Takshashila Skills

The Takshashila Institution has made public a whole bunch of skills, based on their public policy research and writing over the years. I take a few for a spin.

The more perceptive of you will remember that I used to be a “resident quant” with the Takshashila Institution. Maybe I still am - I never formally resigned, and to the best of my knowledge they haven’t appointed another any time.

Now, over the years (I don’t even know when they formally started - they’ve always existed!) they’ve built up a massive repository of knowledge and skills in the public policy domain. And given that they are a not-for-profit institution, they have decided to share some of the knowledge accumulated over the years.

And so they have shared a bunch of skills. Yes, a public policy think tank can have a github page.

I started using them. I saw that there exists a skill to write op-eds, and so pointed that to Claude, and asked it to choose a random post from this blog, and then write it in the form of an Op-Ed.

The result is impressive and can be seen here (this is a public chat). It even offered to write it in the style of ET, or the Hindu Business Line, or Indian Express. Basically it has been trained on a whole load of research and writing output that Takshashila has put out over the years - and is rather good.

There are loads of other skills - to write grants, to synthesise literature, write policy briefs, etc. etc. And all of this, the result of over a decade and half of work, is now in the public domain! Use it, and write better. And write more. You have no excuse now.

OK in the morning I’d written that I’ll not put any AI-generated content on this blog, but I like the output of this skill so much that here are a few paragraphs of the ET OpEd it generated:

Take a mid-sized Indian company running its analytics on PowerBI or Tableau. The annual bill, between licences, integrator fees, and the dashboard developers sitting in some captive or near-shore team, comfortably runs into low double-digit crores. In return, the company gets several hundred dashboards. When I was building Babbage Insight, I interviewed leaders from a hundred companies about how they used data. Five of them looked at their dashboards daily. The rest had given up.

The standard diagnosis is bad design, bad data, or bad users. All three are partly true. But the structural reason is simpler. A dashboard in Tableau or PowerBI has to serve a roomful of people whose questions are not the same. The CFO wants one cut, the COO wants another, the regional head wants a third. The designer’s response has been to bolt on dropdowns, filters, and drill-throughs - so that each user can, theoretically, navigate to their view. In practice nobody does. You end up with a dashboard that serves nobody especially well, and a quarterly review where everyone exports to Excel anyway.

And the same thing in a policy brief form:

Executive Summary

The Government of India spends an estimated ₹2,000-3,000 crore annually on dashboards, BI tools, and data visualisation platforms across central ministries, state governments, and PSUs. Yet basic policy metrics - “unemployment rate”, “MSME”, “active beneficiary”, “school dropout” - carry materially different definitions across the ministries that report on them. A child counted as “out of school” by the Ministry of Education may be counted as “enrolled” by the Samagra Shiksha portal. This is not a technical glitch; it is a definitional one, and it directly degrades the quality of evidence reaching the Prime Minister’s Office, the Cabinet, and Parliament. We recommend that NITI Aayog establish and govern a central semantic layer for Government of India data - a single, authoritative dictionary of policy metrics and their definitions, which all ministry dashboards and AI-generated reports must consume. The marginal cost of building this has fallen sharply with LLMs; the cost of not building it is rising.

Enjoy. They also have a web app (needs your Anthropic key) here. I prefer to use the skills directly from my LLM.

PS: I asked it to write my previous post as an OpEd. I like it:

The MeitY label is asking the wrong question about AI content

In February, MeitY notified amendments to the IT Rules requiring all “synthetically generated information” to carry a visible label covering at least ten per cent of the surface area, or the first ten per cent of the duration. The intent is sound. The framing is wrong. The rule treats AI involvement, of any degree, as the disqualifying signal for authenticity. The actual question is whether a human being has put their stamp on the output.

I learnt this lesson in algorithmic trading. In 2008, I had joined one of India’s earliest HFT firms, set up shortly after SEBI legalised algo trading. The rules were tight - direct exchange access required your algorithms to be audited by the regulator. As a startup with messy code, this was not feasible. So we used a workaround called “one touch DMA”. Our broker employed a trader whose entire job was to hit a single key the moment our algorithm sent an order through. That one keystroke was enough for the trade to no longer count as algorithmic. There was a human in the loop, and that was enough.

The legal and ethical question with AI-generated content is not whether AI was used. It is whether a human has done the equivalent of hitting that key - read the output, owned it, and chosen to publish it.

Consider the spectrum. At one end, someone pastes a ChatGPT response into a WhatsApp group, emojis and all, without reading it. At the other, an artist who cannot draw iterates through twenty prompts, rejects fifteen drafts, and finally publishes the one she is happy with. The MeitY rule treats these as the same thing. Both must carry the ten-per-cent label. This is conceptually wrong. The first case is genuine slop. The second is a person using a tool, no different in kind from Photoshop, autotune, or the grammar check that fixed this sentence.

The objection will be that AI is categorically different because it can generate convincing fakes - the Akshay Kumar deepfake, the manipulated political clip. True. But a labelling regime that targets deception is not the regime that has been notified. The current rule targets generation. A scientist using an AI model to clean up a micrograph must label the figure identically to someone deepfaking the Home Minister. Both creators are flattened into the same category. The signal-to-noise ratio of the label will collapse, and within a year users will ignore the watermark the way they ignore cookie consent banners.

A better rule would test for the one touch. Has a named human taken responsibility for this output? Are they answerable for it under existing defamation, copyright, and electoral law? If yes, no label is required; the law already binds the publisher. If no - if the chain of responsibility breaks somewhere between the model and the audience - then label, and label prominently. The architecture is closer to SEBI’s registered analyst framework than to a blanket content-type tag.

I have started using ChatGPT to make cartoons for my Substack. I cannot draw. The cartoons exist because I write the script, review the output, reject what does not work, and publish what I am happy with. Under the current rule, they must carry the label. The cartoonist at this newspaper, using a Wacom tablet and software more sophisticated than the pencil it replaced, does not. The line between us is not the involvement of software. It is the touch. The MeitY rule cannot see that line. It should be redrafted before the labels start appearing on everything and meaning nothing.