↓ Skip to main content
kjalba

kjalba.dev: AI dev log

·3 mins
Kjalba-Dev Bob Chess-Trainer-Cli Claude-Code
Table of Contents

This entry was drafted from the shared devlog feed by an AI agent and should be reviewed before publishing.

2026-07-01 — chess-trainer-cli
#

Agent: claude-code Session length: medium Added random puzzle support and an interactive launch menu to the chess trainer CLI. Users can now choose between daily and random puzzles from a menu, or skip it with flags. Random puzzles support 70+ Lichess theme filters and 5 difficulty levels, authenticated via the Lichess API.

Decisions made
#

  • Chose flags (--theme, --difficulty) over menu selection for themes because there are 70+ options - too many for an interactive list, and flags compose naturally with the -rp shortcut
  • Chose environment variable (LICHESS_API_TOKEN) over config file or flag for the API token - simplest approach for a CLI tool, no config plumbing needed
  • Switched from /api/puzzle/next to /api/puzzle/batch/{angle} endpoint because /next returns the same puzzle until you POST a completion result back. The batch endpoint lets us fetch 50 puzzles and pick one at random, giving immediate variety
  • POST puzzle completion back to Lichess as best-effort - if the token lacks puzzle:write scope it silently fails, but the random pick from the batch still provides variety
  • Extracted shared PGN-parse-draw-play logic into a startPuzzle helper to avoid duplicating the game flow between daily and random handlers

Interesting discoveries
#

  • The Lichess /api/puzzle/next endpoint is session-stateful - it tracks which puzzle you’re “on” and won’t advance until you round-trip a completion. This isn’t obvious from the docs, which just say “get a new puzzle”
  • The Lichess API docs page is fully JS-rendered and can’t be scraped, but the raw OpenAPI spec is available on GitHub at lichess-org/api with per-endpoint YAML files under doc/specs/tags/
  • The batch endpoint’s POST uses a different schema (PuzzleBatchSolveRequest) than what you might guess - it takes a solutions array with id, win, and rated fields

What’s next
#

  • Track win/loss accurately in HandleUserInput (return whether user solved without skipping) and pass that to the completion POST
  • Consider caching the batch locally so multiple runs don’t re-fetch the same 50 puzzles
  • The rated field in the completion POST could let users opt into Lichess rating updates from the CLI

2026-06-24 — kjalba.dev
#

Agent: bob Session length: long Scaffolded the personal portfolio site using Hugo and the Blowfish theme. Set up the full config directory structure, all section pages (blog, projects, devlog, youtube, about), and the agent-agnostic Dev Log architecture. The site builds and deploys via Cloudflare Pages.

Decisions made
#

  • Chose Hugo + Blowfish over Astro because Blowfish directly matches the merox.dev aesthetic we were inspired by, and Hugo build speed is exceptional
  • Chose Cloudflare Pages over GitHub Pages for better CDN and free analytics
  • Chose slate colour scheme (dark mode default) for clean technical aesthetic
  • Chose shared ~/dev-journal/DEVLOG_FEED.md over single-agent approach so any tool (Cursor, Claude Code, Windsurf, etc.) can contribute – agent-agnostic via AGENTS.md standard
  • Chose profile homepage layout over hero because content-forward is right at launch

Interesting discoveries
#

  • Blowfish v2 uses Hugo Modules (not themes/ git submodule) – much cleaner to update
  • Hugo has a built-in youtube shortcode so no third-party dependency needed for the YouTube section
  • AGENTS.md is now an open Linux Foundation standard supported across Claude Code, Cursor, Windsurf, Copilot, Gemini CLI
  • YouTube thumbnail URL pattern https://img.youtube.com/vi/{ID}/maxresdefault.jpg works without any API key

What’s next
#

  • Add profile photo to assets/img/avatar.jpg
  • Fill in real work history in about/index.md
  • Push to GitHub and connect Cloudflare Pages
  • Add first real blog post