// decisions

Decisions, with their reasons.

Decisions in three lenses: technical (architecture, incidents and development with AI agents), product (what to build and for whom) and business (money and focus). Each record has the context, the options weighed, the choice and what it cost.

2026-11-01
ADR-013

Stable topic IDs, so saved progress survives roadmap edits

context
Topic IDs are positional: section index and topic index. Rewriting a topic in place hands its saved check and note to a different topic, and inserting one shifts every ID after it. Progress saved on Data Engineering before it went from 109 topics to 78 points at the wrong topics, and two other restructures needed a hand-written SQL migration each.
options
Keep positional IDs and write a migration for every change · Snapshot each roadmap version per user · An explicit id per topic, with a one-time map from the old positions
decision
An explicit id per topic, used by the render, the anchors and the logbook tags. A per-roadmap legacy map migrates the saved rows once, read marks included. Removed topics are retired, not deleted: their rows and notes stay and don't count toward totals. Old #nm-x-y anchors keep working, because blog posts link to them.
consequences
Snapshots were rejected because they would move roadmap content from the HTML into the database and give up the prerendered pages. Data Engineering is the pilot. Not built yet.
2026-10-01
ADR-012

One analytics project for every subdomain, through a first-party proxy

context
The main site, the blog and infoleg-graph live on different subdomains, and a visit has to be followed from a blog post to a roadmap. Ad blockers drop requests to third-party analytics hosts.
options
One PostHog project per site · One project shared by all of them
decision
One shared project, reached through a reverse proxy on a geeklogbook.com subdomain. Opening any page once with ?glb_internal=1 takes that browser out of the numbers.
consequences
Every query has to filter by host. More than half of the raw visitors turned out to be noise (referrer spam, scrapers, one crawler, internal visits), so the exclusions are written down and applied before reading any metric. With samples this small, numbers show a direction, not a proof.
2026-10-01
ADR-010

Merge thin blog posts into canonical guides instead of deleting them

context
Months after moving the blog from WordPress, Google indexed about 25 of its 447 posts. An audit found 127 posts under 300 words, 404 posts that no other post linked to, and dozens of near-duplicates, like several posts on the same pandas or S3 task.
options
Wait · Delete or noindex the thin posts · Merge each cluster into one canonical guide and redirect the rest
decision
Merge in batches. The canonical post keeps its slug and its pubDate and gets an updatedDate. Absorbed posts are deleted and 301'd, with and without the trailing slash, and the old WordPress URLs point straight at the canonical. Tag pages and pagination are noindex,follow.
consequences
No unique content is lost, and each merge also adds links between related posts. The first batch merged 10 clusters on 2026-10-15; a log of done merges keeps later batches from repeating them. The redirect list keeps growing, so a script checks it.
2026-10-01
ADR-008

Publishing is a commit: no CMS for the logbook

context
Every TIL or logbook entry is a commit and a deploy. The concern was that it would slow down writing.
options
Entries in Supabase, an /admin editor, and a Pages Function that injects them into the HTML at the edge · A scheduled GitHub Action · Entries stay in the static HTML and a commit publishes them
decision
Keep committing. Commits happen every day anyway, and a push to main is already the deploy.
consequences
No editor, no database and no extra moving parts, and every logbook page stays readable with no backend at all. Drafts live in drafts/, which returns 404, and move into the page when they're complete. Publishing from a phone isn't possible.
2026-10-01
ADR-007

Prerender roadmaps with small scripts, not a framework

context
Roadmap topics were rendered in the browser, so crawlers saw empty pages. Titles, canonicals and the sitemap were inconsistent, and the blog already ran on Astro.
options
Move the site to Astro like the blog · Keep rendering in the browser and hope crawlers run JavaScript · Dependency-free Node scripts that write static HTML between markers, plus a checker
decision
Scripts. One prerenders each roadmap's topics into its HTML, one writes every page's head tags and JSON-LD from a single table, one builds the sitemap with dates from git, and verify-seo fails if a rule breaks: one h1, unique titles of 60 characters or fewer, extensionless URLs, no visible todo.
consequences
No framework and no npm dependencies, and the browser re-renders the same markup. Generated blocks are never edited by hand, and every content change ends with running the scripts, which an agent does reliably (ADR-002).
2026-10-01
ADR-006

Roadmap progress: magic-link login on Supabase, mirrored in the browser

context
Roadmaps needed checks and notes that persist and sync across devices, with full create, read, update and delete, on a static site with no server of its own.
options
A JSON file in the repo · Browser storage only · A managed Postgres with auth (Supabase) · A custom API on Pages Functions and D1
decision
Supabase magic-link login and one topic_state table (user, roadmap, topic, done, note) protected by row-level security. Every change is mirrored to localStorage, and local progress merges into the account at the first login.
consequences
The publishable key sits in a public script by design, because row-level security is what protects the data. Progress works without an account, so if the backend is down the roadmaps still work and only the sync is lost. Topic IDs are positional, so moving topics needs a SQL migration of saved rows (three in October).
update
2026-11-06: the Supabase project was suspended, and the browser mirror is what kept the roadmaps working (incident →).
2026-10-01
ADR-003

Serve infoleg-graph ↗ from Cloudflare D1 on the free plan

context
A navigable graph of Argentine national norms from the InfoLeg open dataset: which norm modifies which. About 425,000 norms and 376,000 links that change a few times a year, read-only for visitors, and no budget for a database server.
options
Precompute everything into static JSON · A managed Postgres (Supabase) · Cloudflare D1 behind Pages Functions, next to the site that already lives on Pages (ADR-001)
decision
D1 with three Pages Functions (search, norm, neighborhood). A Python + DuckDB pipeline turns the CSVs into SQL files. The overview "map of laws" is precomputed into a static JSON file and never touches D1.
consequences
The free plan sets the rules. D1 counts every index entry as a row written, so a 50k-row file cost ~150k writes against a 100k/day quota: files are now sized by write cost, inserts are INSERT OR IGNORE, and the loader resumes the next day. Files load hubs first, so a half-loaded database already shows the most connected norms. Reads stay low with a 24 h edge cache per URL and an FTS5 title index instead of table scans. Updating the data rewrites production, so it is a manual, checked procedure.
2026-10-01
ADR-002

Maintain the site with LLM agents, guided by AGENTS.md

context
Hand-written site with no framework, where most of the rules are conventions: design tokens, EN/ES copy, SEO limits, generated blocks, cache busting. Most changes, from new roadmaps and their resources to logbook entries and copy, are now made with an LLM coding agent, and the agent forgets everything between sessions.
options
Re-explain the site in every prompt · A framework or CMS that enforces the structure · An AGENTS.md in the repo that the agent reads every session, plus idempotent scripts that rebuild and check the result
decision
AGENTS.md is the single source of truth, and CLAUDE.md just points to it. It holds the product definition, the map, what to run after each kind of change, the visual style, the copy conventions and the don'ts. A change that sets a new rule updates AGENTS.md in the same commit. The scripts regenerate the derived parts, and verify-seo must pass before anything ships.
consequences
Changes come out consistent without re-explaining the site, and the docs keep up with the code because they change together. An outdated AGENTS.md misleads every later session, so keeping it current is part of the work. Generated blocks are never edited by hand. What an agent finds, like links, citations and figures, still gets a human check before it is published. AGENTS.md is not public: it is routed to a 404 through _routes.json.
2026-05-01
ADR-001

Host geeklogbook.com on Cloudflare Pages

context
Static HTML site, ~1,000 legacy WordPress URLs to redirect to the new blog, and build scripts living in the same repo.
options
Cloudflare Pages · GitHub Pages · Netlify
decision
Cloudflare Pages: _redirects for the legacy URLs and Pages Functions where redirects fall short.
consequences
Static rules must come before dynamic ones (100-rule dynamic cap). _redirects cannot return 404, so internal paths go through a Function routed by _routes.json.
2026-11-06
geeklogbook.comincident

The progress backend went away, and the roadmaps kept working

The Supabase project behind login and progress sync was suspended, and its host stopped resolving. Checks, notes and read marks kept working for every visitor, because every change was already mirrored to the browser (ADR-006 →).

What broke: login and sync across devices. The sign-in button still shows, and asking for a magic link fails with "Failed to fetch". The options on the table: restore the project, switch on the local-only mode progress.js already has, or move progress to D1 as infoleg-graph does (ADR-003 →).

2026-11-04
geeklogbook.com

Owner opt-out for analytics, and redirects for old URLs still getting visits

Opening any page once with ?glb_internal=1 stops that browser from being counted. Old WordPress URLs that showed up as 404s in analytics, slugs with _ or → that the first redirect pass had stripped, got their 301s.

2026-10-29
geeklogbook.comincident

Data Engineering rebuilt, and read marks and estimated time on every roadmap

Data Engineering went from 109 topics to 78, keeping only the big-tech core (ADR-009 →). Every course, reading and practice link got its own check, and checking the last one completes the topic. Each roadmap shows an estimated time computed from its links, and the ones built on books or long courses got a bibliography.

What broke: topic IDs are positional, so progress saved on the old Data Engineering topics now points at different ones. The fix, stable topic IDs, is planned but not built.

2026-10-24
clientincident

A WordPress site where a push is the deploy

Client work, anonymized. The host deploys every push to main. The first push used the old repo layout, with only wp-content in it, and took the site down. Now the repo is the whole web root: each push lands in a new release folder and a symlink flips to it. Uploads and config live in a shared folder outside the repo, and rolling back means pointing the symlink at the previous folder.

What broke next: PHP's OPcache kept serving the previous release after the symlink flipped. A must-use plugin now clears it when the served path changes.

2026-10-22
infoleg-graph

infoleg-graph.geeklogbook.com ↗ goes live

A graph of Argentine national norms: which one modifies which (ADR-003 →). Pages builds the site from GitHub. The data goes into D1 separately, in SQL files sized to the free plan's daily write quota and ordered from the most connected norms down, so a half-loaded database already shows the useful part.

The same day, a political scientist tried it alongside an AI assistant and sent notes. The main view changed from the graph to a plain page per norm, with a summary, charts and the full list of links. Code and data now deploy independently: the API reads link citations only when the database has them.

2026-10-21
blogincident

Topics page, and analytics through a first-party proxy

The blog got a page that groups posts by area of interest. Both sites now send analytics through a reverse proxy on a geeklogbook.com subdomain (ADR-012 →).

What broke: nothing, but it looked like it. Three minutes after the push, production had no topics page. The commit was on main; the build was just slow, because it renders an Open Graph image for each of 400+ posts. A background check polled the URL until it went live.

2026-10-16
geeklogbook.comincident

The logbook replaces the classes page

TIL, now, decisions, lab, datasets, radar, cheatsheets and glossary, a shared nav with a logbook dropdown, and the EN/ES toggle on every page. AGENTS.md was written the same day (ADR-002 →).

What broke: an outside check said production was stale. A curl with a cache-busting query showed it was live: the checker was reading draft entries hidden in HTML comments. That was still a leak, so drafts moved to drafts/, which returns 404, and verify-seo now fails on any visible todo.

2026-10-15
geeklogbook.comincident

An indexable site, and /materials becomes /roadmaps

Extensionless URLs everywhere, roadmap topics prerendered into the HTML, generated head tags and JSON-LD, a sitemap with lastmod from git, and verify-seo (ADR-007 →). Internal repo files return a real 404 through a Pages Function. The paid pack was removed and /materials now redirects to /roadmaps (ADR-004 →).

What broke: part of the legacy redirects had been silently ignored. Pages stops reading _redirects after 100 dynamic rules, and every rule after the first splat counts as dynamic. Static rules moved to the top (TIL →).

2026-10-15
blog

Content audit, and the first batch of merged posts

An audit of 447 posts (words, code blocks, internal links in and out) proposed 41 clusters of overlapping posts. The first batch merged 10 of them into canonical guides, with 301s for the absorbed posts (ADR-010 →). Tag pages and pagination went noindex,follow, posts got an updatedDate, the 18 posts still loading images from the old WordPress got them fixed or removed, and the blog header now links to the roadmaps.

2026-10-14
geeklogbook.com

Login and progress sync for the roadmaps

Magic-link login on Supabase, with progress mirrored in the browser so it also works without an account (ADR-006 →).

2026-10-12
geeklogbook.com

Redirects for the old WordPress URLs

Months after the blog moved to its own subdomain, Search Console still reported errors on the old URLs under geeklogbook.com. A 945-line _redirects sends each old post to its new address on the blog.

2026-05-22
geeklogbook.com

geeklogbook.com restarts as one static page

A single hand-written index.html on Cloudflare Pages, with no framework (ADR-001 →). Analytics followed the next day. The first tracks, Data Engineering and English, followed on 2026-05-26.

2026-05-01
blog

The blog moves from WordPress to Astro

blog.geeklogbook.com ↗: Astro on the AstroPaper theme, posts in Markdown, Cloudflare Pages, and Pagefind search built at deploy time. That freed geeklogbook.com for the roadmaps.

2026-11-07
memoryverification

It reads the project list at the start of every session

problem
Ideas for this site were scattered in a spreadsheet that tracks every project, and they only came up when someone opened it.
change
A memory tells the agent to read the sheet when a session starts and say which projects fit Geek Logbook: a roadmap topic, a logbook section or a post. It names the source of every claim.
result
The first review oversold twice: an empty draft presented as a finished entry, and "already built" copied from a stale row. Now it checks the repo, on disk and on GitHub, before calling anything built.
2026-11-05
memory

Ask for the design reference before writing any UI

problem
On a client site, the agent picked a palette and a style on its own, and had to be stopped halfway to receive the reference that already existed.
change
Before it styles anything on a new project, it asks whether there is a design reference, or looks for one in the repo.
result
No more first versions built in a direction nobody asked for.
2026-11-04
mcpverification

Analytics over MCP, with the noise written down first

problem
The agent was asked to improve the site from its metrics. The raw numbers showed hundreds of visitors and a wave of rage-click "bugs".
change
PostHog is connected over MCP, and before reading any metric the agent applies a written list of exclusions: referrer spam, scraper fingerprints, one crawler's day of clicks, and internal visits. ?glb_internal=1 keeps internal browsers out from now on.
result
More than half of the raw visitors were noise, and the rage clicks were the crawler. Small clean samples are read as direction, not proof (ADR-012 →).
2026-10-29
memory

Corrections become memories, with the reason

problem
Granular roadmap links had been requested before, and the agent linked a whole course again. The same correction was repeated across sessions.
change
After each correction, the agent saves one memory file with the rule, why it exists and how to apply it. Some of them: one lesson, chapter or page per link, 20 minutes or less, found by opening the destination. Never link an unrelated project. No emojis.
result
Corrections stick. A memory that turns out wrong gets deleted, and one that names a file or a flag gets checked against the code before the agent relies on it.
2026-10-29
agents.md

A parking lot for ideas the agent must not start

problem
Big ideas came up in the middle of small tasks. One example: stable topic IDs, so saved progress survives roadmap edits. That was too much to do right then, and too good to lose.
change
AGENTS.md has an "Ideas & improvements" section with the problem, the plan and the rejected options, under the rule "only pick one up when asked". Pending work goes in a TODO.md that lives in the repo and returns 404 in production.
result
The plan is written down when requested, and the agent doesn't start it on its own.
2026-10-23
review

Production stays human-run: the agent writes the script, a person runs it

problem
On a server reached over SSH, the agent's permission checks (rightly) blocked it from writing to production, and the errors from manual runs were hard to paste back whole.
change
The agent writes the script, a person runs it, and every script appends its output to a logs.txt that the agent reads.
result
Only a person touches production, and the agent still sees exactly what happened.
2026-10-22
prompts

Call the fluff by its name

problem
Asked to refocus a roadmap, the agent renamed topics ("Crop cycle" became "The season as a unit of data") and kept the same content. Difficulty badges like core and expert were in the same family: adjectives with nothing behind them.
change
The request says it plainly ("this is renaming, go straight to the point") and asks what each piece gives a learner. The copy rules in AGENTS.md ask for short, plain, professional text with no first person and no marketing tone.
result
The topics were rewritten around the data each activity actually produces, and the badges are gone from every roadmap.
2026-10-22
agents.md

One AGENTS.md per repo, starting with what to check first

problem
Each repo has chores that should never be skipped, and they kept being forgotten.
change
Every repo has an AGENTS.md, and CLAUDE.md only points to it, so any agent finds the same file. Each one opens with the routine: in infoleg-graph, ask the open-data API whether a newer dataset is out before touching anything; on the blog, run a taxonomy report after every new post.
result
The routine runs without being asked. Anything that writes to production still needs an explicit OK, and that rule lives in the same file.
2026-10-21
review

The agent doesn't touch git

problem
The agent committed to new branches, and work got merged before anyone had read it.
change
It edits on main and stops. A person reads the diff and commits, and the push deploys.
result
The human commit is the review. The rule is in this site's AGENTS.md and in a memory for the other repos.
2026-10-21
verification

Watch the deploy instead of guessing

problem
"Production shows nothing new" right after a push, and the guesses started: a preview branch, a failed build, a cache.
change
The agent checks that the commit is on the remote, compares it with what production serves using curl and a cache-busting query, and if the build is just slow, polls a URL in the background until it changes.
result
Two "failed deploys" were a slow build (an image per post) and a checker reading HTML comments (deploys →).
2026-10-20
prompts

Real users review what the agent wrote

problem
The AI for lawyers roadmap looked complete, but a model wrote it, and from the inside the useful parts were hard to tell from the filler.
change
A practicing lawyer read it, and the notes went to the agent as they were, asking it to keep only what gives value in practice. On infoleg-graph, a user's notes became a NEXT_STEPS.md that the agent keeps up to date.
result
The roadmap went from 30 topics to 23, and the US edition got the same cut. infoleg-graph's main view changed from the graph to a plain page per norm.
2026-10-16
agents.md

The site explains itself to the agent

problem
The agent forgets everything between sessions, and this site runs on conventions: tokens, EN/ES copy, SEO limits, generated blocks.
change
An AGENTS.md with the product, the map, what to run after each kind of change, the style and the don'ts. A change that sets a new rule updates it in the same commit.
result
Changes come out consistent without re-explaining the site (ADR-002 →).
2026-10-15
promptsverification

Plan in a chat, build in the repo, audit in production

problem
Big jobs, like an SEO overhaul, a content audit of 447 posts or a new site, don't fit in a quick instruction.
change
The brief is written in a chat with the model: context, numbered tasks and hard limits ("never modify pubDate", "do not change the visual design", "local only, no deploy"). The coding agent runs it in the repo step by step. Afterwards the chat audits the live site.
result
The brief gets reviewed before anything changes. The audit caught a real leak and also raised false alarms, so the coding agent checks production itself before acting on a report.
2026-10-15
mcp

The agent gets the database through MCP

problem
For roadmap progress, the agent wrote SQL that was copied by hand into the database dashboard, and the result was reported back.
change
The Supabase MCP server, configured per project, lets the agent inspect the schema and the row-level security policies and run the SQL itself.
result
Fewer round trips through the dashboard. The SQL files still live in the repo as the record (ADR-006 →).
2026-08-24
memory

Fix only what was asked

problem
In a data pipeline at work, right after fixing the reported bug, the agent started proposing a new layer and changes to code that worked. Don't fix what isn't broken.
change
It fixes the reported bug and stops. A related problem it notices is mentioned as a possible follow-up, never started.
result
Diffs stay small enough to review, and nothing changes that wasn't asked for.
2026-05-30
skills

Where it started: guides to Claude Code, then skills

problem
The goal was a real workflow with the agent, not a chat that edits files.
change
Four guides (hooks, slash commands, MCP servers, custom agents) were written and briefly sold them as a pack (deploys →). Later, the brand rules and the content planning became skills: the format, voice and wordmark of every post load on their own when writing for Geek Logbook.
result
The pack is gone, and what writing it taught is this page.
2026-10-01
ADR-011

Blog topics: curated, with boundary rules, not raw clusters

context
The blog had 135 tags, with duplicates (DataEngineering, Healthcare Data, HealthcareAnalytics) and many used once. Clustering the tags by co-occurrence (Louvain) gave groups, but some of them reflected tagging gaps, like Python as a "topic", rather than real subjects.
options
Use the clusters as they came · Tags only · A short list of curated topics, each with an explicit "goes in / stays out" rule
decision
Curated topics with boundary rules. Single-use tags fold into a broader one, and renamed tags get redirects. Each field (tech, healthcare, finance, IoT) gets a topic map of neighboring subjects, where hollow nodes are posts not written yet.
consequences
Every new post runs a taxonomy report, written into the blog's AGENTS.md, that flags new tags and near-duplicates and shows what the post newly covers on the maps. The maps double as a list of writing ideas.
2026-10-01
ADR-009

Roadmap links: one lesson each, 20 minutes or less

context
Topics linked whole courses or docs home pages, which says nothing about where to start. Topics also carried core, advanced and expert badges. Data Engineering, meant for top-1% work at big tech companies, mixed generic Python and live coding with distributed systems.
options
Keep course-level links and add more badges · Link one lesson, chapter or page per topic and move whole works elsewhere
decision
Each link is one lesson, chapter or doc page that takes 20 minutes or less, found by opening the destination. Whole books and courses go in a per-roadmap bibliography, cited by chapter. The badges were removed: they told a learner nothing they could act on. Data Engineering kept only the big-tech core as topics, with its baseline moved to external courses under "other resources", plus practice and building sections at the end.
consequences
Every link gets its own read check, and a roadmap's estimated time comes from its links. Data Engineering went from 109 topics to 78. Topic IDs are positional, so progress saved before the change points at the wrong topics; stable IDs are planned, not built.
2026-10-01
ADR-005

Define the product before reading the metrics

context
The home had grown one section at a time (roadmaps, logbook, blog, social links) with no order, and analytics was about to decide what to change. With no definition, any number could justify any section.
options
Let the metrics decide · Make the logbook the product, since that is where the time goes · Define the product first
decision
Geek Logbook is learning roadmaps, built by someone who learns in public. The roadmaps are the product, the logbook is the proof and the blog is the explanation. The first reader is someone who wants to learn a field and doesn't know where to start.
consequences
The home follows that order, and Roadmaps and Blog are top-level nav, never logbook sections. A home block about current work was cut (it only restated what its author already knew), and in its place one real topic is shown as a learner sees it. The definition opens AGENTS.md, so a change that doesn't fit gets raised before it's built.
2026-11-15
pa38.geeklogbook.com

PA-38 Cockpit Lab on its own subdomain

An interactive cockpit of the Piper PA-38 Tomahawk, built from its flight manual: instruments, controls, a simple flight model, the manual's diagrams and a cockpit quiz. It lives at pa38.geeklogbook.com ↗, with its own repo and deploy, and links both ways with the blog posts on each instrument. How it was built is in /lab →.

2026-11-06
geeklogbook.com

Home: marker highlights, margin notes and a closing band

The last phrase of each heading gets an ember marker stroke, short margin notes say what each part of a sample topic gives a newcomer, and the home ends on one saturated band that sends to the roadmaps.

2026-10-30
geeklogbook.com

Data Modeling and AI Agents on Business Data

The question was whether three parts of the job deserved their own roadmap: pipelines that move millions of transactions, the data model a business runs on, and AI agents that answer with that data. Pipelines were already Data Engineering. The other two became Data Modeling → (42 topics) and AI Agents on Business Data → (32 topics).

2026-10-23
geeklogbook.com

Math for Finance and Oil & Gas Data, and the datasets page reworked

Math for Finance → collects the math under recurring finance questions. Oil & Gas Data → goes from well logs to SCADA, with open data from Vaca Muerta. /datasets → now leads with the sources actually used and what was built on them, and /papers → got the papers already read from the reading folder.

2026-10-22
blog

Topic maps and a tags dashboard

Each field (tech, healthcare, finance, IoT) has a map of neighboring subjects. Hollow nodes are posts not written yet. The 135 tags, full of duplicates and one-post tags, were merged, and every renamed tag got a redirect (ADR-011 →).

2026-10-22
geeklogbook.com

Drones for agriculture, in English and in Spanish

Two separate roadmaps, each with its own country's rules and resources in its own language: Drones for Agriculture → (US) and Drones para el agro → (Argentina). Both end with the economics: whether it is viable, and drone against airplane against ground rig. Agro + IoT and Minería + IoT were rewritten around the data each activity actually produces, and saved progress moved to the new topics with a SQL migration.

2026-10-20
geeklogbook.com

AI for lawyers, cut to what a lawyer uses

A practicing lawyer went through the Spanish edition and pointed at the filler. The theory shrank to three topics, a section on writing prompts was added because that was what was missing, and case-law research moved to the end. The Spanish edition went from 30 topics to 23, and the US edition got the same cut, from 33 to 24.

2026-10-19
geeklogbook.com

AI for professions, papers, and a home built around the product

Roadmaps for journalists → and lawyers →, plus a Spanish edition for Argentine lawyers. A /papers → section. The home was reordered once the product was defined in writing (ADR-005 →).

2026-09-30
geeklogbook.com

Finance and US Healthcare tracks

The first two outside data engineering, English and math. Back then they were "tracks" under /materials.

2026-10-01
ADR-004

Free roadmaps: drop the paid materials and the classes page

context
From May, the site sold a paid pack of Claude Code guides next to free learning tracks, and from June it had a checkout and a classes page. Geek Logbook is a learning log, and selling on it pulled in another direction.
options
Keep both · Keep the classes as a service and drop the pack · Free roadmaps plus a public logbook, with classes reduced to a contact link
decision
Free roadmaps and a logbook. /materials became /roadmaps, the paid page was removed, and the logbook sections took the place of the classes page. Private classes are a single link in the footer.
consequences
The old /materials and /classes URLs needed redirects. With nothing to sell, every page can be written for the learner, which led to defining the product (ADR-005).
2026-05-30
geeklogbook.com

Paid materials: a Claude Code skills pack

Four guides (hooks, slash commands, MCP servers, custom agents) sold as a $9 pack next to the free tracks. A checkout and a classes page with a contact form followed on 2026-06-13. All three were removed on 2026-10-15 (ADR-004 →). Classes live on as a link in the footer.

geeklogbook.com

blog.geeklogbook.com ↗

infoleg-graph.geeklogbook.com ↗