A skill is a folder of instructions an AI assistant loads when a task calls for it. A general-purpose model applies US market conventions to Indian data by default, which quietly produces wrong adjusted returns, wrong risk figures and wrong tax treatment. Seven open skills published by Artha carry the Indian conventions instead, and they need no Artha account.
Key takeaways
- A skill is instructions, not data access: it changes how an assistant reasons, not what it can read.
- Indian conventions differ from US defaults in adjustment, settlement, trading-day count and tax treatment.
- The seven cover data literacy, P&L, concentration, risk metrics, fund overlap, filings and taxation.
- They are MIT licensed and public, so the instructions can be read before they are trusted.
- They explain method and state their limits; none of them recommends a security.
What is a skill?
A skill is a folder of written instructions an assistant loads when the work calls for it: how to approach a class of task, what to check, what to refuse to guess at. It is prompt engineering that lives in a file and is shared, rather than being retyped into every conversation.
What a skill is not is access. It reads no account and fetches no data. If an assistant has no way to see your holdings, a skill will not give it one — that is the job of an interface like the one described in What is an MCP server.
What does a general-purpose model get wrong about Indian markets?
The defaults. A model trained mostly on US material assumes US conventions, and most of the resulting errors are silent rather than obvious.
A price series that is not adjusted for a bonus issue shows a crash that never happened, which is the trap described in why do two sources show different market caps territory and in Artha's own data-quality material. Risk figures computed on 252 trading days and a US risk-free rate do not describe an Indian portfolio. Gains are taxed under one of five heads of income here, with holding periods that differ by asset class. None of these produce an error message. They produce a confident, wrong number.
What each skill covers, the install command, and the three rules all seven follow.
What do the seven cover?
They are layered from the data upward. `indian-market-data-literacy` covers reading the bhavcopy NSE publishes after each session and why adjustment matters. Three cover what you own: `portfolio-pnl-review`, `portfolio-concentration-review`, and `portfolio-risk-metrics`, the last on Indian conventions. Two cover what you are buying: `mutual-fund-overlap-review`, for funds that hold the same stocks twice over, and `reading-indian-filings`. One covers what you keep: `indian-investment-taxation`.
They work with or without Artha. Point an assistant at the repository and ask a question about Indian markets; the assistant loads the skill the task calls for.
What will they not do?
They will not recommend. A skill that teaches an assistant to measure concentration teaches it to report the measurement, not to call a portfolio good or bad, and none of the seven produces a buy, sell or hold verdict.
They also state their limits. A skill that computes a risk metric on too short a sample is written to say so rather than to quote a number that looks precise. That is the difference between a method and an answer, and it is the part a general-purpose model is least likely to supply on its own.