public digest · 5 picks
This week: IPFS blogging, resurrected road trip maps, and three agent skills
Five picks this week that split cleanly into two piles: software that does one infrastructure job well, and the fast-growing genre of 'agent skill' repos that are really just markdown wrapped around a slash command. Both piles are worth your time, but they reward different kinds of scrutiny.
// pick 1 of 5
Planet is a macOS SwiftUI app for writing a static site or blog, building it, and publishing it to IPFS under an IPNS key that can be set as an ENS contenthash. It also reads RSS, Atom, and JSON Feed, and can follow ENS-hosted sites. It's aimed at people already working in the ENS and IPFS ecosystem who want a native client instead of a web tool.
A native macOS app that bundles a Kubo IPFS node so you can write, build, and publish a blog to IPNS without running a separate daemon, and optionally point an ENS name at it via contenthash. The two-way following — pinning sites you follow so they survive offline, plus plain RSS support so a site doesn't need ENS to be followed — is the detail that makes this feel like a real client instead of a tech demo.
What to know going in: this is early-stage software by its own README, and losing your IPNS key means losing your site's identity, so back it up before you rely on this for anything real. The Kubo binaries ship via Git LFS, and a plain clone leaves you with useless pointer files instead of binaries — an easy first-run trap. macOS only, and gas costs for updating contenthash aren't discussed anywhere, which matters if you publish often.
View on GitHub → Our full take →
// pick 2 of 5
A restoration of Randal Olson's 2015-2018 'optimal road trip' interactive maps, which broke after Google required billing for Maps API calls. This rebuilds every map as a fully static page using OSRM-routed OpenStreetMap data and Protomaps vector tiles instead of live Google Maps calls, so it's for anyone who wants the nostalgia maps working again, not really for anyone building new software.
This exists because Google put Maps behind a paywall and broke a decade of blog embeds. The fix is unglamorous and correct: pin an OSM snapshot per trip, vendor the Protomaps basemaps and fonts, and keep every original URL working so old bookmarks still resolve. Licensing is handled per-source instead of one blanket license slapped on top, which is more care than most archival projects bother with.
What to know going in: the actual pipeline that generated this data is private, so you're trusting the README, not auditing a build process. The repo is heavy — a 51MB basemap plus hundreds of font glyph files committed straight into git, no LFS — and there's no code to speak of, so 'Language: HTML' undersells what this really is, a data archive with a software repo's trappings.
View on GitHub → Our full take →
// pick 3 of 5
This is a plugin marketplace manifest for Claude Code — a thin JSON catalog that points to four of obra's own repos (a skills library, an Elements of Style writing guide, a plugin-dev reference, and a journaling MCP server) so they can be installed with `/plugin marketplace add` and `/plugin install`. It's for Claude Code users who want one-command installs instead of manually wiring up plugins.
A thin JSON manifest that lets you install four of obra's own Claude Code plugins — a skills library, a writing-style guide, a plugin-dev reference, a journaling MCP server — with two slash commands instead of manual wiring. The separation between manifest and plugin code is the right call: each plugin can ship independently without touching this repo.
What to know going in: there's no code here to evaluate, and the README doesn't say whether the manifest pins specific versions of the linked plugins or tracks their default branch — that distinction matters if a linked repo ships a breaking change. 'Marketplace' is generous framing for four plugins from one author with no visible process for outside submissions; treat it as a convenient installer, not an ecosystem yet.
View on GitHub → Our full take →
// pick 4 of 5
A pair of 'agent skills' (installable into Claude Code, Cursor, Codex, etc.) that tell a coding agent how to build Expo/React Native mobile screens matching the UI patterns of top-grossing apps. One skill drives research against Appllama's paid MCP design library, the other is a build-quality bar (native controls, navigation rules, animation discipline, simulator verification). It's aimed at teams using AI agents to build mobile app UI who want a shortcut past generic-looking screens.
Two agent skills for building Expo/React Native screens that actually encode mobile UI opinions worth having — push vs. replace navigation, keeping animations off the JS thread, springs tuned to finger velocity. That's more specific than the usual 'make it look polished' skill prompt, and it installs the same way across 70+ agent clients via the skills.sh convention.
What to know going in: the actual selling point — a library of top-grossing app screens to crib from — sits behind a paid MCP with a credit system, so the open repo is a prompt wrapper around a closed backend you can't inspect. There's no code, no tests, no versioning in this repo, just markdown, so there's no way to tell if an edit changes agent behavior before it ships. The design skill does work standalone without the paid tier, which is the one honest free path in here.
View on GitHub → Our full take →
// pick 5 of 5
Caveman is a set of agent skills plus a local CLI proxy. The skill makes a coding agent answer in terse fragments while leaving code, commands, and error text untouched, and the proxy compresses logs, JSON, CSV, YAML, and test output before the agent reads them. It is for developers who already pay real token bills in Claude Code or a similar agent and can live with a clipped voice.
A terse-response skill paired with a local proxy that compresses logs, JSON, CSV, YAML, and test output before an agent reads them, with originals kept on disk so the agent can fetch full detail back when it needs to. The HONEST-NUMBERS doc is the standout: it shows where the tool actually loses, including a 9.9% regression on HTML sessions, instead of only publishing the wins. The line it draws for the skill — strip articles, never touch negations, numbers, paths, or payloads — is the right place to cut for a tool whose output includes code.
What to know going in: the marketing says 65% token savings, the README's own numbers say 33.2% input and 8.5% output, so quote the lower figures. The rule set adds roughly 1,000 tokens to every call, meaning short exchanges can get worse, not better. The CLI phones home usage telemetry — commands, token counts, install ID, OS, IP — by default, opt-out, which you'll want to know before putting it on a work machine. The project also spans five languages and 30+ agent profiles, so expect uneven maintenance across the edges.
View on GitHub → Our full take →
If you want these picks in your inbox before we write them up, the email signup is at the bottom of the page — one email a week, no filler.
Get this in your inbox →