets cited in Claude, ChatGPT, Perplexity, and Gemini answers. Data, not theory. Includes the raw export.
Billing hourly in 2026 is a moral hazard.
When an operator plus Claude can do a week of classical consulting work in a day, the hourly rate incentivizes slower delivery. How we price fixed-scope and what it took to make that work.
Extended thinking is a research budget, not a toggle.
When to turn on extended thinking, when to pay for it in tokens, and how to know if your prompt actually needs it. With benchmarks from our own engagements.
The IAM attack path every Series A SaaS eventually has.
One policy, one role, one bucket - the same misconfiguration shows up at seventy percent of the Series A SaaS companies we audit. What it is, why it persists, and how to detect it in fifteen minutes.
Citations are the new backlinks.
In Google SEO, a backlink is a vote. In answer engines, a citation is a vote. The overlap is smaller than you think. A tactical guide to being the source that LLMs quote, not just the page they read.
A consulting firm's P&L when the model is a team member.
Our actual cost structure in the first year, broken down line by line. API spend, tooling, labor, rent we don't pay. Published so the next founder doesn't have to guess.
How we write templates that survive a model upgrade.
Our template library has been through three Claude versions without a rewrite. The principles that make a template portable across model generations, illustrated with our actual library.
Claude is not a pentester. It is a very fast junior analyst.
Claude will miss the exploits that matter if you let it drive. But it will correlate, summarize, and document faster than any human. How we use it without letting it run the engagement.
Why we stopped chasing Google featured snippets.
Featured snippets used to be the apex traffic prize. With ChatGPT and Perplexity absorbing the top-of-funnel, the snippet game is a loser's game. What we're doing instead.
W4 Subscribe
One long piece a week. No newsletter filler.
You'll get one email, every Tuesday, with one piece in one of the four pillars. No “interesting links,” no sponsor spots, no roundups. If we do not have something worth sending in a given week, we skip it.
W5 Editorial notes
A few things about how we write.
Since we are transparent about our operating stack, we will be transparent about the editorial stack too.
- 01Every piece is drafted by a named operator and Claude, and edited by at least one other human. Every piece is signed.
- 02We do not publish marketing dressed as journalism. When a piece is also an advertisement for an engagement, we say so at the top.
- 03We cite sources inline. Every quantitative claim has a footnote. Where our data is proprietary, we say so and offer the anonymized export.
- 04Corrections are logged at the top of each piece. We do not silent-edit. If the web is not forever, our commits are.