Sledge

Wait we can just plug in a Mac Mini to our Teslas?? 👀👀😱😱

Finally built a "true" harness in Cursor and improved successful chats by +7.4pp Before - AGENTS.md repo memory files - Session protocol docs - Skill routing table - Task-specific skills - Knowledge capture rules - Context-doc hygiene rules - Canonical path guidance - Early chat quality scoring - Mostly passive instructions Chats became way more effective at bringing in context, but I didn't have a lot of control over how they started their work. After W15/W16 – added a true harness - Task / Scope / Skill orientation rule - Skill-first hard gate - Hardened skill routing quick-checks - Harness audit/bootstrap scripts - Hook scaffolding - Cursor health canvas - Repeatable smoke/E2E skills - More client/operator skills - Success rose from 73.8% to 81.2% - Skill-first compliance rose from 15% to 43.5% Reviewed a few different articles about harnesses and scraped all of Cursor workshops to understand a bit more about best practices. Realized I wasn't using hooks and linters and was just praying that the agent followed my rules and routed accordingly. I was already analyzing chats, but canvases made it much easier to visualize what was going on. I didn't want to invest in building a dashboard just to understand this, so those really helped. During this time, I also increased my usage of Composer 2 which I think was the real reason why I had to improve the harness. I noticed that the effectiveness of my chats were obviously decreasing compared to Sonnet, so I wanted to see if there was a way I could supplement that degradations with a better harness. After Today - Re-scored 622 chats with GPT-5.5 - Built failure diagnosis report - Built harness scorecard - Built audit Canvas - Hardened first-action gate - Added stronger scope confirmation - Expanded routing - Archived audit outputs for future incremental runs - Added deterministic harness health check I was originally using 4o-mini to analyze chats. That was a terrible idea. I was just being too cheap. It wasn't really understanding the full context of an iteration in the chat and it was marking everything as failures. Which was fine if everything is graded that way, but I needed something a little bit more rigorous to explain why things were failing. Completed work for context W15 - Rebuilt the Levered site and free-audit funnel - Built the voice-agent eval harness - Shipped client sites and migrations - Pushed consumer app closer to launch W16 - Added the first real agent harness layer - Built the Cursor spend pipeline - Shipped client landing pages and variants - Hardened consumer app W17 - Built the Work OS dashboard for Levered - Hardened the Cursor quality loop - Codified more Levered operator skills - Advanced client ops and tracking workflows Day-to-Day Workflows - Pulling client meeting context from Granola - Filing call notes, summaries, and transcripts - Updating client masters, timelines, and standups - Turning meetings into next actions - Syncing open items into Work OS - Tracking client commitments and blockers - Preparing weekly debriefs and Monday prep - Running analytics / ads / SEO check-ins - Capturing site migration open items - Running consumer app e2e tests - Debugging lead alerts, voice, and SMS flows - QAing pages before launch - Building readouts, decks, and client briefs - Keeping knowledge-base daily logs current - Preserving session handoffs across chats

Using skills saves you money – fix your harness

deleting linear again lol

Cursor multitask is 🔥

btw Zoom is here now too

build a better harness they said

@lingxi Please make a hotkey to move to the next unread chat that also works with the sidebar closed

shoot me any feedback about the keybindings (keyboard shortcuts) in cursor 3 agent window. be critic and harsh please. all ears!! 👂

🥺 looks like I have to try this

skills are so powerful my Cursor agent just prevented a future Cursor agent from making a mistake by leaving a detailed skill write up

Sequoia's thesis that the next $1T company will sell work, not software, is the most important reframe in AI right now. The argument: if you sell a copilot, you're competing with every new model release. But if you sell the outcome — books closed, contracts reviewed, claims handled — every AI improvement makes your margins better, not your product obsolete. The key insight most people miss: for every $1 spent on software, ~$6 is spent on services. The entire SaaS playbook was about capturing the software dollar. The AI playbook is about capturing the services dollar — at software margins. Not "AI for accountants." The AI accounting firm. Not "AI for lawyers." The AI law firm. The companies that figure this out won't look like SaaS companies. They'll look like services firms rebuilt on software infrastructure. That's a fundamentally different company to build, fund, and scale. And most founders are still building copilots.

Sick! Can’t wait to try this out

You have to become more efficient with token usage or it’s not worth the investment. If you find that you’re having ineffective sessions (e.g. complete restarts, iterations, unused code) there’s something wrong with your process – not the model, not the IDE. I spend at least 20-30% of my time in Cursor updating my “out-of-session” context, tweaking operations, and measuring performance. I’ve used Loveable, Cursor, Codex, Claude, TraeIDE, NotebookLM, all of them. I can confirm the key to a successful output mostly relies on the user, not the tool. The Don’t Repeat Yourself (DRY) principle is more important than ever right now when you’re billed by the character. DRY in practice makes me view things like portable knowledge bases and skills as no-brainer must haves out the gate. When I work on anything, it’s already designed to be accessible, independently iterable, and a source of truth. This helps manage context much more efficiently and creates more successful outcomes.

but I still run into this a lot

Composer 2 is much faster than Comet Browser Assistant I just wish it could sign in with Google

Tried linear 2-3 times and still just keep going back to regular to things 3 for everything. Maybe I’m using it wrong

wonder what this will do

chats are ephemeral anyway – we should just be able to jump to the next "ready" convo or the last convo what's the point of grouping them for storage

@daradoescode @cursor_ai yeah that was really killing me – literally the only reason I didn't use glass when it launched

I love Cursor and I think the ppl there are building something great, which is why I have to say this: @cursor_ai the side bar in Glass, genuinely makes no sense rn why the hell does sorting my threads by recent remove the entire workspace grouping? And the workspace grouping absolutely doesn't make sense either I have 2 folders pointing to diff branches on the same repo, why are they in the same group even though they are different folders and different branches ?? (you should not group workspaces by repo entirely, ik this is bec of cloud agent stuff but it makes actually using it Glass rlly frustrating) funny thing is you can't even tell what folder you're in unless you choose to open the terminal Also design mode does not work when you're starting a new thread, weird but okay? you guys have something great here, which is exactly why this is so frustrating

going to try Composer 2 only for the next 30 days will explore adding some prompt wrapping with Superwhisper to see if it can synthesize my voice-to-text ramblings better and provide clearer instructions and context