Resources
Guides, methods, benchmarks and datasets by topic. Each one says what it is for, how it works and where it stops.
Building products with AI agents
Running agent work through gates
A guide to running coding agents through a scheduler, a gate and a main check, so that one person still decides what goes live.
Governance rules for coding agents
A guide to a rulebook for coding agents: one entry file, a reading order, safety limits, a task for every change and a check that runs before the merge.
A guide to an eval suite for agent output: one decision boundary per case, self-contained prompts, adaptive repetition, changes on repeated evidence only.
Search intelligence
Measuring a search snippet by width
The method behind the search snippet preview: a title and a description are measured in pixels against the result's width and line limits, not counted.
Sizing and reading an A/B test
The method behind the A/B test calculator: a two-proportion z-test sizes a test from the baseline and the smallest lift worth detecting, then reads it.
Classifying search queries by topic
A method for classifying search queries with term lists: topic, intent and confidence per query, reportability apart from the match, and a review queue.
Ranking domain name candidates
A method for ranking domain name candidates: fluency, relevance, length and age in one score, availability first and a language model as a second opinion.
Building a competitor landscape corpus
A method for a competitor landscape corpus: seed landscapes, frontier expansion and deterministic seed selection; membership is discovery, not proximity.
Evidence separation rules for SEO data
A method for evidence classes in SEO data: an actual is not an estimate, a rank is not an average position, a goal is not an outcome, missing is not zero.
Blind evaluation of a domain-similarity model
A method for blind evaluation of a domain-similarity model: a stratified pair sample, blind labels, integrity checks, confidence intervals, disagreements.
A fail-closed market snapshot contract
A method for a market snapshot contract: a complete marker and run id, a hash-bound source set, immutable once verified, and a parity benchmark.
Forensics on defective SEO exports
A method for forensics on defective SEO exports: the state file checked against the bytes, defects classified before ingest and the raw export untouched.
Keyword taxonomy with QA gates
A method for a keyword taxonomy: signals before intent, location as a facet only, a person-name audit as diagnosis only and growing samples as the gates.
UX and content for large sites
A guide to demos with no server: one export per snapshot, allowlisted fields, a closed schema, a redaction scan and a page that works without the file.
Design tokens from JSON to CSS
A guide to design tokens: decisions in one JSON file, generated CSS custom properties, no raw colours, tested contrast pairs and vendored source tokens.