Your First 30 Days — A Practical Plan That Doesn’t Break Your Process

Start with the honest boundary, because everything else depends on it. At the end of thirty days you will not have a system that tells you what to buy, a model that anticipates NIFTY, or any statistical evidence that your decisions improved. Thirty days is far too short a window to say anything about outcomes even if the process were perfect, and the process will not be perfect. Anyone selling you a month-long path to better returns is selling the month, not the returns.

What a month is long enough for is habit formation and elimination. You can run the same prompt on a dozen real documents, discover which parts of the output you actually read and which you skip, and freeze a template around that. You can develop the reflex of opening the source PDF before writing a number into a note. You can find out, concretely, that a particular task you assumed AI would help with does not help at all — and stop, which is a real result even though it looks like a failure.

The distinction that runs through every article in this module applies here too. AI compresses, structures and drafts. It does not forecast prices and it does not know what happened after its training cutoff. A thirty-day plan built on the first set of capabilities is buildable. One built on the second is a plan to be disappointed on a schedule.

One more boundary before the plan. Nothing in this month changes who is accountable. Every position, every decision and every consequence stays with you, exactly as it did before you opened a chat box. A model produces text; it does not carry responsibility and there is no arrangement under which it can. That is the subject of the guardrails article at /learn/ai/guardrails-sebi-and-the-human-in-the-loop, and it is scheduled inside week one for a reason.

A month of experimentation with an unfamiliar tool needs boundaries written down in advance, when you are calm, rather than improvised later when an answer looks unusually good. There are four, and none of them is negotiable during the month.

First, your existing process keeps running untouched. Whatever you did before — your charts, your scans, your reading, your notes — continues exactly as it did. The AI pass sits alongside it as a second opinion you can throw away. Replacing a step you trust with an unproven one in week two is the single fastest way to turn a learning exercise into an expensive one.

Second, nothing enters a prompt that should not leave your machine. No broker credentials, no client information, no live position sizes, no unpublished research belonging to anybody else. Read /learn/ai/data-privacy-and-what-never-to-paste before day one, not after an incident, because a prompt box looks private and is not.

Third, no number reaches a note or a decision without being checked against the primary source it came from — the exchange filing, the annual report page, the transcript paragraph, your own chart. This is the habit the whole month is really about, and it has to be in place from day one. Bolting verification on afterwards never works, because by then you have spent three weeks learning that the output is usually fine.

Fourth, keep a running log of what you tried. Two lines per session: what you asked for, and whether the output was worth the time. At the end of the month this log is the evidence you review, and without it the review collapses into a general impression, which is exactly the thing that cannot be argued with or acted on.

The four weeks map onto the five groups of this curriculum, in order. Foundations and the safety rules first, because knowing what the machine is stops you asking it for things it cannot do. Then research craft, which is where nearly all the durable value in this module sits. Then one — only one — application from the market group. Then the workflow layer and the review.

Two groups sit deliberately outside the month. The Quant and Machine Learning group covers a genuinely different technology with a much higher bar for doing it responsibly, and squeezing it into week four would produce exactly the overconfident backtest those articles exist to warn you about. The agents and automation material is out for a related reason: nothing in a first month should be taking actions on your behalf. Both are mapped in the section on what thirty days does not get you.

The table below is the plan in one view. Everything after it is detail. If you read nothing else on this page, the table plus the month-end review checklist is enough to run the whole thing.

Week one produces no research output at all, and that is intentional. You are building the mental model that determines whether every later week is useful or dangerous. The single most expensive error people make with these tools is category error: asking a text-completion machine for tomorrow’s level, asking an image describer for an exact price, expecting a fixed rule to notice that the tape has changed character.

Read the four foundation articles in order. /learn/ai/what-ai-is-in-trading separates the four unrelated technologies that share the label. /learn/ai/how-llms-work-for-traders explains next-token prediction, training cutoffs and why a confident wrong answer is the normal behaviour of the machine rather than a malfunction. /learn/ai/ai-capability-line draws the boundary you will keep referring back to. /learn/ai/choosing-your-ai-stack helps you decide which kind of model suits the job you have in mind.

Then read the two safety articles out of curriculum order, because you need them before you paste anything. /learn/ai/data-privacy-and-what-never-to-paste tells you where your text goes. /learn/ai/guardrails-sebi-and-the-human-in-the-loop is the accountability frame: investment advice in India is a regulated activity carried out by a registered, named, accountable person, and no amount of fluent output changes that. Whatever you build this month is research support for your own decisions, and it stops being defensible the moment it is presented to anyone else as a recommendation.

The one hands-on exercise this week is diagnostic, not productive. Take a company whose latest results you have already read carefully, and ask a model about them with no document pasted. Compare its answer to what you know. You are looking for the specific texture of an invented figure — right format, right decimal places, confident tone, wrong number, no signal that anything is missing. Seeing that once, on material where you know the truth, is worth more than reading about it five times.

On model limits, check the current published constraints of whatever tool you use rather than trusting anything you read about capacity, pricing or version behaviour — including here. Those numbers change frequently and quietly, which is precisely the failure mode /learn/ai/model-drift-and-silent-failure is about.

Week two is where the durable skill lives. Everything else in this module is application; this is the craft. You will read /learn/ai/prompting-for-market-research and /learn/ai/filings-and-concalls-with-ai, and then run one prompt structure repeatedly on real Indian market documents until it stops surprising you.

Pick your source material from primary sources only: the exchange filing on the NSE or BSE website, the annual report from the company’s own investor-relations page, the concall transcript as published. Not a news summary of a filing, and not a forum post about it. The whole design of the prompt is that the model reasons over the text you supply, so the text you supply has to be the real thing.

The prompt structure has four parts and does not change: the exact scope of what you want, the pasted source, the fixed output format, and explicit permission to answer "not stated in the source". That last part is what people leave out and it is what makes the difference. A model asked for six risks will return six risks whether or not the document contains six, and the invented ones are phrased identically to the real ones.

Verification is part of the same session, not a later chore. Read /learn/ai/hallucination-and-the-audit-trail during this week and adopt the habit immediately: every figure gets opened in the source, found in a sentence, and written into your notes with the document and page or section beside it. If it takes you thirty seconds a claim, that is the correct cost, and it is the cost that makes the note worth keeping.

Also read /learn/ai/context-cost-and-limits this week, because it explains why a 300-page annual report cannot be dealt with in one prompt and why your fifteenth question in a long thread quietly comes back worse than your third. Split the document. One section, one prompt, one fresh thread.

Week three is where most people overreach. The temptation, having got a filing summary that worked, is to add chart reading and news triage and a screening layer and a journal review all in the same week. Do not. Add one. If you add five, you will not know which of them helped, and at the end of the month you will have five half-formed habits instead of one you actually kept.

Choose the one that matches the bottleneck in your existing process, not the one that sounds most interesting. If you are drowning in headlines, take news triage. If you have a long watchlist and not enough hours, take chart triage or screening. If your problem is repeating the same mistakes, take the journal. If you suspect your positions are less diversified than the number of names suggests, take portfolio structure.

Whichever you choose, read its article properly first, because each one has a specific failure mode that is not obvious from outside. /learn/ai/ai-for-charts-and-patterns explains that a vision model reads pixels, not price — so every level it names has to be confirmed on your own chart before it enters a note. /learn/ai/ai-assisted-screening keeps numeric filtering where it belongs, in a real scanner, and uses AI only for the language layer. /learn/ai/news-and-sentiment-at-machine-speed makes the case that triage and deduplication are the value, and that a sentiment score is a compression of language, never a judgement about whether an event matters. /learn/ai/ai-trading-journal-and-review is about structuring a review so the model argues with you instead of agreeing, which it will do by default. /learn/ai/ai-for-portfolio-review surfaces concentration and overlap, with the numbers supplied by you and never recalled by the model.

There is a sixth option for readers who write their own scans or indicators: /learn/ai/ai-assisted-coding-for-traders. Take it only if you already read code, because the whole point of that article is that generated code which looks correct and quietly repaints is worse than no code at all. If you cannot audit what comes back, this is not the week-three job for you.

Whatever you pick, run it on situations where being wrong is free. Closed trades for the journal. A watchlist name you do not hold for chart triage. Last week’s news rather than this morning’s. You are testing whether the tool helps, and a test you have money riding on is not a test.

Three weeks in you have a template, some verified notes and one applied job with a verdict attached. Week four turns that from a set of things you did into a thing that runs. Read /learn/ai/building-an-ai-research-workflow — it is the structural companion to this article and the closest thing the module has to a blueprint.

The workflow is three assets and no software: a template library holding one fixed prompt per recurring task, a sources folder you paste from, and a notes file of verified, dated claims that becomes the grounding for every future session on the same company. The chat conversation is deliberately disposable. The test is simple — could you close every open AI conversation right now and lose nothing? If the answer is no, the extraction step has not been done.

Then read /learn/ai/model-drift-and-silent-failure, because the thing you have just built will degrade without telling you. Providers update models. Data feeds change format. Market regimes shift under a template written for the previous one. None of that produces an error message; the output simply gets worse while looking identical. The defence is a fixed reference document you re-run every month or two and compare against a stored answer you already verified.

Finally, read /learn/ai/evaluating-an-ai-trading-tool before you spend a rupee on anything. Having spent a month doing this by hand, you are in an unusually good position to interrogate a vendor: you know what the tasks actually cost in effort, you know which of the four technologies is plausibly underneath the marketing, and you know that an accuracy percentage published without a sample size, a period and a cost assumption is a number with no content. Check the SEBI registration status of anyone offering anything that resembles advice, and treat performance claims without disclosed basis as a reason to walk away rather than a reason to negotiate.

One thing week four does not include is automation. Nothing you have built should be taking an action on its own behalf at the end of a first month. /learn/ai/ai-agents-and-automation-limits explains why reversibility is the right test, and reading it now is useful precisely so that you know what you are choosing not to do.

This section matters more than the plan, because a month of small successes is the exact condition under which people start believing things that are not true. Four things sit outside this month, and three of them sit outside the tool entirely.

It does not get you prediction. Nothing in the four weeks above, and nothing in the twenty-three articles before it, produces a forecast of where a price goes. A language model completes text. A vision model describes pictures. A fitted numeric model expresses a statistical tilt learned from history, which is a different object from a forecast and a much weaker one. If your month ends with you asking a chat box where NIFTY closes tomorrow, the plan failed regardless of how good the filing summaries were.

It does not get you a quantitative modelling capability. The Quant and Machine Learning group — /learn/ai/machine-learning-vs-llms, /learn/ai/features-labels-and-overfitting, /learn/ai/backtesting-an-ai-strategy and /learn/ai/model-drift-and-silent-failure — describes work that takes months to do responsibly, and whose most common outcome is discovering that an attractive backtest was look-ahead bias, survivorship bias or memorised noise. Read those articles by all means. Do not attempt to compress the practice into week four.

It does not get you automation. An agent that takes actions is a different risk category from a model that reads, because reading is reversible and an order is not. That subject has its own article and its own prerequisites, including kill switches that have to exist before anything is switched on.

And it does not get you evidence about your own results. Thirty days is far too few decisions to distinguish a better process from a favourable stretch of tape, in either direction. The honest measurement horizon for a change in process is long, and anyone who tells you a month settled the question is describing a coincidence with confidence.

What remains after all those subtractions is still worth the month: a template you can run in seconds, a growing file of dated and sourced notes, a verification reflex, and a clearer sense of which parts of your process a machine can genuinely take off your hands. That is a small claim, and it is one that holds up.

On day thirty you sit down with the experiment log and go through it line by line. The review has one bias built into it deliberately: the default is to drop things. Anything that did not clearly earn its place goes, because an unused half-habit costs attention every time you remember it exists and returns nothing.

Ask three questions of each thing you tried. Did it save time you can actually name — a specific number of minutes on a specific task? Did anything it produced survive being checked against the source, consistently? And would you notice if it disappeared tomorrow? A tool that fails the third question was never part of your process, whatever the log says.

Then look at the notes file itself, because that is the real output of the month. Count the verified claims in it. Ten checked, dated, sourced lines about companies you follow is a genuinely good month. Two hundred lines of unverified model output is a worse position than you started in, because it looks like knowledge and is not.

Finally, write the month up in a paragraph and date it. What you kept, what you dropped and why, and what you would do differently. Next quarter, when you are tempted to re-adopt something you already tested, that paragraph is what stops you spending the same three weeks twice.

The module opened by taking one marketing word apart into four unrelated technologies, on the argument that you cannot calibrate trust in a machine you cannot name. Everything since has been the working-out of that: what each one does, where each one fails, and what has to be true before any of it touches a decision. The month you have just run is that argument turned into habits.

What you own now is unglamorous and durable. A frozen prompt template. A folder of primary-source documents. A short file of dated, sourced, personally checked claims that will be more useful next quarter than it is today, because the comparison against last quarter is where its value actually appears. A reflex that opens the source before writing a number. And a list of things you tested and discarded, which is the part nobody advertises and the part that keeps a process clean.

What you do not own is any advantage in the market, and it is worth ending on that rather than softening it. The tools in this module remove drudgery from the reading half of research. They do not tell you what to buy, they do not know what happens next, and they never carry the consequence of a decision. That still sits entirely with you — which is also the reason the verification habit was built first and the reason it has to outlast the enthusiasm.

If you want a next step, it is not another tool. It is the second pass on the same company: pull up the notes you verified this month, put them alongside the next quarter’s filing, and ask what changed, what went quiet and what is new. That single exercise is what the whole workflow exists to make possible, and it is the point at which a month of setup starts paying you back in a form you can see.