Michael Ewens

@polisisti @aniketapanjwani @DAcemogluMIT This is the skill summary. I'd ask Claude to create a skill for extracting+analyzing an academic paper. "I want you to first convert the PDF to Markdown using MarkItDown" ...and then add your own instructions on how you like the paper summarized including thing to watch out for,

@startupecon @aniketapanjwani @DAcemogluMIT In what file does "line 87" appear? Is this the code for one of the packages, or a private file you have?

How useful is Claude Code's 1 million context window for academics? Yesterday, I tried pushing Claude Code to do a literature review of as many as @DAcemogluMIT 's papers as would fit in context. Here's what I learned: > CC will try to "get around" actually fitting the full text of papers into its context. It did this silently by reading parts of some papers (often the first 20 pages) and none of other papers, and only revealed it did this after I pressed it after being puzzled by some behavior. > However, the implied token usage by even reading the first 20 pages of many papers should have been much higher. I'm still not 100% sure what's going on - either some sort of auto compaction, or it's just lying. > I got it to "spend" more tokens by first extracting the text from PDFs (imperfectly), and then reading the text directly rather than doing an operation on the PDF. > Context visualization from /context seems broken. After going through this exercise, it shows 481k tokens used, but 962k free space remaining. > I don't see this *at all* as replacing reading papers. The bottleneck is still *human* understanding. I think of the output as a rough map I do not fully trust, but which nevertheless speeds up the pace at which I can understand papers. > I think the optimal approach is something like this: 1. Do a full text extraction with a dedicated skill per paper, one subagent per paper. 2. Use other sets of subagents to do specific types of summaries analyses per paper. 3. Use a final agent or set of agents to combine those summary analyses in a way particular to you. > I also think there are going to be gains from experimentation combining CC/Codex, perhaps Gemini, and also - if you have lots of $ - gpt 5 pro by API via opencode. Made a YouTube video on the new 1M Context - youtu.be/sns-ks-bUJM?si… - the exercise with Acemoglu's papers starts about 4 minutes in the two literature review pdfs I produced in the exercises are her - ai-mba.io/tutorials/clau…

@aniketapanjwani Thanks! You confirmed the basics of my plan and reduced my hesitancy about giving it a shot.

you can solve this by asking codex to create symlinks from agents.md -> claude.md and from your .agents/skills -> .claude/skills then any edit to your claude.md is manifest in your agents.md, and if either tool tries to read or write to your agents.md, it is just reading your claude.md . same idea for skills

interesting results. I've been 90% codex since gpt 5.3 . I only use Claude Code for design tasks now which are irrelevant to economists if you've got CC and also a ChatGpt subscription, try out the Codex desktop app and compare a bit the results you get openai.com/codex/

@ipeirotis Great post. I use Claude Code tasks (the built-in) that persist across sessions (attached). The md task storage system is prone to errors when the context or task is long. The custom system below also allows across-project viewing and is viewable with ctrl+t.

New blog post: "Let's Work on the Next Task: Claude Code, GitHub, and the Most Diligent Project Manager I've Ever Had” A beginner's guide on how to set up Claude Code on the Web to be your project manager, organize your task list (and often handle the tasks in them), and give you a "chief of staff"+"team of RAs" that you did not know you could have. Knowledge of basic GitHub concepts that are commonly used in software engineering (branches, pull requests, and reviewing) is useful but not necessary to get you started. Because Claude Code is not (just) for coding. behind-the-enemy-lines.com/2026/03/lets-w…

@arpitrage "fire the employees" = some % of workers laid off in PE buyouts in the first N years post-close?

@startupecon Happy to take the other side, as I’ve talked to PE folks already doing this!

We are going to see the emergence of a new Private Equity strategy: buyout, fire the employees, and replace with AI

@alexolegimas This must expose your computer/local network to the world (subject to some key sharing), right?

Ugh I spent like 2 hours programming this exact thing using claude code on saturday (but mine was...less good).

@DimitrisPapail Try Zed (multiple projects is great)

Starting to think VS Code is the best GUI for Claude Code.

Claude Code or Codex in Zed (multiple projects in one window)

@ernietedeschi Claude’s new memory solution helps here. It is a small file and can be project specific

In this vein I've started having Claude keep track of its mistakes in a shared markdown doc; every time it makes one, I have it append what happened, why it happened, and general strategies for avoiding the mistake in the future. Claude re-reviews the markdown every session.

@lugaricano Same! I moved everything to Zed code editor (zed.dev), that let's you have multiple projects open with active Claude terminals/chats. All-in-one window has been amazing. Coincidence of a project in the screenshot :)

One of the things that makes working with AI super stressful is continuous context switching. This morning I have worked on six different projects at the same time (while something gets done, work on something else). The cognitive load is huge and it is kind of exhausting.

@PradyuPrasad @aniketapanjwani I force Sonnet[1m] for subagents in settings.json: "env": { "ENABLE_TOOL_SEARCH": "auto", "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-sonnet-4-5-20250929[1m]" } though have not run an A/B test yet.

@aniketapanjwani "how does function X work" "makes sub agent" alt could have been three greps

just ban claude code from using sub agents. experience improves quite a bit

@alexolegimas Agree! Slowly working through “ Rousseau and Revolution”. So jarring!

@startupecon The think the other issue is that so much of the writing we're consuming is now AI, even if you're doing the writing, you might start sounding like AI. Hence the reading old fiction haha.

This essay “The Last Temptation of Claude” by Harry Law is excellent. This part stuck out: de skilling doesn’t just happen. People dont make one choice—it’s hundreds and thousands of little choices. Off load a memo, off load a section of the paper, etc. they all seem small but then you wake up one day and hey that thing that used to be easy is now quite hard, or even impossible. The temptation is real, so as with all temptation, it’s important to be deliberate and have rules. Everyone will be different, I have mine (write everything on my own, read old fiction)—but they might sound dumb or not work for others. open.substack.com/pub/cosmosinst…

Check out the early draft here: unlockingalts.com/paper

Some interesting implications for major investors (e.g., pension funds, endowments): if retail capital grows as a share of AUM, expect strategy drift toward distribution-friendly, shorter-duration assets. The selection evidence already points in that direction.

Many major PE firms are building a retail wealth channel. With Jacob Faber, we built a dataset of 114 registered funds (totaling $115B+) that offer access to private assets. The findings are nuanced for the 401 (k) +retirement expansion case. New paper: unlockingalts.com