/resources · 103 entries · 8 shelves
Resources
The things I use, build and keep coming back to: skills, repos, tools, agents, prompts, reading, terms and experiments.
Each one says why it is here, and most link to where I used it.
Free to take
No sign-up. Download it, use it, change it.
- Claude Skills (free download)GEO / AI-Search Audit (Claude Skill)
Auditing a site for AI search (GEO/AEO) before or after a redesign, and checking that structured data never outruns what the page shows.
- Claude Skills (free download)Human Check (Claude Skill)
Checking posts, READMEs, cover letters and LinkedIn drafts before publishing.
- Claude Skills (free download)Crawler or Scanner? (Claude Skill)
Answering “was I hacked?” from access logs, and building honest server-side AI-crawler alerting.
- Claude Skills (free download)pgvector RAG Doctor (Claude Skill)
Debugging retrieval quality and latency on pgvector before blaming the model.
- Claude Skills (free download)LLM Cost Tracer (Claude Skill)
Knowing what each AI feature costs and noticing a dead key that a fallback is quietly covering for.
- Claude Skills (free download)Deterministic LLM Fallback (Claude Skill)
Assistants and RAG features that must survive outages, budget caps and key rotation.
- Claude Skills (free download)Honest Evidence Writer (Claude Skill)
Portfolio case studies, project write-ups, résumé bullets and launch posts.
- Claude Skills (free download)Frontend Perf Pass (Claude Skill)
Slow pages, weak Lighthouse/PageSpeed scores and launch prep.
- Claude Skills (free download)Design Token Enforcer (Claude Skill)
Rebrands, dark-mode leaks, theme switchers and design-system clean-ups.
- Claude Skills (free download)Ten-Persona Review (Claude Skill)
Portfolio and landing-page reviews, pre-launch checks and conversion audits.
- Claude Skills (free download)Award-Readiness Audit (Claude Skill)
Award submissions and honest, scored design critiques.
- Claude Skills (free download)LinkedIn Post Packager (Claude Skill)
Repurposing blog posts and case studies for LinkedIn without losing the click-through.
- Claude Skills (free download)CV from Content (Claude Skill)
A downloadable CV that stays in sync with a portfolio, plus an ATS-safe version.
- Design tokens / UI kitDesign tokens
Generated from the site's live Theme DNA, so the file always matches what you see here.
- Claude skill13 free Claude Skills
Each one is the checklist I wish I'd had before the problem it solves.
Skills
Claude Skills I wrote and give away, plus the wider list of tools and methods I work with.
See all (77) →- Frontend & interfaceAccessibilityPublic site components.
- AI quality & evaluationAI EvaluationContent scoring before publish and eval cases for Alex OS's reasoning surfaces.
- AI systems & retrievalAnthropic Claude APIAlex OS's model router calls Claude directly over HTTP for grounded answers, guidance generation and content review.
- Infrastructure & deliveryAzure DevOpsProfessional delivery workflows.
- Architecture & practiceClaude-Assisted DevelopmentDesigning and building Alex OS, Alex CMS, Theme DNA and the analytics and tracing layers on this site; the same loop at work.
- Infrastructure & deliveryCloudPanelManaging the production VPS.
Repos
Public code. Each one exists because I kept solving the same problem by hand.
See all (3) →- retrieval-evalI shipped a hybrid retriever on by default without measuring it. This harness showed it was the weakest of five configurations.
- multi-collection-ragI kept rebuilding the same ingestion and query code per project. This is the version I copy in instead.
- product-recommendation-agentA small, readable example of an agent as an explicit graph: each step is a node you can test on its own.
Toolkit
What is actually open while I build, and why it earned its place.
See all (4) →Agents
Agents I built: what goes in, what comes out, and how much they do on their own.
See all (4) →- Assistant / chatAlex OS site assistantVisitors ask the same questions in different words; this answers them from the source instead of a static FAQ.
- Workflow automationProduct recommendation agentTo build an agent as an explicit graph, where each step can be tested and swapped on its own.
- Workflow automationLinkedIn post drafterPosts with outbound links reach fewer people, so the idea goes in the post and the link goes in the comment.
- Workflow automationRelationship suggestionsThe links between projects, skills and notes are what make the site navigable, and keeping them up to date by hand doesn't scale.
Prompts & Recipes
Prompts and snippets I reuse, with when to reach for them.
See all (4) →- PromptClaim rules for an assistant that answers about a person or productUse when: Any assistant answering questions about a real person, product or company from retrieved content.
- PromptLinkedIn native post briefUse when: Repurposing a blog post or case study for LinkedIn without losing reach to an outbound link.
- Code snippet / recipepgvector: dedupe without killing the HNSW indexUse when: Whenever you need DISTINCT ON or GROUP BY over a pgvector similarity search.
- Code snippet / recipeVerify a crawler with forward-confirmed reverse DNSUse when: Before trusting a bot's user agent in analytics, alerts or rate limits.
Library
Reading that changed how I build, with the one thing I took from each.
See all (4) →- “A user agent is a claim. Forward-confirmed reverse DNS is the check, and it's what my crawler alerts use now.”Article · Verifying Googlebot and other Google crawlers
- “Markup has to describe what the page visibly shows. My FAQ existed only as JSON-LD, which broke exactly this.”Article · General structured data guidelines
- “An HNSW index only serves ORDER BY distance LIMIT on the table itself. Wrapping it in DISTINCT ON quietly turned it off for me.”Article · pgvector
- “It's the format I packaged my own checklists in, so the same steps run in the same order every time.”Article · Agent Skills overview
Glossary
Terms I use, explained plainly, each tied to a real example from my work.
See all (4) →- HNSW indexRetrieval & RAG
- A graph-based index that finds approximately nearest vectors quickly, without comparing the query to every row.Why it matters: If retrieval is slow, the first question is whether the index is being used at all.
- Forward-confirmed reverse DNSWeb
- Looking up an IP's hostname, then resolving that hostname back, to prove the IP really belongs to the domain it claims.Why it matters: It's how you tell a real crawler from a scanner using its name.
- GEO (generative engine optimisation)Web
- Making a site easy for AI answer engines to find, understand, trust and quote.Why it matters: More questions are answered by assistants before anyone clicks a link.
- Circuit breaker (for LLM calls)AI & LLMs
- A switch that stops calling a failing model provider for a while, so requests fail fast and fall back instead of waiting on retries.Why it matters: Without it, a dead key makes a feature slow as well as worse.
Experiments
Small tests I ran on my own work: the question, what happened, and what it does not prove.
See all (3) →- Does the shipped hybrid retriever beat simpler ones?→ Hybrid + rerank, the shipped default, scored worst: MRR 0.882 against 0.925 for semantic search alone.Comparison
- How much of the global CSS does the homepage use?→ 8.7% of the CSS was used. Pruning removed 116 KB of source; the built bundle went from 51 KB to 35 KB gzipped on every route.Benchmark
- Does the Human Check flag my own published posts?→ 0 of 68 posts flagged. The scanner is lenient on long technical prose.Eval