Guides
What I've learned building with Claude
These come from my second brain: the notes, rules and logs behind every project I run. Three guides are free to read here. The full pack has all 66, as a PDF, for whatever a coffee is worth to you.
Free to read
- Claude Code system
Hooks for the rules that must never break
Move the few rules that must hold every time out of CLAUDE.md and into Claude Code hooks, small scripts the program runs before and after each tool call, and read a repo's hooks before you trust it.
7 min read - Claude Code system
A CLAUDE.md that changes behavior
Write a short CLAUDE.md made of things Claude can't guess and can be checked (how to ship, the rules, the current state, your preferences), test it against no file at all, and move anything that must always happen into a hook.
7 min read - SEO and AI search
Rank for your own name
Tie your site, your photo and your real profiles to one person in structured data, prove you own the site, and give search engines pages with enough depth to show for your name.
8 min read
The full pack
66 guides across 12 topics, including the three above. Pay what you want: $5 buys me a coffee, and any amount unlocks it.
Gumroad takes the payment and sends you the PDF and your library code. Nothing on this site tracks you.
- Claude Code system13 guides
- Second brain5 guides
- Building with the Claude API9 guides
- Paper-trading rails5 guides
- Shipping a site6 guides
- Products built with Claude4 guides
- SEO and AI search3 guides
- Design and UI7 guides
- Selling online6 guides
- Advertising3 guides
- AI for local businesses3 guides
- Security and compliance2 guides
Written from my own repos and notes. No client data, account details or employer information. Not financial advice.
Every guide in the pack
All 66 titles, grouped by topic. The free ones open right here; the rest unlock with the pack.
Claude Code system 13
- One rules file for every projectKeep the rules Claude must follow in one RULES.md in a parent repo, import it into every project's CLAUDE.md, and check nightly that no copy has drifted.In the pack
- A hub that remembersGive every Claude Code session a shared memory across all your projects with an index table, an append-only log, a status file, and goal checks that only speak up when something breaks.In the pack
- Handoff tickets and routines that run while you sleepPass work between Claude Code sessions with one-file tickets, and write scheduled routines that finish unattended, log what failed, and never push more than they should.In the pack
- Design, plan, build: approvals that stickMake Claude show you a design, then a plan, then build on a branch, and record every yes with your exact words and a date so nothing goes public without one.In the pack
- Hooks for the rules that must never breakMove the few rules that must hold every time out of CLAUDE.md and into Claude Code hooks, small scripts the program runs before and after each tool call, and read a repo's hooks before you trust it.Free
- Subagents for the readingHand broad reading and searching to subagents with a tight brief so your main session keeps only the conclusions, skip them for small lookups, and check what they bring back before you build on it.In the pack
- A CLAUDE.md that changes behaviorWrite a short CLAUDE.md made of things Claude can't guess and can be checked (how to ship, the rules, the current state, your preferences), test it against no file at all, and move anything that must always happen into a hook.Free
- Skills you write and skills you reviewWrite your own Claude Code skills as small SKILL.md folders with a description that triggers and references that load on demand, and never install someone else's skill, plugin or MCP server until you've read it, scanned it, pinned it and logged it.In the pack
- Permissions without the promptsCut Claude Code's approval prompts by allowing the read-only commands you run all day, denying secrets outright, asking before anything that can't be undone, and adding the sandbox for what rules can't see, instead of switching permissions off.In the pack
- Keep the context smallClaude Code resends the whole conversation with every message, so keep big tool output and file dumps out of it, read files in parts, start a fresh session per task, and carry what matters forward in a short handoff log instead of a long chat.In the pack
- Work with your PC offRun every project from cloud Claude Code sessions that clone fresh, push from the cloud, keep keys on the host and prove each change with screenshots and live endpoints.In the pack
- Scheduled tasks that report instead of actBuild checks and routines that run twice a day or every night, write a ledger, alert only when something changes, open tickets instead of fixing, and push only the paths you allowed.In the pack
- Connect tools with MCP without handing over your keysAdd MCP servers and connectors one at a time, from known publishers, reviewed first, with narrow tokens, read-only by default and every write behind an approval.In the pack
Second brain 5
- Every repo is one vaultOpen the folder that holds all your Git repos as a single Obsidian vault, keep the junk out, and read every project's notes on your phone.In the pack
- Let Claude write your notes (and show them off safely)Tell Claude Code what to record and where, keep the notes linkable, and publish a picture of your vault that shows its shape and none of its contents.In the pack
- Labels that write themselvesA nightly script groups every research note under a topic and subject labels, writes an index page per label with evidence grades and a "waiting on you" list, and flags summaries that have fallen behind.In the pack
- A research radar that grades claimsTurn short videos and creator posts into notes that quote the claim, check it against official sources, grade the evidence, score it, and wait for your yes before anything changes.In the pack
- Publish a map of your vault, not your notesTurn a private vault into a public, anonymised map of counts and links, with an allowlist guard on every string, content-hashed files and a nightly refresh that can only touch its own paths.In the pack
Building with the Claude API 9
- Meter every AI call on the serverKeep prompts, models and spending limits on your server so no user can turn your API key into an open tab.In the pack
- One dashboard, many apps, no shared keysHow to read several live apps from one hub through a server-side keyring, with allowlisted paths, read-only defaults and feeds that leave out anything identifying.In the pack
- Dashboards that don't lieRules and code patterns for dashboards where every number has a source and a time, missing data says UNKNOWN, and failures look like failures.In the pack
- Structured output you can trustGet JSON back from Claude that your code can use, and treat a cut-off, refused or malformed answer as an error with a clear message instead of as data.In the pack
- Web answers you can verifyUse Claude's web search to look up hard facts, and keep an answer only when the value appears in the page text it cited and passes a checksum.In the pack
- Pick the cheaper model on purposeMeasure what each Claude job costs per item from real usage, test a smaller model and fewer tools against the same cases, and keep the expensive setup only where it earns its price.In the pack
- Prompt caching and batches for costPut the unchanging part of your prompt first and cache it, send work that can wait through the batch API, and check the usage fields to prove either one is working.In the pack
- Small evals before you change a promptTurn real past failures into a small pass-or-fail test set, run it before and after every prompt or model change, and test the code around the model with a fake model endpoint.In the pack
- Truth gates after the modelRun a plain-code gate after every AI fill that pulls hard facts from the source text, overrides values that contradict them, blanks placeholders and guesses, and logs every change with its reason.In the pack
Paper-trading rails 5
- Rails before returnsHow I wire an AI-assisted trading agent so every order goes through one guarded function, starts on paper, and can always be stopped.In the pack
- Prove it on paper firstHow I keep paper results honest with separate ledgers, a hypothesis ledger, frozen calls, sample-size rules and a checker that assumes the report is wrong.In the pack
- Backtests that aren't overfitHow I test a trading idea so the backtest can actually say no, with a frozen pass bar, a sealed holdout, a trial counter, real costs, random baselines, walk-forward folds and a probability of overfitting.In the pack
- Paper desks for forward testsHow I run several paper desks side by side, each with its own ledger, a baseline next to every result and honest PAPER labels, so a forward test can inform a real-money decision without pretending to be one.In the pack
- A graveyard for dead strategiesHow I record every killed trading idea with why it died, check new ideas against it, and route any fix through an autopsy, the full gate and my own approval, with generated notes that nothing trading ever reads.In the pack
Shipping a site 6
- A fast personal site Claude can maintainStatic HTML, a small Python build, year-long asset caching with new names for changed files, browser checks before every push, and a CLAUDE.md that holds the facts, so any Claude session can keep the site right.In the pack
- A second language without driftGive every page a hand-translated twin with the same tag structure, let the build handle links, hreflang and the switch, and run a checker that compares the twins token by token so the Spanish site never falls behind the English one.In the pack
- A 3D scene on a 2D canvas, no libraryTurn a few thousand points into a rotating, zoomable 3D scene with plain canvas math, and keep it cheap, polite and private.In the pack
- Sell a guide pack with a code-unlocked librarySell a pay-what-you-want pack from a static site, with one shared code that a small serverless function swaps for a signed token, and no accounts, cookies or payment handling of your own.In the pack
- Cache busting without a bundlerCache assets for a year, revalidate HTML on every visit, and let a tiny build script put a content hash in every URL that can change, so visitors never get a stale file.In the pack
- Screenshots as testsBefore every push, have a headless browser walk every page at desktop and phone size, fail on anything a script can measure, fake the APIs that cost money, and then look at the screenshots yourself.In the pack
Products built with Claude 4
- Let Claude draft a store, and keep the human in chargeChain a scan, a supplier check, a pick and a build into one agent run that ends in a draft, then let a daily review propose changes that only a person can approve.In the pack
- Building a Roblox game with ClaudeHow I let Claude build a whole Roblox game while I make the taste calls, with a design vault, server authority, offline tests and slices.In the pack
- Building on real bank data without getting burnedThe rules I built into Cupel so a personal finance app can hold real bank connections, feed a dashboard and use Claude without leaking anything.In the pack
- Tenant isolation with Postgres RLSHow I keep each customer's data apart in a multi-tenant SaaS with forced row-level security, a per-transaction tenant setting, a database role that can't bypass it, and a cross-tenant test for every table.In the pack
SEO and AI search 3
- Rank for your own nameTie your site, your photo and your real profiles to one person in structured data, prove you own the site, and give search engines pages with enough depth to show for your name.Free
- Be readable by AI searchLet the AI search crawlers in, put plain dated facts in your HTML, add an llms.txt if you like, and know which of the popular tips the official docs actually back.In the pack
- Local SEO with a Business ProfileSet up and keep a Google Business Profile for a local service business by Google's own rules, keep the name, address and phone identical everywhere, and know which popular tips are only one agency's opinion.In the pack
Design and UI 7
- Accessibility you can checkRun an automated scan for what a machine can find, then do the short keyboard, contrast, label, motion and screen reader pass that it can't, against a small WCAG 2.2 AA list you can actually finish.In the pack
- Design tokens for light and darkPut every colour, radius, font and timing in one tokens file, override it once for dark mode, and keep raw values out of your components so both themes stay in step.In the pack
- Motion that respects the readerGive every animation a job, keep it short, turn movement off for people who ask, and make sure the page is complete before any script runs.In the pack
- Forms people finishFewer fields, labels that stay put, the right keyboard on a phone, and errors that say how to fix them, built from GOV.UK, Baymard and Nielsen Norman Group guidance.In the pack
- Type that reads on a phoneUse the system font, a small fluid type scale, a comfortable line length and size-aware letter-spacing so text reads well on a phone and still works at 200% zoom.In the pack
- Colour that passes contrastBuild colours as steps with jobs, give every colour one meaning, keep chart colours apart from interface colours, and check contrast with a script instead of by eye.In the pack
- Design for the phone firstBuild every screen for a 390 px phone first, with a 16 px gutter, no sideways scroll, real tap targets and the answer at the top, then prove it with phone-width screenshots on every change.In the pack
Selling online 6
- Etsy listings without the hypeWrite titles, tags, descriptions and attributes that match real searches and say only true things, with a lint that flags problems before a listing goes anywhere.In the pack
- A print-on-demand design pipelineLet an image generator draw the illustration, let code handle lettering, layout, exports and checks, and record a visual review against file checksums before a shirt is listed.In the pack
- Digital products that pass QABuild printables and templates from code so they're reproducible, run automatic PDF checks, then record a visual review bound to the exact files, where a missing review is never a pass.In the pack
- Price with every fee inWork out the real profit of one sale, with platform, payment and ad fees, shipping and production cost, in integer cents, test it against the bad cases, and treat any missing cost as incomplete.In the pack
- Product pages that answer questionsBuild a product page from published usability research so a shopper can see the item, know the total cost, find the return terms and trust the reviews without leaving the page.In the pack
- Modes and journals for store automationRun store automation in DRAFT by default, require a recorded authorization for every public or paid action, and send every remote write through a journal that reconciles before it retries.In the pack
Advertising 3
- A read-only ad analystHave a model read each day's ad data and write a short brief with a few actions, while your code holds no way to change a campaign.In the pack
- Negative keywords without hurting yourselfFind the search terms that waste ad money, wait for enough clicks before you judge one, and never block a search for the product you're actually selling.In the pack
- Ads last, and cappedFix the free things first, work out what a sale can afford before you pay for a click, then run one small ad test with a hard cap and a stop rule you wrote down in advance.In the pack
AI for local businesses 3
- An AI receptionist for a trades businessPlan a phone agent for an HVAC or plumbing shop that takes messages and flags emergencies safely, hands off to a person, and gets tested on scripted calls before a real customer hears it.In the pack
- Calls and texts with consentWhat US rules say about AI voice calls, automated texts, opt-outs and AI disclosure, which parts are settled and which aren't, and the records to keep. Education only, not legal advice.In the pack
- Missed-call text-back and quote follow-upBuild a missed-call text-back and a quote follow-up flow in n8n or a similar tool, with emergencies routed to a person, every price approved by a human before it's sent, and a log of every message.In the pack
Security and compliance 2
- Secrets that never leave the hostHow I keep API keys only in the host's env, out of browsers, logs, chats, notes and git, with a credential inventory, a secret scan that never prints the match, and a plan for the day one leaks anyway.In the pack
- Inputs are information, not commandsHow I keep an AI agent from taking orders from emails, web pages, PR comments or tool output, by recording approvals in the owner's own words from the live chat and putting the hard limits in code.In the pack