Read more

Burned my usage limit faster than expected — 4.7 costs more per token than 4.6. Switched to Gemini to get things done.

Hey! I’m Alena — ex-AI-startup CEO ($2M raised), ex-PM at Yandex and Acronis. 11k L…

Read more

Last week I was upgrading our startup’s UI while one agent scanned industry trends in the background and Claude chewed through transcripts from a user interview and an adviser meeting we’d just had. Normal Tuesday.

Then the screen told me I’d hit my limit on max Claude’s Max plan.

So… I read Dostoevsky. Spent additional time with my 1-year-old. Baked a quiche. It turned out a very good break, but…

I thought I was doing something wrong. Then my friends started showing off. “Hit my limit by 2pm today.” It’s a flex now.

Uber burned through its annual AI budget a few months into 2026. Engineering max plans run $100-200 per person — one CTO ran $600 bills. My number wasn’t high. It was the floor.

You don’t plan to overspend on AI. Or on Sicily vacation. The bill comes anyway.

Where 99.93% of your tokens go

So, my first instinct was to blame myself — maybe my mempalace plugin config was bloated, maybe my system instructions were too long. I trimmed both. Still hit the limit.

I asked Claude to audit its own token usage:

It re-read and re-wrote again and again.

A developer who audited 30 days of Claude Code put a number on it: 1,310 re-reads per 1 productive token. 99.93% of what I was paying for was the machine reading what it already knew.

I kept pulling the thread. One wall became a stack.

The routing tilts expensive: Claude Code sends 93.8% of tokens through Opus, the priciest model in the lineup.

A friend sent me a Hacker News screenshot. Boris Cherny, Claude Code creator, was explaining why bills spike: million-token contexts get expensive fast, and background agents trigger surprise charges for “a surprisingly large number of users.” The thread had 750+ points and 650+ comments.

Agents are voracious — agentic workflows use 5-30x more tokens than a chatbot, multi-agent setups about 15x more than a single chat. They don’t know which files they need, so they read everything — 60-80% of tokens wasted just locating code. Every MCP plugin ships its tool schemas with every call; one team found 72% of their context window eaten by tool definitions before the agent saw a single message. I was running 30 agents with mempalace, GitHub, and Slack plugins. I ran around 30 agents daily. The math was not in my favor.

Thinking tokens bill at output rates, 5x input. Self-attention scales quadratically: double the context, quadruple the compute. Each layer multiplies through the others faster than I could trim.

Economists call this Jevons Paradox. a16z calls it “LLMflation”. Cheaper units don’t reduce consumption. They unlock it.

At the same time Anthropic pays even more.

One user calculated their virtual cost at $1,274 a month — on a $200 plan. Anthropic eats the difference today; that subsidy won’t last.

DeSight Studio’s analysis puts the platform-side number at roughly $5,000 of inference compute burned per $200 Max subscriber. That’s not a small gap a price hike closes; it’s 25x. Every heavy user is being underwritten by venture capital and the implicit promise that someone, eventually, pays full price.

Designing around the bill

When my token bill kept climbing, I stopped trying to push it down and started designing around it.

So I tried the obvious fixes first.

Killed half my skills and plugins — the ones I’d installed once and forgotten. Cleaner config, it helped a bit. The overhead wasn’t in the tools I’d stopped using; it was in the ones I used every day.

Compacted earlier, cleared more often. Middle of a pitch deck final, I ran /compact too soon — the thread I’d been chasing got summarized into a sentence, and the investigation was gone. I spent the next hour rebuilding what I’d just discarded.

Routed the cheap work to Haiku. Extracting action items from a user research transcript seemed like pattern-matching. It wasn’t. Haiku returned generic bullets that missed the three decisions the customer had actually announced. I reran on Opus, ate two passes instead of one, plus the time I’d lost trusting the first output.

Switched to Cursor one day when Claude locked me out again — same folders, same files, two hours. It was fun, I realized that I’m truly model agnostic as well as my work, but I still prefer Claude.

The token-by-token war was unwinnable. Every fix bought back a few percent that the next agent run spent twice over. I needed a discipline.

Every cheap technology creates an expensive discipline

You read this far — here’s a small thank-you. If you want to keep reading (this post + every past post), your first month is $5.20, that’s 60% off. Feel free to cancel in the first 30 days and you pay nothing more.

The $5 link is open until June 1st. After that, it goes back to full price.

→ Keep reading for $5.20

Read more

Everything you build around the model — the instructions, the tools, the error handling, the pipeline — developers call this the harness. Mitchell Hashimoto, who built Terraform, gave it a simple rule: every time your agent makes a mistake, engineer a fix so it never makes that mistake again.

All of this research comes from the engineering world. But I applied the same principles to my trendwatching agent — and it brought me 85K impressions on LinkedIn in 3 days by finding the hottest.

The Context

(skip if you want to apply harness asap)

When there’s no marketing budget, you ARE the marketing. For me that meant LinkedIn — and LinkedIn meant knowing what’s trending before everyone else. That was 3 hours of my morning, every morning. I tried to automate it. Ten times.

The first six versions were a slow education in what doesn’t work. I started by dumping everything into one Claude conversation — worked until the context window filled up and the model forgot what it had already read. Then I vibe-coded Python scripts with a scoring system, but spent more time calibrating scores than the system was saving me. Then I added 9+ sources across languages and the volume drowned the model in noise — hallucinations, missed items, everything still crammed into one agent’s context.

Version 7 broke me. I tried giving every single content piece its own dedicated agent. Three hours of compute that never resolved. Then the opposite — blind parallel pipelines with no monitoring. 45 minutes, garbage output. Both extremes failed for the same reason: I was engineering around the wrong problem.

Version 9-10 was a full rebuild. Monitored background agents instead of blind subprocesses. Evals at every phase with fallback instructions for when things break. An orchestrator that takes over when subagents fail. The specifics are the seven principles below.

Runtime went from 45 minutes to 15, reliability went from “maybe works” to something I’d actually stake my LinkedIn on — and it’s still the same Claude.

Here are 6 harness principles I applied:

1. Give the AI less, not more

Version 7 ran for 3 hours and never finished because the agent was doing everything — fetching data, routing APIs, handling files, analyzing, scoring, formatting output. I kept throwing more at Claude because that’s what felt natural — more tools, more responsibility, more context — and three hours later I had nothing.

I split the work — Python fetches, Claude analyzes. The scoring agent sees only the scoring rubric and the raw items, no pipeline logic, no historical context, and scores improved immediately.

Vercel learned this too. Their text-to-SQL agent had 17 tools. They deleted 15, gave the model bash and file access. 100% accuracy. 3.5x faster. 37% fewer tokens. The model wasn’t confused by the task — it was confused by the choices.

Birgitta Bockeler from Thoughtworks: “Increasing trust and reliability requires constraining the solution space rather than expanding it.”

2. Evals and Plan B

You read this far — here’s a small thank-you. If you want to keep reading (this post + every past post), your first month is $5.20, that’s 60% off. Feel free to cancel in the first 30 days and you pay nothing more.

The $5 link is open until June 1st. After that, it goes back to full price.

→ Keep reading for $5.20

Read more

A year ago I opened a terminal for the first time since high school — dark screen, blinking cursor, zero buttons.

I don’t think I’m alone in this.

Quick context: a terminal is the text window on every computer (Terminal on Mac, PowerShell on Windows). A CLI is any tool you control by typing instead of clicking. When companies say “we shipped a CLI” — they mean you can run their product from that window.

Everything in this article happens there.

What happened in March

Google put Gmail, Calendar, Drive, Sheets, Docs, and Chat behind a single text interface — a CLI with over 100 agent skills. It hit #1 on Hacker News and pulled nearly 5,000 GitHub stars in three days.

Then Resend shipped 53 CLI commands for email — the founder’s tweet got 310,000 views. What caught my attention: they designed it for both humans and AI agents at the same time.

Visa’s crypto lab made a CLI for AI agents to do card payments without API keys. Sendblue gave agents their own iMessage number, Kapso did the same for WhatsApp.

Then Stripe released Projects: hosting, database, auth, analytics, and payments — each set up with one command.

All of this in four weeks. I don’t know if these companies coordinated or if they all looked at the same usage data and reached the same conclusion.

But six major launches in one month was hard to ignore.

What it means

Stripe didn’t kill its dashboard — you can still click through it. But they built Projects because agents can’t click buttons, and agents need text. So Visa kept its existing interface and added a CLI alongside it, and Google did the identical move with all six Workspace apps.

Old UIs aren’t going anywhere. But every major platform added an entrance for agents next to the one for humans. Which door gets more traffic? Too early to tell. But when Stripe, Visa, and Google all ship the same architecture in the same month, they’re not guessing.

One detail convinced me this isn’t a developer-only trend anymore: Amp launched an ad-supported free CLI, and companies like Axiom, Chainguard, and WorkOS are paying for ad placements inside a terminal.

When advertisers spend money on a CLI, the audience is big enough to matter.

AI coding hit $7.4 billion — so ads in a command line shouldn’t surprise anyone. But it surprised me.

So what does this mean if you’re not a developer?

Who this is for (and who it isn’t)

If you’re using Lovable, Replit, or Cursor with the visual editor and it covers what you need — you probably don’t need a terminal right now. Those products are improving fast and they exist specifically so you never touch the command line.

But if you’re already talking to Claude, Copilot, or ChatGPT and hitting walls — can’t point it at your actual files, can’t chain tasks, can’t let it run while you do something else — that’s the gap terminal fills.

There’s also a timing factor. New capabilities land in terminal first — sometimes weeks before they reach the desktop apps. Right now I count over 15 features that are terminal-only. So if you want early access to what’s coming, not what shipped last month, that’s where it lives.

I’m not a developer — never needed a terminal, and every visual interface I had covered my work fine.

Then I started wanting things they couldn’t handle. Like feeding Claude a folder of 40 meeting transcripts and saying “build me a table of every commitment anyone made.” You can’t paste 40 documents into a chat window.

This reminds me of spreadsheets. In 2005, most managers had someone else handle their Excel. Five years later, if you couldn’t build a pivot table, meetings moved on without you.

Nobody planned that shift. The tool crossed a line where not knowing it cost you.

Onboard yourself

Time investment: about 20 minutes for your first session, install included.

You won’t need to google commands — press / and Claude lists every option, or ask “how do I do X” and it explains. A software engineer friend told me the same thing: with an AI-powered CLI, nobody memorizes syntax anymore. The tool knows it.

Budget a month before the whole experience feels natural.

Open your terminal on Mac and paste this: curl -fsSL https://claude.ai/install.sh | bash

On Windows PowerShell: irm https://claude.ai/install.ps1 | iex

Enter claude and you’re in.

At a new job you get a first assignment and figure everything out from there — so why not treat this the same way?

Organize your screenshots. Everyone has a folder of hundreds of files named Screenshot 2025-11-26 151454.png. Tell Claude:

Look at every screenshot in my screenshots.
Rename each one based on what's in the image.
Group them into subfolders by topic.

It opens each image, figures out the content, renames it, and sorts it into subfolders. A hundred screenshots takes about 8 minutes and 10 parallel agents — so you point it at the directory and go make coffee.

The process:

The result:

Now your Downloads folder. Enter /plan first — this switches Claude to read-only mode where it previews what it would do without touching anything:

/plan Go through my Downloads folder.
Group everything by type.
For anything older than 6 months create a folder "Archive".

It lays out the proposal — here’s what I’d move, here’s the directory structure. You review it, adjust if something’s off, then switch back to normal mode and approve.

The Downloads folder has ~2,500 files and ~113 subdirectories accumulated over years — that's what Claude told me.

The Plan she created to organize it:

Took 3 mins to sort it.

Now look at what it found. There are probably meeting transcripts in there — recordings, notes, PDFs from calls. This is where it gets interesting, because you probably have 20 or 30 of them and reviewing each one manually would take hours.

This is where agent teams come in — an experimental terminal feature where Claude automatically spins up parallel agents to divide the work. To enable: ask Claude to set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in your terminal.

Then tell Claude:

Find and read all meeting transcripts.
For each one: who attended, action items, who owns each one,
deadlines, open questions, conflicts between projects.
Combine everything in one table.
Then create a task list grouped by owner.
For anything with a deadline this week, draft a calendar invite.

Claude suggests creating a team and waits for your OK. Or you can ask it directly. Once you approve, it assigns a lead agent that spawns teammates — each one picks up a few transcripts, they message each other directly through a shared task list, and the results merge into a single table.

My team:

The difference from regular subagents: subagents just report back to one main agent. Teammates actually coordinate with each other — one finds a conflict between two projects, pings another who’s reading the related transcript, and they resolve it together.

You can cycle between them with Shift+Down to watch progress, or run each in a separate pane if you use tmux.

That’s one use case. I wrote about [the most popular ones people build — link TBD] and the number one was a full second brain: organized directories, agents that maintain your notes, weekly synthesis, and automatic tagging. It starts exactly like this — one messy folder, one assignment, then it grows.

Commands worth knowing

You read this far — here’s a small thank-you. If you want to keep reading (this post + every past post), your first month is $5.20, that’s 60% off. Feel free to cancel in the first 30 days and you pay nothing more.

The $5 link is open until June 1st. After that, it goes back to full price.

→ Keep reading for $5.20

Read more

Dave Ricks, CEO of Eli Lilly (a leading pharmaceutical company with market cap $934 billion in early 2026) uses Grok to ask scientific questions.

Wade Foster, co-founder & CEO of Zapier, uses it “to help source under the radar talent is Grok… We’re trying to find some good social media candidates…”:

Grok is like a people finder — is a really helpful tool… You can do the same way if you wanted to find people who just might have product feedback for me… or for influencer marketing.

People agree – ChatGPT lost 35% of active mobile audience last year and Grok grew from 0 to 15%.

Why so?

Grok has real-time access to X (Twitter) data + web search in the same model. On X people think out loud, before they leave reviews, before they update their LinkedIn. It's the closest thing to unfiltered opinions you'll find online. So market research that used to take 4-6 hours now takes minutes.

And they just launched 4.20 – a multi-agent assistant.

Instead of one model thinking sequentially, Grok 4.20 runs 4 agents in parallel that debate each other before responding: Grok (Captain), Harper (Research), Benjamin (Logic), Lucas (Creative).

Honestly, I don’t really see the difference. It was pretty good at social search and it still is.

These 6 prompts are designed specifically for Grok. They won’t work the same way in ChatGPT or Claude because they depend on X data that only Grok can access.

How Grok’s DeepSearch works (so you understand why these prompts are structured this way):

• It breaks your question into sub-searches, pulls from 10+ sources across the web and X, checks them against each other, and gives you a sourced report with the reasoning visible.

• It naturally includes source links — but you need to explicitly ask for structured tables, otherwise it defaults to prose with inline citations.

• Each prompt below includes a follow-up you run in the same thread to deepen the initial findings.

1. Trend watching & industry pulse

The prompt

Research the most significant emerging trends in {{YOUR INDUSTRY}}
over the last {{14 / 30}} days.

I'm a {{YOUR ROLE}} making decisions about {{WHAT: product roadmap /
investment / content strategy}}. I need signal, not noise.

Search specifically for:
- X threads where {{INDUSTRY}} practitioners debate new approaches
- Recent funding rounds, acquisitions, and product launches
- Shifts in how people talk about the space (new terminology,
  new complaints, new excitement)
- Signals from both web sources AND X conversations — compare
  what media covers vs. what practitioners actually discuss

EVIDENCE TARGET: Find at least 20 distinct data points
(X posts, articles, announcements, threads). Each must have
a source link. If you find fewer than 20, note which searches
returned thin results so I can adjust.

FORMAT YOUR RESPONSE AS:

**Trend table** (one row per trend, aim for 8-12 trends):
| # | Trend | What's happening | Signal strength (Strong / Emerging / Noise) | Key evidence | Source with link |

Signal strength criteria:
- Strong = multiple independent sources + practitioner discussion on X
- Emerging = few sources but high-quality signals
- Noise = media coverage but little practitioner engagement

Then after the table:
1. **Overhyped**: One trend with lots of buzz but weak evidence
   when you look at actual practitioner reactions. Explain why.
2. **Underrated**: One signal most people are probably missing.
   Link to the source that convinced you.
3. **Evidence count**: State how many distinct sources you found
   and how many were from X vs. web.

Follow-up prompt (same thread)

Now go deeper on {{MOST INTERESTING TREND FROM THE TABLE}}.

Search X for the most substantive posts and threads about this
in the last 2 weeks — practitioners with skin in the game,
not media accounts.

TARGET: Find at least 15 more X posts or threads I haven't
seen yet. Different authors, different angles.

Present as a numbered list:
1. [Author name / handle] — [Key point they made] — [Link to post]
   Context: [Why their perspective matters]

Then summarize: Who agrees? Who disagrees? What's the strongest
argument on each side?

Running total: You should now have {{20 from first pass}} +
15 new = 35+ evidence points on this topic.

Result:

2. Find your audience

You read this far — here’s a small thank-you. If you want to keep reading (this post + every past post), your first month is $5.20, that’s 60% off. Feel free to cancel in the first 30 days and you pay nothing more.

The $5 link is open until June 1st. After that, it goes back to full price.

→ Keep reading for $5.20

Read more

We all know the drill. You need to understand what real people think about a topic — not the landing page version, not the founder tweets. The real stuff. What people say in X threads when they’re frustrated, excited, or comparing tools at 11pm.

Normally this takes half a day.

Search, scroll, read replies (the real opinions are always in the replies), copy quotes into a doc, try to see patterns. By the end you’re fried and you’ve barely covered one angle. And you still don’t know what real people think — just the landing page version, not the 11pm X threads where people are actually honest.

Last week I typed one sentence into the Claude browser extension and went to make coffee.

Came back to a good enough breakdown. Pain points, what people love, what’s broken. Pulled from real threads.

Here’s the setup, what worked, and what you should watch out for.

What are AI browsers

An AI browser browses the internet like you do. Clicks, scrolls, reads pages, follows links. You give it a task in plain language. It figures out the steps.

Two worth trying:

Claude in Chrome — Anthropic’s extension. Side panel. Type a task, approve the plan, it runs. Can schedule tasks to repeat weekly. Needs a paid Claude plan.

Comet by Perplexity — Standalone browser. Same idea, free for Perplexity users.

Others are popping up fast (Fellou, Sigma). This workflow works with any of them.

Key thing: these aren’t scrapers. They navigate pages like a human — so they work on sites with no API at all. Like X.

The AI workflow loop

Most people think the hard part is writing a good prompt. It’s not. The hard part is knowing what to ask next.

You don’t write one perfect prompt. You run short iterations — get something back — use that to decide the next one. Each iteration sharpens the next.

5 stages: Explore → Narrow → Deepen → Lock → Maintain

Here’s how it played out when I researched no-code AI agent tools on X.

1. Explore

Open x.com, launch the Claude extension, and type something like:

“On x.com find what people are saying about no-code AI agent tools — complaints, praise, questions, comparisons, everything”

No angle. No hypothesis. You’re mapping the conversation, not searching for a specific answer.

Claude shows a plan before doing anything:

1. Search X for threads about no-code AI agent tools

2. Try additional queries like “no-code AI agents experience” and “no-code agents honest review”

3. Read through threads and replies across all sentiments

Read the plan. If it makes sense, approve it. (Why you always read the plan — danger zone section below.)

Go make coffee.

Come back to a map. Mine showed:

It won’t be perfect. But now you know what people actually talk about — not what you assumed they talk about.

2. Narrow

Pick the most interesting cluster from Explore. Zoom in.

My Explore showed the criticism cluster was the richest and most specific — users burning $50 in credits before an agent even runs, security teams worried about no-code tools shipping faster than threat modeling, platforms that help you launch but not scale, an AI agent publicly accusing a project of fake vulnerabilities. Four different problems, all with names and numbers attached.

So:

“Go deeper on the criticism around no-code AI agent tools. Focus on cost, reliability, and what happens after launch. Find what specific things break, what the actual costs look like, and who’s most affected”

3. Deepen

You read this far — here’s a small thank-you. If you want to keep reading (this post + every past post), your first month is $5.20, that’s 60% off. Feel free to cancel in the first 30 days and you pay nothing more.

The $5 link is open until June 1st. After that, it goes back to full price.

→ Keep reading for $5.20

Read more

Become a Founding Partner

Why it matte…

Read more

Hey! I’m Alena, former AI startup CEO ($2M raised), Yandex and Acronis Sr. PM with 10k LinkedIn followers, 57% of my followers are senior leaders from big tech, good company ;)

There are tasks that should be delegated to AI.

And then there are tasks where you need to build a custom project – one that works for you, and only you.

(I’ve included the full free template for AI personal assistant at the end of this post. I spent over 50 hours debugging the prompts so you can set this up in 5 minutes.)

There are about 100 popular AI-assisted to-do lists popping up right now.

Teresa Torres, author of Continuous Discovery Habits and an internationally acclaimed coach, explains this proliferation simply:

“How we manage our tasks is so idiosyncratic that this is exactly the type of thing that you should build for yourself because it can work exactly the way you want it to work.”

I couldn’t agree more.

I’ve been setting personal and professional goals for the last 10 years. I rarely achieve even half of them. Usually, things lose relevance, or I simply forget they exist. This year, I decided to outsource the executive function to AI.

Imagine having a personal assistant with the skills of an executive coach, ready to help you 24/7 for $0 a month. Top CEOs have Chiefs of Staff. Professional athletes have performance coaches. Why do we expect ourselves to manage our complex lives alone?

The assistant that doesn’t just manage your calendar. It manages you. It knows when you’re lying to yourself about “just checking email.” It knows when you’re catastrophizing. And it knows exactly how to get you back on track without making you feel guilty.

But first, I did some research to find the most scientifically validated techniques for goal setting and tracking.

The science

Part 1: How to set and track goals

It turns out, simply “set SMART goals” isn’t enough. Here are the four evidence-backed frameworks I’m using:

1. Implementation Intentions (Gollwitzer, 1999)

The structure is simple: “If [Situation], then I will [Action].”

• Example: “If it is 9:00 AM on Monday and I open my laptop, then I will write for 25 minutes before checking email.”

2. WOOP (Oettingen, 2014)

Wish → Outcome → Obstacle → Plan

• Wish: What do you want?

• Outcome: What will it feel like?

• Obstacle: What internal barrier stops you?

• Plan: If [Obstacle], then [Action].

Fun Fact: Positive visualization without visualizing the obstacles actually reduces the likelihood of achieving the goal. You need to prepare for the resistance.

3. Learning vs. performance goals (Locke & Latham, 2002)

I personally love this concept.

• Performance Goals: Good for familiar tasks (e.g., “Increase conversion rate by 2% in Q1”).

• Learning Goals: Essential for new or complex territory.

  • Bad: “Hit 10k subscribers.” (You can’t fully control this).

  • Good: “Run 5 experiments to understand what drives subscription growth.” (I will learn/understand X).

Fun Fact: The SMART methodology lacks a strong evidence base for complex, ambiguous goals. It often forces false certainty.

4. Written Goals + Accountability (Matthews, 2015)

The formula: Write the goal → Share with 1 person → Schedule weekly check-ins.

Part 2: How to track goals & maintain motivation

The Weekly Check-in (Harkin et al., 2016)

Action: Every week, ask yourself: “Am I closer to the goal than I was a week ago? If not, why?”

If you lose motivation → Goal Gradient (Kivetz et al., 2006)

Action: Break the goal into 5-10 stages. Show progress as “X% done” (not “Y% remaining”). Make the first 10-20% easy wins to build momentum.

If you forget to act → Implementation Intentions

Action: Write “If [Situation], then [Action]” and place a reminder at the physical trigger point.

If you have too many goals → Sequential Focus

Action: Maximum 3 goals per quarter. 1 MIT (Most Important Task) per day. Finish one before adding another.

If you’re stuck and don’t know why → WOOP Diagnostic

Action: Ask yourself: “What INSIDE ME is stopping this? What feeling am I avoiding?”

Part 3: How I picked goals

To start, I wrote down 100 wishes. It was hard. My imagination ran dry after the first 20, but that’s when the interesting stuff started coming out (e.g., “start taking golf lessons” or “take a course at the Art Institute”).

For the final list, I divided all goals into 3 groups:

1. One-off Projects: Finite tasks with a clear end.

2. Pillars: Ongoing areas requiring constant engagement (Startup, Consultancy, Personal Brand).

3. Side Quests: Everything else – from learning to make pizza to visiting my 41st country.

Important: I intentionally did not separate “Work” and “Life.” Work-life balance is one of my goals this year, and I plan to achieve it by freeing up more time for rest using AI.

The build

Here is how I actually built it:

1. Created a new project. I use Cursor, but you can use Obsidian, Antigraviti, Claude Projects, ChatGPT Projects or just a folder on your desktop. The magic is that it’s just text files – no vendor lock-in.

2. Uploaded my goals. I literally took a photo of my notebook and had Claude transcribe it.

3. Created a Context File. This defines work, family, time constraints, and cognitive distortions.

4. Uploaded Frameworks. I fed the AI the scientific frameworks for goal setting and tracking (WOOP, etc.).

5. Tasked Claude Code. I asked it to organize the project structure based on all inputs.

6. Tasked Gemini 3 Pro. I asked it to review the structure.

7. ...Several iterations of edits...

8. ...2 weeks of debugging... (I added a simple database for task tracking and a calendar database).

Profit.

The final structure

goals-template-2026/
├── Vision_2026.md          # The main map (generic categories)
├── Memory.md               # AI Instructions (Persona & Rules)
├── README.md               # Setup Guide
├── Pillars/
│   └── pillar-template.md  # Template for recurring systems
├── One-off projects/
│   └── project-template.md # Template for finite projects
├── Side quests/
│   └── quest-template.md   # Template for hobbies/fun
└── frameworks/
    └── project_template.md # The master template to copy

The context

I added two critical layers of context to my system:

1. Cognitive Distortions: The psychological patterns I’m currently working on.

2. Life Constraints: My reality – like having a small child, which means plans often change unexpectedly.

I also applied the concept of Progressive Disclosure when designing my AI coach.

The result

I now have a Coach/PA that helps me:

• Set goals in a way that makes them achievable.

• Track those goals without the manual headache.

• Identify blockers (both external and internal).

• Overcome demotivation.

• Work through my cognitive distortions in real-time.

• Keep all related notes and information related to each project.

• Balance work and life.

• Work through my cognitive distortions in real-time.

• My most favorite feature so far: organize meeting transcripts into tasks and allocate to appropriate projects.

You read this far — here’s a small thank-you. If you want to keep reading (this post + every past post), your first month is $5.20, that’s 60% off. Feel free to cancel in the first 30 days and you pay nothing more.

The $5 link is open until June 1st. After that, it goes back to full price.

→ Keep reading for $5.20

Read more

This week I went deep on AI usage research — and the data is surprising. Turns out 52% of open-source AI usage is... roleplay. Gemini has better retention than ChatGPT. And “act as an expert” prompting? Wharton proved it doesn’t actually work. Microsoft dropped a similar usage report. It’s benchmark panic season.

Read more

Interviews are forgiving — you adjust questions on the fly, you follow up, you rephrase when someone looks confused.

Surveys are unforgiving.

You can’t follow up. You can’t rephrase. You can’t run the same survey twice on the same person. If something’s broken, you find out after the data is useless.

You want to ship a viable version on the first try. Like an MVP, but actually viable.

But there’s negative time allocated for this. So people do it half-heartedly, cross their fingers, and hope the data makes sense.

Hey! I’m Alena, former AI startup CEO ($2M raised), Yandex and Acronis Sr. PM with 10k LinkedIn followers — 57% of my followers are senior leaders from big tech, good company ;). I’m writing about use-case first AI implementation and tools. This issue: how to do it fast without losing quality — using AI.

Subscribe now

AI-native flow

1. Design — Goal + draft questions

2. ✨ Validate with AI — Synthetic personas answer → blind spots → refine

3. Validate with humans — Friends/colleagues → refine

4. Launch & learn — Real data → blind spots → iterate

5. ✨ (bonus) Talk to your data

6. (bonus) Present to stakeholders

The acceleration happens at the design stage. The magic — on synthetic user validation and the ability to literally talk with your data.

Step 1: Design

Traditional: Write research question → align with manager → draft questions.

AI-native: Use AI to structure goals → align with manager → use AI to draft questions

Goal Setting — AI as thinking partner

My favorite framework is Working backwards but you can choose yours, the main idea to ask AI to challenge your ideas. Do not fall into infinite feedback loop though. 3 iterations is max.

Start with context: “I want to use [your framework] to set objectives for my survey. Product: [your product]. Description: [your product]. Target audience: [fill]. My goal: [fill].” Attach all relevant docs you have.

You’ll have more grounded and bulletproof case for your manager.

Frameworks: PR/FAQ, JTBD

Example of conversation for our market research and social listening AI agent.

Draft creation — AI generates, You curate

AI knows survey design rules. You validate.

1. Start with context: Feed final output from previous step.

2. Clarify constraints

3. Ask about best framework to fulfill the goal

4. Start new conversation and feed in the full context (to avoid context rot)

5. Simply ask to create questionnaire.

6. Start new conversation, fill in context and ask to review (AI call flags issues: “This might be leading because...”)

Principle: Split jobs. The more granular, the better the results. Do not do more than 3 iterations.

My example:

✨ Step 2: validate with AI (magic 1)

You read this far — here’s a small thank-you. If you want to keep reading (this post + every past post), your first month is $5.20, that’s 60% off. Feel free to cancel in the first 30 days and you pay nothing more.

The $5 link is open until June 1st. After that, it goes back to full price.

→ Keep reading for $5.20

Read more

You’ve surely heard about MCP over the last year and about Skills for the past 2 weeks, both concepts by Anthropic, both can be used to increase your AI capabilities. MCP was still a pretty “dev required” thing—you need an API key at least. Turned out that Anthropic realized that MCP doesn’t really fulfill user’s needs and came up with the Skills concept. And like MCP, it can be used with any AI; I tested it in Cursor. So you don’t really need to lock into Claude—think about it as an instructions template that is highly consumable for AIs.

In this issue I want to show you what Skills are at their core, best practices and limitations to consider, and how to find, build, customize and manage skills.

What’s Inside

• The Matrix moment — Why Skills changes everything you know about AI automation

• Token economics — How to cut costs 80% while fixing hallucinations

• Skills vs MCP — Why Anthropic quietly abandoned their own protocol

• 30-minute build — Your first Skill that actually works (Meeting Notes → Action Items)

• Security blindspots — What malicious Skills can do + protection protocol

• The Cursor edge — Why some teams get 3x better results

• Early adopter notes — Current limitations & workarounds that work

What are skills from the first principles

Remember how Trinity in The Matrix downloaded a helicopter piloting Skill? That’s pretty much it. If your Agent (can be a general agent like ChatGPT or specialized like Claude Code) has a skill, whenever you set a task that can be solved with that skill it will find, download and use it on demand.

If you have repetitive tasks or prompts that you store somewhere and paste every time you need to perform a specific task, Skills is the solution. Instead of explaining your process every single time you need something done, you create a Skill once – and Claude remembers it forever. Need a weekly status report? There’s a Skill for that. Want consistent brand colors in every presentation? Create a Skill. Have a specific way you analyze customer feedback? Package it as a Skill.

A Skill is a folder containing:

1. A SKILL.md file - This is your instruction manual. It tells Claude what this Skill does, when to use it, and exactly how to do it. You write this in plain text using simple formatting.

2. Optional extras - You can add template files, reference documents, example outputs, or even small scripts if you’re comfortable with that.

At its core, it has the idea of progressive disclosure (like in The Matrix) meaning that at every single moment it uploads only the necessary amount of information (=tokens), which saves you from two main problems:

→ Hallucinations and mistakes because it doesn’t contain huge unrelated context

→ High costs because it doesn’t upload all context, only what’s necessary

Here’s how it works in practice. When you start a conversation with Claude, it sees a short description of each available Skill (like “Creates weekly team reports with wins, blockers, and priorities”). These descriptions are tiny – they barely use any of Claude’s “memory” for the conversation.

When you ask for something that matches a Skill description, Claude thinks “Oh, this person needs a weekly team report!” and then – and only then – it opens up that Skill folder and reads the full instructions. It’s like having a filing cabinet where Claude can instantly grab the right procedure when needed.

What makes Skills different from other features:

• Custom instructions - Those brief, always-active settings apply to everything. Skills are detailed procedures that activate only when relevant.

• Projects - These are workspace-specific and accumulate context over time. Skills are portable recipes you can use anywhere.

• MCP (Model Context Protocol) - This connects Claude to external data sources like your CRM or Google Drive. The main problem with MCP (Anthropic just admited it in this article is token consuptions – it was the main reason to shift paradigm from MCP or projects (also called Agents in Claude Code) to Skills.

Use Cases

You read this far — here’s a small thank-you. If you want to keep reading (this post + every past post), your first month is $5.20, that’s 60% off. Feel free to cancel in the first 30 days and you pay nothing more.

The $5 link is open until June 1st. After that, it goes back to full price.

→ Keep reading for $5.20

Read more

How’s it going?

Last week I tested Claude Skills - more handy and transferable than projects. Projects are still more manageable for complex work. The interesting part: these skills work as portable prompts you can adapt for other models. Also discovered a benchmark showing AI agents fail 97.5% of real freelance work. Reality check deliv…

Read more

Following up Lenny’s LinkedIn poll on how people use Claude Code for non-coding tasks. I think it was brilliant conversation and we can learn so much from that thread. Check the original post here.

Teresa Torres (bestselling author) wrote this:

I now write all of my content with Claude Code in VS Code. We iterate on an outline, it helps me improve the hook, it conducts research for me and adds citations to my outline, and it reviews and gives feedback on each section as I write.

Teresa doesn’t code. She writes books. Using a “coding tool” for everything.

So I scraped 159 comments from professionals using Claude Code. Categorized every single one. Let’s take a look which story data tells us:

What’s Claude Code and how to get started

Claude Code is Anthropic’s AI coding assistant that lives in your terminal or VS Code. But here’s the twist - it’s not just for coding. It reads, writes, and organizes any files on your computer. Think of it as having a smart assistant who can see your entire file system and help you work with it.

Getting started takes 3 minutes:

1. Open your terminal and install Claude.

2. Open it anywhere

• In your regular terminal: Just type claude

• In Cursor: Open the terminal tab (inside Cursor) and type claude

• In VS Code: Same thing - open terminal, type claude

3. Log in (first time only)

4. Talk to it

Just describe what you want:

• “what does this code do?”

• “fix this bug”

• “add a login button”

Done! Claude edits your files right there.

Everyone is talking about context engineering, I believe it’s the easiest way to start use it as your tool.

Who’s using it

🥇 Product Managers: 38 people (24%)

🥈 Founders/Co-founders: 30 people (19%)

🥉 Engineers for non-coding tasks: 24 people (15%)

What they’re building

1. Content Creation, 31 mentions

2. Knowledge Management (aka Second Brain), 28

3. Product Management, 24

4. File Organization, 18

5. Research & Analysis, 17

6. Data Processing, 15

7. Personal Productivity, 14

8. Business Operations, 12

9. Workflow Automation, 11

10. Documentation, 10

Let me know your thoughts below and I’ll feature best comments in the next issue on this topic.

Subscribe now

Now let’s dive into exactly how people use each category - with real examples you can implement today.

5 most popular use cases

01 Content Creation & Marketing (31 Users, Highest Engagement)

The Problem

Writing takes forever. Research is scattered. Quality is inconsistent.

Teresa Torres isn’t alone. 38 people shared similar workflows.

12 agents working sequentially from newsletter research to publish-ready article with gated PDF/Notion template...Repurposing all content (published/research) for Twitter, Substack notes, Medium, Reddit through specific agents

Ayush Poddar

Framework based on comments.

Stage 1: Capture → Everything becomes markdown

Stage 2: Process → Specialized agents extract value

Stage 3: Create → Multi-iteration refinement

Stage 4: Multiply → One source → many formats

Stage 5: Compound → Everything feeds back → Next content uses accumulated knowledge

Edgar’s Quality Gate System

• Draft from docs → Rate 1-10 → If <9, loop back with fixes

• Bans overused AI words (”seamless”, “leverage”)

• Publishes directly to Notion with proper formatting

File structure example:

content-system/
├── 00-inbox/              # Raw captures
├── 01-research/           # Sources, transcripts, notes
├── 02-ideas/              # Brainstorms, outlines
├── 03-drafts/             # Work in progress
├── 04-published/          # Final pieces
├── 05-templates/          # Proven frameworks
├── agents/                # Agent instructions
└── analytics/             # Performance data

Your agent team

Core Agents

1. Librarian Agent: Organizes files, creates cross-links

2. Research Agent: Web search, competitor analysis, pattern extraction

3. Writer Agent: Knows your voice, past content, frameworks

4. Editor Agent: Enforces quality rules, scores output

Specialized Agents (task-specific)

• Hook Optimizer: Tests 5-10 opening variations

• SEO Agent: Keyword research, optimization

• Visual Agent: Quote cards, diagrams, thumbnails

• Distribution Agent: Formats and posts to platforms

• Analytics Agent: Tracks performance, identifies patterns

Results

• 4-6 hours → 45 minutes per piece

• Teresa Torres: Iterates on outlines, gets research with citations, receives real-time feedback on each section

• Ayush Poddar: 12-agent pipeline from research to multi-platform distribution

• Edgar: Quality gates ensure consistent 9/10+ output before publishing

But the most valuable source from the discussion is a GitHub repo containing a fully functional LinkedIn content creation machine that you can copy:

https://github.com/mslavov/linkedin

In the next sections, I’ll break down the remaining 4 popular use cases. There’s a special offer – a 60% discount and the LinkedIn Optimization Playbook for free. The methodology has already helped 3,000 professionals increase recruiter visibility up to 8x and land their dream jobs, from new grads to CEOs.

Feel free to unsubscribe anytime, no hard feelings!

02 Your Second Brain: Knowledge Management (28 Users)

You read this far — here’s a small thank-you. If you want to keep reading (this post + every past post), your first month is $5.20, that’s 60% off. Feel free to cancel in the first 30 days and you pay nothing more.

The $5 link is open until June 1st. After that, it goes back to full price.

→ Keep reading for $5.20

Read more

Subscribe now

Read more