Talk 2 Stocks: Your Personal AI Assistant for Market Intelligence

Opening

Most market AI demos start with a promise. Real research starts with a workflow.

A market intelligence assistant has to do more than answer. It has to investigate.

Most market AI demos open with a single line: ask any question about any stock. Real research rarely works that way. A practitioner does not ask one question and stop. They understand the market state. They pick the right data mode. They retrieve evidence. They inspect technical context. They compare sectors. They check global cues. They read a filing or check an event. Then — only then — do they synthesise a view, with appropriate risk framing.

Agent Adda was built around this realisation: the value is not in the chat box. It is in the disciplined loop around it. Talk 2 Stocks, surfaced through nse_agent.py, is that loop made into a working assistant.

This document is a point of view on that pattern, applied to NSE-first market research. Throughout, the research-only frame is explicit. Talk 2 Stocks is a research and learning assistant. It is not a trading recommendation engine. The value is repeatable investigation, not prediction.

Who This Is For

Three audiences, one document

This POV is written to be read by three groups. Each will find a different centre of gravity in the same text.

Executive Summary

Five propositions, in plain language

The thesis of this document, compressed into the smallest number of statements that still carry weight.

The remainder of this document unpacks each proposition through the Talk 2 Stocks system: its terminal surface, its five-layer architecture, its operating principles, its eight named RICs, its prompt library, the artifacts it emits, and a worked anatomy of a single session.

Section 01 · The Problem

Section 01 · The Problem

Market research is fragmented. The friction is everywhere except inside the question.

Indices live one place. Stock-level data another. Sector rotation a third. Filings somewhere else.

Even when the data exists locally, the workflow is repetitive. Open reports. Scan screeners. Compare symbols. Inspect indicators. Write notes. Remember what to check next. Each repetition of this pattern costs time and consistency.

Large language models help with synthesis — they are very good at narrating a clean read of structured inputs — but a chat box alone is not enough. The model does not know which mode of data is appropriate (intraday or end-of-day). It does not know which screeners are relevant. It cannot navigate filings on its own. It does not preserve research artifacts the user can revisit.

What is missing in most assistants

• Tool-using behaviour, not just text generation
• Mode awareness — when to read live data, when to read end-of-day
• Context routing — knowing which tool answers which kind of question
• Reusable recipes for recurring research patterns
• Artifact emission — outputs that survive the chat session
• A discipline of caveats and invalidation, not just confident answers

Talk 2 Stocks combines these missing pieces into one assistant, surfaced through a single terminal.

Section 02 · What It Is

Section 02 · What It Is

A terminal-first AI assistant for NSE-first market research

One window. 150+ commands. A single operating discipline.

Agent Adda is a terminal-first AI market intelligence assistant for NSE-first workflows, with expanding support for US and global market context. It combines natural language questions, slash commands, curated prompts, Recursive Investigative Conversations, technical screeners, intraday scanners, background monitors, watchlist alerts, forensic analysis, options chains, deep document search, and local data tools — all accessible from one interactive terminal: nse_agent.py.

The surface is deliberately broad. Most market intelligence workflows require five to ten different capabilities in sequence. A practitioner opening research on a stock needs a live quote, a technical read, a fundamentals check, a recent-news pass, a peer comparison, and a corporate events lookup — then a synthesised view with caveats. Having each of those capabilities in a different tab, at a different keyboard shortcut, with different data freshness, is exactly the fragmentation the terminal was built to remove.

Twelve Intelligence Surfaces

Each surface earns its place by removing a specific kind of friction. Together they make nse_agent.py a complete operating environment for market research — not a chat wrapper around a single data API.

The Command Surface

150+ slash commands are organised into intelligence domains. The table below shows how the surface is structured — what each domain covers and how much depth it carries. The user does not need to know all 150 commands to begin; natural language handles most queries. The commands are there for practitioners who want precision and reproducibility.

Section 03 · The Research Foundation

Section 03 · The Research Foundation

Weinstein. Minervini. O'Neil. Synthesised.

Talk 2 Stocks is not a collection of indicators. It is a specific investment philosophy, quantified and operationalised.

The sector rotation report and the Stage 2 tracker — the two most-used visual artifacts in the system — are grounded in a body of market research that spans thirty years. Stan Weinstein's 1988 work on market stages, William O'Neil's CANSLIM framework, Mark Minervini's trend templates, and Jegadeesh and Titman's 1993 academic work on momentum (Journal of Finance) form the intellectual core. The system synthesises them into one quantitative pipeline: top-down macro, then sector rotation, then stage analysis, then execution timing.

Three Forces Drive Every Stock Return

Cohen and Polk's research on sector attribution establishes a striking decomposition of stock returns. The implication is direct: most stock selection errors are not stock-level errors at all — they are macro and sector errors committed before a single stock is analysed.

The Weinstein Four-Stage Cycle

Every stock — and every sector index — cycles through four identifiable stages defined by price behaviour relative to the 30-week moving average (approximately the 200-day MA). The cycle exists because institutional accumulation and distribution take time: a fund cannot deploy or redeploy ₹500 crore in a week without moving the price.

The CANSLIM-Weinstein-Minervini Synthesis

The three frameworks are not alternatives — they operate at different layers of the same decision cascade. Weinstein defines which stage to be in. Minervini's trend template defines what quality within Stage 2 looks like (8 criteria: all EMAs aligned, RS ≥ 30-week high, 200-DMA rising, etc.). O'Neil's CANSLIM ensures the fundamental quality of the move — that the earnings and institutional sponsorship are there to sustain it.

Together, they implement: top-down macro → mid-level sector rotation → bottom-up stage analysis → execution timing. Each layer enforces a discipline that compensates for a known behavioural weakness. Order matters: no stock analysis begins until the macro and sector gates are passed.

Section 04 · Architecture

Section 04 · Architecture

Five layers, one disciplined loop

Talk 2 Stocks is easiest to understand as a workflow assistant with five layers — input, orchestration, data, analysis, output.

Each layer earns its place. Input expresses intent. Orchestration routes it. Data answers honestly. Analysis builds evidence. Output preserves the work as artifacts.

The architecture is deliberately practical. The assistant does not try to replace every market tool. It coordinates the repeated steps that usually sit between raw data and a usable research view.

A worked view of how a user sentence becomes an artifact: Input → intent and mode → tool routing → emit. Trace mode preserves the path.

Section 05 · Operating Principles

Section 05 · Operating Principles

Six tenets we hold to

Architecture is downstream of principle. The tenets came first; the layers and the loop followed.

A working assistant is the visible artifact of a set of choices made earlier — about what the system should refuse to do, what evidence it owes the user, and where the line between LLM and tool falls. The six tenets below are the operating discipline of Talk 2 Stocks.

Section 06 · Recursive Investigative Conversations

Section 06 · Recursive Investigative Conversations

Why RIC matters

A one-shot prompt can answer a narrow question. It cannot guide a user through an investigation.

RIC stands for Recursive Investigative Conversations. The point is simple: market research often requires sequence. A RIC starts with a broad question, breaks it into sub-questions, retrieves evidence, compares alternatives, refines the hypothesis, and produces a final view with caveats.

The six-step recursive loop's discipline is not just in the steps; it is in the refusal to skip them.

Eight Named RICs

Each RIC encodes a recurring research pattern. The user does not have to remember the sub-questions or the order — the recipe holds them. Every RIC ends in an artifact and applies caveats and invalidation by default. The eight RICs cover the full arc of practitioner work — from morning session prep to earnings-season management.

Section 07 · Anatomy of a Session

Section 07 · Anatomy of a Session

What a Sherlock RIC actually looks like

An assistant earns trust by showing its work. A trace of one investigation, end to end.

The trace below is a stylised view of a Sherlock RIC running against HDFCBANK after market close. It compresses what the trace mode would show: routing decisions, tool calls, evidence collected, synthesis, and the artifact emitted at the end. The detail is condensed for readability; the shape is faithful.

[USER] > sherlock HDFCBANK
──────────────────────────────────────────────────────
[ROUTING] intent = stock_deep_dive · symbol = HDFCBANK
mode_resolver: /auto → /eod (post-close session window)
session_state: last_symbol = HDFCBANK · last_ric = sherlock
──────────────────────────────────────────────────────
[STEP 01] tool: company_lookup(HDFCBANK)
evidence: profile, latest filings index, recent corp actions
[STEP 02] tool: technical_read(HDFCBANK, mode=eod)
evidence: stage = Stage 2 (3w) · RS = 63 · 50DMA > 200DMA
[STEP 03] tool: peer_compare([ICICIBANK, KOTAKBANK, AXISBANK])
evidence: relative strength table, sector rotation context
[STEP 04] tool: events_calendar(HDFCBANK, window=±21d)
evidence: results date, dividend record, board meeting
[STEP 05] tool: risk_radar(HDFCBANK)
evidence: concentration · event risk · weak-evidence flags
──────────────────────────────────────────────────────
[SYNTH] narrative written · caveats applied · invalidation set
invalidation: weekly close < 1620 ends Stage 2 read
[OUTPUT] reports/sherlock_HDFCBANK_2026-05-09.html (artifact emitted)
[DONE] session preserved · trace inspectable via /trace

Three things in the trace are deliberate. The mode resolver makes its decision visible — the user can always see which data window was read. Every step names the tool and the evidence it produced; the LLM never invents a number. The synthesis ends in an explicit invalidation level and an HTML artifact, not a chat paragraph that disappears with the scroll. The trace is the workflow made legible.

Section 08 · Continuous Surveillance

Section 08 · Continuous Surveillance

The terminal watches while you think

A research assistant that only answers questions is a half-finished product. The durable value is in what it watches for without being asked.

Talk 2 Stocks has two continuous-surveillance surfaces: background monitors and watchlist alerts. Both run independently of the main chat loop. Both queue alerts that render in the terminal every few seconds during market hours. Together they convert the assistant from a query tool into a persistent research presence.

Background Monitors — 14 Strategies

The /monitor command launches auto-running scan workers that sweep an index at configurable intervals and queue alerts for anything that matches a strategy. The practitioner can run one strategy or all fourteen in parallel. Alerts display direction (BUY / SELL / WATCH), entry, target, stop-loss, risk-reward ratio, and a confidence level.

Usage example: /monitor start breakout NIFTY 500 15 buy → runs every 15 min on NIFTY 500, BUY signals only. /monitor start all 5 → all 14 strategies, 5-minute sweep.

Watchlist Alerts

The /alert command manages manual price and RSI triggers on specific symbols. Input is natural language: "RELIANCE breakout above 1580 in 15min" or "TCS RSI below 30". An LLM parser converts this to a structured alert; a positional fallback handles cases where parsing fails. The background alert monitor polls every five minutes during market hours and surfaces any triggered conditions into the active session.

Section 09 · Intelligence Library

Section 09 · Intelligence Library

Sixty prompts. Eight RICs. One teachable surface.

The library is the interface between the practitioner and the command surface. It makes 150+ commands discoverable without memorisation.

A prompt library is not a convenience feature. It is a product design choice with a specific purpose: making the surface teachable. A new practitioner opening the terminal for the first time faces 150+ slash commands. The library gives them a starting point — sixty named research workflows, organised into ten categories, runnable with a single keystroke. Each prompt teaches by example: what a good market-intelligence question looks like, how to frame a sector comparison, what evidence a stock deep dive should gather.

Ten Categories, Sixty Prompts

The prompt library is organized into ten research domains. Each prompt is a complete, opinionated research question — not a template to fill in, but a live workflow to run.

The prompt library is browsable with /prompts and filterable by category. Any entry can be run with a single-keystroke shortcut (p1 through p60). The library is not a catalogue of features — it is a curated set of research starting points, each reflecting a real practitioner workflow.

Section 10 · Evidence Surfaces

Section 10 · Evidence Surfaces

An artifact you can revisit

A chat answer disappears with the scroll. An artifact lives long enough to be checked, audited and shared — and the most important ones are grounded in thirty years of investment research.

The report generation surface covers eight report types across three output formats (HTML, PDF, Markdown). Two of those reports deserve particular attention because they implement the CANSLIM-Weinstein-Minervini synthesis end-to-end: the Sector Rotation Report and the Stage 2 Tracker.

Two Reports, Two Time Horizons

The two reports work at different cadences and different levels of abstraction. Neither is useful without the other: the sector report tells you where to look; the tracker tells you when to act.

Ranked by Rotation Score (0.35×RS_1M + 0.25×Return_1M + 0.20×RS_5D + 0.10×RS_3M + 0.10×RS_6M). Top sectors on 9 May 2026: Capital Markets (14.8), Defence & Aerospace (14.1), Energy/Power (13.0).

Confirmed Stage 2 names ranked by Investment Score (0.38×Technical + 0.27×RS + 0.25×Fundamental + pattern and Supertrend bonuses). Every row carries an explicit invalidation level.

The Five-Step Selection Funnel

The sector rotation report is not a screener that filters once. It is a five-step cascade that starts with the entire NSE universe and narrows to a conviction-weighted portfolio of approximately ten positions. Each gate is necessary; skipping any one produces a materially worse candidate set.

1. NSE Universe (~2,000) — load_comprehensive_analysis() — all NSE stocks with prices, technical scores and fundamental metrics from the daily refresh pipeline.

2. Hot Sector Filter (Top 6) — rank_rotating_sectors().head(6) — Rotation Score = 0.35×RS_1M + 0.25×Return_1M + 0.20×RS_5D + 0.10×RS_3M + 0.10×RS_6M. Only stocks inside the top 6 ranked sectors advance.

3. Stage 2 Hard Gate (Stage 2 only) — enrich_with_stage(): price > SMA50 > SMA200, both moving averages rising. Stage 1, 3 and 4 stocks are dropped. The gate protects against ~70% of historical drawdowns.

4. Investment Score Rank (Top 8/sector) — 0.38×Technical + 0.27×RS + 0.25×Fundamental + bonuses for pattern, Supertrend, trend signal and stage quality. Top 8 per sector advance.

5. Final Portfolio (~10 positions) — compute_position_sizes() — conviction-weighted, capped at 15% per position, normalised to 100%. Action buckets: BUY_WATCH, HOLD_TRAIL, BREAKOUT_WATCH, WAIT_FOR_PULLBACK, AVOID.

Why the Step-by-Step Approach Raises the Win-Rate

Each gate raises the base probability of a successful outcome independently. Multiplied together, they produce a candidate with 4–5× the base rate of a random NSE pick.

Worked Example: AEROFLEX, 9 May 2026

Every gate applied to AEROFLEX on 9 May 2026 — from live report output, not simulation. Every number is from the actual pipeline run. The trade plan is mechanical and reproducible.

Two Forensic Surfaces the Market Misses

Beyond technical and fundamental surfaces, Talk 2 Stocks embeds forensic accounting intelligence made accessible in a single command.

CANSLIM Quality Scoring

The /canslim command runs O'Neil's 7-point quality framework. STRONG BUY ≥ 6/7; AVOID < 4. A quality gate that runs before a technical setup is considered.

• C — Current Earnings — EPS growth ≥ 25% YoY, revenue confirmation. Quarterly.
• A — Annual Growth — EPS ≥ 25% for 3+ years, ROE ≥ 17%, stable margins.
• N — New Catalyst — New product, management change, or near 52-week high.
• S — Supply & Demand — Float, volume pattern. Smaller float with demand surge scores higher.
• L — Leader — RS rank, sector outperformance, Stage 2 status. Laggards do not qualify.
• I — Institutional — FII / DII / MF stakes increasing, bulk deals from quality institutions.
• M — Market Direction — Bull, bear or correction regime. Follow-through day required.

Section 11 · How It Saves Time

Section 11 · How It Saves Time

Time saved is consistency gained

The point is not speed. The point is repeatability.

Talk 2 Stocks saves time by reducing repeated navigation and synthesis work. It brings commands, reports, scans, prompts and research flows into one terminal. It turns recurring analyst patterns into repeatable recipes. It generates structured summaries instead of leaving the user with raw tables. It preserves report outputs that can be reviewed later. It supports a faster move from question to evidence-backed view.

But the more durable value is consistency. A repeatable workflow is easier to audit, improve and teach than an improvised series of browser tabs and ad hoc notes. When the same investigation always runs the same way, two things follow: results become comparable across days, and gaps in the workflow become visible — they can be fixed instead of being absorbed by the analyst.

Section 12 · The Builder's Takeaway

Section 12 · The Builder's Takeaway

The chat box is the smallest part

For builders, this is the lesson: the durable value in applied AI agents is the workflow around the conversation, not the conversation itself.

For market intelligence, that means an assistant that can route intent, choose data modes, run screeners, read reports, produce artifacts, and keep the user inside a disciplined research loop. Talk 2 Stocks is a working example of that product pattern.

The same pattern generalises. A workflow assistant for life sciences research, manufacturing operations or regulatory intelligence has the same anatomy: input layer expressing intent in many forms, an orchestration layer that routes, a data layer that fails honestly, an analysis layer where tools build evidence and a model writes synthesis, and an output layer that emits artifacts.

Hold. Think. Then act.

That editorial line is the operating discipline of Talk 2 Stocks. The system holds — by reading mode, state and intent before answering. It thinks — by decomposing investigations into sub-questions and gathering evidence. It acts — by emitting an artifact the user can use. The order is the product.

Appendix A · Glossary

Terms used in this document

A reference for non-finance readers and for builders new to the NSE-first workflow and its analytical frameworks.

RIC — Recursive Investigative Conversation. A multi-step research recipe that decomposes a question into sub-questions, gathers evidence tool by tool, refines a hypothesis, and emits a structured view with an invalidation level and a persistent artifact.

Stage 2 — Weinstein-style trend stage: confirmed advance after a base. Typically requires price above the 30-week moving average with the average rising and relative strength confirming. Stage 1 = base, Stage 3 = distribution, Stage 4 = decline.

Breadth — Share of stocks participating in a market move: advancers vs decliners, percent above 50-DMA, new highs vs new lows. A move on weak breadth is treated as fragile; broad breadth supports a thesis.

RS — Relative Strength. A symbol's price performance versus a benchmark index or peer set, shown as a 0–100 percentile rank. Distinct from RSI (Relative Strength Index, a momentum oscillator bounded 0–100 by formula).

Regime — Overall market state — risk-on, risk-off, neutral — inferred from indices, breadth, volatility and global cues. Used as session-opening context; influences whether the assistant frames answers defensively or constructively.

Invalidation — The price level or condition at which a thesis is proved wrong. Talk 2 Stocks does not present a setup, stage call or investigation conclusion without one. A read without an invalidation is a guess.

DMA — Daily Moving Average. Common reference levels: 20, 50, 100, 200 DMA. The 50/200 crossover and the slope of the 200 DMA are standard trend filters.

/live /eod /auto — Data mode controls. /live prioritises intraday real-time NSE API data. /eod uses end-of-day historical CSV and SQLite snapshots. /auto routes by session window — intraday when the market is open, historical otherwise.

Artifact — A persistent output (HTML, PDF, Markdown) generated by a session, independent of the chat scroll. Designed to be re-read, audited, shared, and compared against future sessions.

VCP — Volatility Contraction Pattern. A technical setup where price and volume contract in a series of tighter and tighter ranges before a breakout. Associated with Stan Weinstein and Mark Minervini's work.

ORB — Opening Range Breakout. An intraday strategy that identifies the high and low of the first 15–30 minutes of trading and trades breakouts from that range on volume confirmation.

CANSLIM — William O'Neil's 7-point growth quality framework. C = Current quarterly earnings, A = Annual growth, N = New product/high, S = Supply & demand, L = Leader, I = Institutional sponsorship, M = Market direction. Scored 0–7.

Beneish M-Score — Eight-variable probit model for earnings manipulation risk. A score above −1.78 flags potential manipulation. Uses financial statement ratios including Days Sales in Receivables Index, Gross Margin Index, and Accruals.

Piotroski F-Score — Nine-signal financial health model scored 0–9. Signals cover profitability (4), leverage/liquidity (3) and efficiency (2). Score 7–9 = strong health; 0–3 = weak. Used as a quality filter, not a timing tool.

Altman Z'-Score — Emerging-market distress and bankruptcy risk model. Z' below 1.1 = distress zone; above 2.6 = safe zone. The prime variant (Z') is calibrated for non-US markets, making it directly appropriate for NSE-listed companies.

PCR — Put-Call Ratio. Total put open interest divided by total call open interest for an options chain. PCR above 1.2 is considered bullish (excess put hedging); PCR below 0.7 is considered bearish (complacency).

Max Pain — The options strike price at which the maximum number of open contracts (puts and calls combined) would expire worthless. Theoretically, market makers are incentivised to move price toward this level near expiry.

Supertrend — A trend-following indicator derived from ATR (Average True Range) and a multiplier, plotted above or below price. A buy signal fires when price closes above the Supertrend line; a sell signal fires when price closes below it.

nse_agent.py — The terminal entry point for Talk 2 Stocks. The single interactive window that exposes all 150+ commands: chat, slash commands, prompt library, RICs, scanners, monitors, reports, options, and forensic tools.

Mode resolver — The orchestration component that decides which data mode (/live, /eod) applies to a query. Reports its decision in trace mode so the user can always see which data window was read.

Trace mode — An audit view activated via /trace showing the routing decision, tools called, evidence cited and artifact emitted for any answer. Makes the investigation auditable and reproducible.

ADR — American Depositary Receipt. A US-listed instrument representing shares of a non-US company. Overnight ADR moves on Indian companies (INFY, WIT, HDB, etc.) are inputs to the India read-through in the global context report.

Pivot — A reference price level for breakouts and breakdowns — typically a prior swing high or a base boundary. Paired with an invalidation level in every setup read produced by Talk 2 Stocks.

About Agent Adda

Agent Adda is a practitioner-first platform that publishes points of view, working systems and design language for applied AI engineering. Its editorial line — Hold. Think. Then act. — is also the operating discipline of the products it ships.

Talk 2 Stocks, the system this document describes, is one of those products. It is the Agent Adda thesis applied to NSE-first market research: a terminal-first assistant whose value lives in the workflow around the conversation, not in the conversation itself.

Editorial line — Hold. Think. Then act. The order is the product.

Position — Research and learning publication. Not a recommendation engine. Not investment advice.

Series — Practitioner POVs on architecture, RAG, agentic systems, governance and applied workflows.

Disclaimer — Agent Adda and Talk 2 Stocks are for research, education and workflow exploration only. They are not investment advice, trading recommendations, financial planning, or a substitute for independent judgement, risk management or regulatory obligations. Market data can be delayed, incomplete, stale or incorrect. Users should verify evidence independently before making financial decisions.

Download

This article is available as a full PDF for offline reading and sharing.

Download PDF →