Perpenda — Decision-grade fluency in LLM systems

The thesis

A trade-off, not a topic.

Most AI-fluency content treats PMs like coding-bootcamp students. Perpenda treats you like the decision-maker you are. Every unit teaches one concept through the same three-question scaffold — because that's how product decisions actually get made.

§ When this matters

Which feature, which call, which Monday.

Each unit names the concrete decisions the concept changes — cost forecasts, latency budgets, vendor comparisons, scoping calls. Not "important to know," but "load-bearing for these decisions."

§ When this breaks

The way most teams get it wrong.

Failure modes, named. Estimating cost in words instead of tokens. Treating latency as one number. Shipping streaming UX where output is consumed atomically. The mistakes you can recognize after, but ought to recognize before.

§ What it costs

What you'd have to give up to be right.

The discipline the right call demands — naming three latency metrics instead of one, asking eng for benchmarks on your prompt shape, measuring before promising. Cheaper than the failure mode it prevents.

The path

Twenty units. One curated order.

One path, in v1. LLM Systems for PMs. Fifteen units are published; the last five — the operating-phase units — lock from real closed-beta signal rather than from the armchair.

§ 01

Tokenization

Settled

Published

§ 02

Context window

Settled

Published

§ 03

Latency

Settled

Published

§ 04

Evals

Settled

Published

§ 05

Model selection

Contested

Published

§ 06

Prompt design basics

Settled

Published

§ 07

Hallucination + reliability

Contested

Published

§ 08

Cost dynamics at scale

Contested

Published

§ 09

Fine-tuning vs. prompting vs. RAG

Contested

Published

§ 10

Vector search / RAG fundamentals

Contested

Published

§ 11

Streaming UX

Contested

Published

§ 12

Tool use · function calling

Contested

Published

§ 13

Multimodal · vision basics

Contested

Published

§ 14

Agents · multi-step reasoning

Contested

Published

§ 15

Safety + content moderation

Contested

Published

§ 16–20

Operating units — monitoring · vendor risk · A/B · fallbacks

Locked

From beta signal

The app

A briefing, not a course app.

Two core screens — the path home and the unit reader. Editorial typography, hairline rules, no decorative imagery, no progress bars within a unit. The visual restraint is the design stance — and it makes the writing earn its space.

01 · Path home

The next unit, one tap away. Optional spaced review surfaces alongside the path — never as a gate.

02 · Unit reader

One scroll. Same order, every unit. Definition, bite, framing, depth, prompt, calibration, sources.

Calibration

Every claim, sourced and tagged.

Each unit's claims carry a tier: settled, contested, or unsettled. Sources sit after the decision prompt — never before — so the consensus doesn't prime your answer. The grader will tell you when it doesn't have enough signal to grade fairly.

Settled

Models bill and meter context windows in tokens, not characters or words.

Settled

BPE-style subword tokenization is the dominant scheme across frontier models.

Contested

Whether token-level pricing is the right unit of cost for product decisions long-term — vs. characters, semantic units, or compute-time billing — is unsettled as the market evolves.

Unsettled

Whether next-generation small-model latency improvements will obviate the streaming-vs-blocking distinction.

Confidence as a feature, not a softening.

Built for the call you have to make on Monday.

Phase 3's first fifteen units have shipped and the grader regression sets hold their gate. Closed beta opens next. Leave an email and we'll write the moment there's something to use.

Join the beta Read the code

No streaks. No badges. No daily nudge.