developing pro-human tools and systems • background in international non-profit work • col 3:23-24 † • current idea: reclaiming screen time ( ios native)
@mctavish9579@ClaudeDevs I literally had a note of drafted prompts and was trying to sort them into tiers of importance since I probably had 1-2 left when the reset hit
Shipped something small and useful today, .md Manager, an open source, one command tool to keep tabs on the state of your claude.md/agents.md files etc.
The design constraints were the fun part:
• Fully local, with zero network requests and no telemetry
• One file and one dependency in ~600 lines
• Read-only by default (editing is an explicit opt-in)
• My favorite: agent friendly for custom install and config
As with my other tools, I built this for my own use after struggling to keep up with outdated .md files. Posting it in case anyone else is able to benefit.
https://t.co/3BAZ4UNHqd
@sickdotdev Is the desktop app not super slow for you all? I find that CLI is way faster. I’ll generally use CLI (Ghostty) for most work, and desktop for anything that requires multisession worktree flows with live previews
Memex. v2.1 is live today. Memex serves as a connected knowledge system for Claude Cowork with persistent memory, tiered context loading, full [[wikilink]] integration, and cross-workspace federation. All in pure markdown with zero runtime dependencies and 30 second install time as a plugin.
What a Cowork day looks like with it:
[Morning] Open a new session. Memex auto-briefs from yesterday's closing state: status, in-flight items, blockers, last decisions. 30 seconds.
[Mid-session] Claude consults per-hub closets (typed-field summaries) to decide which files a question actually needs. No "load everything and hope." Cross-hub queries hit /memex:search; cross-workspace hits /memex:cross-search.
[End of session] Auto-captures status, appends to session-log, refreshes closets, verifies wikilinks.
Under the hood:
• Pure markdown + Python stdlib. No DB, no embeddings, no API keys.
• Obsidian compatibility. Open as a vault, see the graph.
• 90.1% Recall@5 on LongMemEval-S, within 0.5pp of full-content keyword search at ~1/10th the size.
https://t.co/jTrkXY4FKt
@garrytan Just in time context is so important. The more you chain skills vs speccing out 1 large skill, the more room you leave for drift. I imagine that this gap will close over time, I just hit it's limitations as of today.
One of the main problems with the idea in this post is that when the chain gets long enough, the fallout presents itself not in the correct sequence of skills (if built well), but all the additional context that said skills may require along the way. Having fewer, bigger skills, has given me better results in terms of pulling in the right context at the right time.
@Flomerboy is there any program available to get more Claude Design usage? I would love to test the boundaries of this more, but am finding that I hit weekly usage extremely fast.
Hey @felixrieseberg! I'm finding it difficult to use both cowork and claude code in desktop, as my MCP connections I use on cowork don't appear to be lazy loading for claude code. CLI Is lazy loading and running much leaner with the same MCPs/plugins. Is this user error on my end?