Renuka Adnani

Had a folder of studio headshots from a shoot 8 months ago, never used them past the headshot itself. Tonight I fed 10 to Higgsfield and trained a face model. Asked for me at the ranch with a horse. First try, ship-ready. Old photos are now infinite source material.

My AI didn't know who "Eric" was on the call I recorded. So I told it: cross-reference Google Calendar by recording date, pull attendees, save the emails. Small skill update. The kind that adds up.

3 calls in an afternoon. 90 minutes of audio. By tonight: 3 structured vault notes, 6 memory entries, an alignment email shipped, 5 LinkedIn DMs drafted. The compression isn't the AI. It's the AI plus the system around it.

Wired up a monthly cron to audit my Obsidian vaults for mis-filed notes. Twist: the cron only proposes. It posts candidates to Slack, I approve in-thread, a paired skill executes. Autonomous where it's safe, human where it counts.

I had a rule in my AI memory: "never use em-dashes." Been there a month. It still shipped 4 today in an email I sent. Memory tells the rule. Grep enforces it. Lesson noted.

Recorded a 5-min Pocket note talking to Claude tonight about a content framework from IG. By the end of the session we'd taught the system to flag strategy convos and produce full vault notes for those. Building tools mid-task.

Asked Claude to inventory my Hormozi material today. 3 books, 6 PDFs across 4 folders, 5 vault notes, plus emails the inbox-triage was silently dropping. 6 months of capture, zero consolidation. Booked Thu for an agent to translate the corpus into PL + RG B2B rules.

the steer feature on codex is sooooo underrated my adhd brain is just stacking todos like anything

codex is killing it - this seems like a serious threat to all the vibe coding platforms. building and sharing apps without going through the hassle of db / auth / deploying, was maybe the only moat vibe coding apps had. time to dive in...

Set out today to render the Maya talking-head clips and hand them to my editor for the actual Reel assembly. 30 min in, asked Claude 'can you do the whole Reel in ffmpeg?' 4 hours later: finished 31-sec IG Reel. Captions, B-roll PiP, kinetic phrase text. One Python script. "AI does prereqs, humans do craft" keeps shrinking.

Built an AI persona with Claude today. HeyGen Avatar IV docs claim 20 credits/min. Real burn on a 25-sec vertical render: 69. ~7× the math. Run one test render before you plan a content cadence. Pipeline shipped by midnight.

Open tracking on cold email looked free. Then read the 2026 data: Apple MPP inflates opens 50%, Gmail flags pixels with a Report Spam button, Woodpecker shows it cuts reply rates 68%. Turned it off. Reply rate is the only signal left.

Discovery call prep this week: AI ran 7 research sweeps. I wrote a one-page thesis. Then I ran it through 9 adversarial lenses myself. Steelman flipped the wedge. Red Team surfaced a gap. The pitch that survived isn't the one I started with.

Wanted to put my face on a LinkedIn cover. AI could not render me accurately. After 7 attempts I dropped the face entirely. The metaphor carried the post better than a portrait would have. Founders default to putting themselves in the picture. The work usually deserves that spot.

I marked a doc "LOCKED" yesterday. Today I found a factual error inside it. The lock kept me from editing. It didn't keep me from being wrong. Lock-status is process state. Validation status is whether it's actually true. Two different things. Easy to confuse.

The pre-mortem question that killed half my plan. Two campaigns launching in 12 days. Almost added 4 LinkedIn lanes on top. Asked: "imagine this failed in 6 months. why?" Answer: I split attention from the launch. Killed 3 lanes. Pre-mortem is the cheapest plan fix I know.

Built my outreach list. Researched 20 companies. Enriched contacts. Wrote sequences. Then realized one of my top targets is my co-founder's employer's sister company. Removed them from every list, every file, every sequence. No gray area. Outreach rule #1: check for conflicts before you fall in love with a prospect.

Built an Obsidian vault system where each co-founder gets their own workspace with shared inboxes. Then wrote a playbook documenting exactly how the skills, routing, and async handoffs work. Shared it with both co-founders. Not because they asked. Because if I get hit by a bus, the system should still run. The strategic part: when your co-founders can self-serve context without scheduling a call, you stop being the bottleneck.

@claruswrite "Judgment about what kind of memory", exactly. I'm doing the same with my agent. It synthesizes transcripts against my vault rules, then asks me when context is genuinely new. The retrieval works. The *classification* is what separates signal from noise.

Connected Read.ai to my Claude setup last night. Woke up to 187 meeting transcripts downloaded to a local folder. Every call I've had this year. Every decision. Every "let's circle back on that." Now they feed into my Obsidian knowledge base. Every meeting becomes searchable context instead of a fading memory. The setup took 10 minutes. The payoff is months of recovered context.