Build Your Own Stock Analyst With ChatGPT and Instinct

1. Key Themes

Theme 1: AI collapses the institutional research cost stack by ~99.8%

The article's central claim is that the infrastructure behind a hedge fund analyst can be replicated for a tiny fraction of the cost.

The institutional baseline is expensive and built on learnable frameworks

• "Bloomberg Terminal: about $28,000 a year / Sell-side research: $600 to $1,500 a month / Expert network calls: around $500 an hour / A junior analyst for the data work: $150K fully loaded"
• "Roughly $250K a year per analyst, built on frameworks anyone can learn from a CFA study guide."
• The workflow burden is also heavy: "A buy-side analyst at a Manhattan hedge fund covers 15 companies, and every Monday she spends about two hours per stock... That's 30 hours a week before she starts thinking."

The replacement stack costs ~$35/month

• "ChatGPT Plus ($20/month) runs GPT-6 Astra inside ChatGPT Work and Codex, a data connector adds about $15, and Instinct is free during its invite-only beta. About $35 a month in total."
• Exhibit 1 title: "Running the research workflow on ChatGPT and Instinct costs about $420 a year, roughly 0.2% of a hedge fund's cost to equip one analyst." The cost breakdown: junior analyst ~$150K, expert network ~$60K, Bloomberg $28K, sell-side ~$13K versus ChatGPT Plus 0.24K, Financial Modeling Prep ~0.18K, Instinct $0 in beta.
• Caveat in the exhibit footnote: "the Pro plan raises the total to about $1.4K. Instinct is in an invite-only beta at $0, and its pricing can change."

Theme 2: Split the analyst job into "analysis" (LLM) and "watching" (agent)

The system divides labor between a reasoning model and a persistent monitoring agent, rather than expecting one tool to do both.

Analysis and monitoring are separate jobs with separate tools

• "ChatGPT does the analysis, Instinct does the watching, and every decision stays with you."
• Instinct "can now take over the other half of an analyst's week: tracking earnings dates, filings, prices and deadlines, then texting or calling you when something needs attention."
• Exhibit 2 maps ownership: ChatGPT owns screening, valuation (three-scenario DCF), thesis reviews and position size recommendations; Instinct owns earnings calendar/IR filings and price levels/deadlines; Financial Modeling Prep owns prices, statements and estimates.

Proactive, push-based alerts are the differentiator

• Exhibit sample from Instinct: "Morning. ACME reports after the close on Wednesday, with consensus at $1.12 a share on $4.8B of revenue. I'll call you at 8am that day unless you'd rather I text."
• "Six standing jobs for Instinct, written as the exact texts you send it."

Theme 3: Human-in-the-loop guardrails are built into the architecture

The system deliberately keeps execution authority and brokerage access away from AI.

Buy/sell decisions and brokerage access stay human-only

• Exhibit 2 title: "Each tool owns a distinct part of the analyst's job, while buy and sell decisions and brokerage access stay with the investor."
• "Instinct is kept out of brokerage accounts and set to $0 spending for investing tasks. ChatGPT has no trading access in this setup."
• "Six standing jobs for Instinct... with rules that keep it away from your brokerage account."

The agent needs explicit boundaries because it holds your logins

• The package includes "the boundaries for an agent that holds your logins."

Theme 4: Prompting best practice is shifting toward goal-based, shorter prompts

Model upgrades are invalidating older prompt-engineering habits.

Long step-by-step prompts now degrade output on GPT-6 Astra

• "OpenAI's own guidance says it works best with fewer instructions, so the long step-by-step prompts from that era now make it over-check and over-explain."
• New format: "All 12 rewritten for Astra, each stating a goal, a deliverable and a done condition. They're shorter than the old versions and the output is sharper."
• Shared rules are centralized: "The analyst contract: one block you paste into a ChatGPT Project so all 12 prompts share the same rules on sources, verification and when to ask you."

2. Contrarian Perspectives

Institutional research is a framework problem, not an information moat

The article challenges the idea that Wall Street's edge requires expensive infrastructure. "Roughly $250K a year per analyst, built on frameworks anyone can learn from a CFA study guide." Evidence: the stack's cost is about 0.2% of the institutional figure, and the analyst's week is dominated by mechanical work ("pulling earnings, updating models and skimming sell-side notes... 30 hours a week before she starts thinking"), which is precisely what automation absorbs. Caveat: the piece is promotional, and what is demonstrated is workflow replication, not performance parity.

Better models mean less prompting, not more

Against the common practice of piling on detailed instructions: "OpenAI's own guidance says it works best with fewer instructions... the long step-by-step prompts from that era now make it over-check and over-explain." The authors rebuilt their own prior top-performing premium piece from May for this reason, implying prompt libraries decay quickly with each model generation.

3. Companies Identified

OpenAI / ChatGPT (GPT-6 Astra)

• Description: Maker of ChatGPT; GPT-6 Astra is its latest model.
• Why mentioned: The "brain" of the stack, handling screening, DCF valuation, thesis reviews and sizing.
• Quotes: "GPT-6 Astra launched on September 3, and OpenAI's own guidance says it works best with fewer instructions." / "ChatGPT Plus ($20/month) runs GPT-6 Astra inside ChatGPT Work and Codex."

Instinct

• Description: Personal AI agent (invite-only beta) created by Noah Shinn.
• Why mentioned: The monitoring layer that tracks holdings and contacts the user by text or call.
• Quotes: "Instinct, Noah Shinn's personal agent, can now take over the other half of an analyst's week: tracking earnings dates, filings, prices and deadlines, then texting or calling you when something needs attention." / "Instinct is free during its invite-only beta."

Financial Modeling Prep

• Description: Financial data provider/API.
• Why mentioned: The data connector (~$15/month) supplying prices, statements and estimates.
• Quotes: Exhibit 2 lists it as owning "Prices, statements and estimates"; Exhibit 1 shows "Financial Modeling Prep: ~0.18" (thousand per year).

Anthropic Claude

• Description: AI model used for the original May version.
• Why mentioned: The prior iteration of this workflow.
• Quotes: "In May we rebuilt that workflow with Claude, and it became one of our best-performing premium pieces."

Bloomberg

• Description: Maker of the Bloomberg Terminal.
• Why mentioned: Benchmark for the incumbent cost structure being disrupted.
• Quotes: "Bloomberg Terminal: about $28,000 a year."

4. People Identified

Noah Shinn

• Description: Creator of the Instinct personal agent.
• Why mentioned: His product powers the monitoring and alerting layer.
• Quotes: "Instinct, Noah Shinn's personal agent, can now take over the other half of an analyst's week."

Ruben Dominguez

• Description: Author of the post at The AI Corner.
• Why mentioned: Byline; builder of the workflow.
• Quotes: "In May we rebuilt that workflow with Claude, and it became one of our best-performing premium pieces."

5. Operating Insights

Write prompts as goal, deliverable, done condition

• Replace long step-by-step instructions with three elements. "Each stating a goal, a deliverable and a done condition. They're shorter than the old versions and the output is sharper."

Centralize shared rules in one "contract" inside a Project

• Paste a single block so every prompt inherits identical standards. "The analyst contract: one block you paste into a ChatGPT Project so all 12 prompts share the same rules on sources, verification and when to ask you."

Constrain agents with hard permission limits

• Give monitoring agents read/alert authority only. "Instinct is kept out of brokerage accounts and set to $0 spending for investing tasks." Also run a fixed cadence: "the weekly schedule, the 25-minute due diligence chain, the live-data setup."

6. Overlooked Insights

Prompt libraries have a short half-life

The piece was rebuilt only months after the original, driven by a model launch on September 3. For anyone selling or operating prompt-based workflows, this implies recurring maintenance costs and a recurring content/product opportunity with each model release.

The agent's value is attention allocation, not intelligence

The sample Instinct message offers a choice of channel ("I'll call you at 8am that day unless you'd rather I text") and ties to specific catalysts such as consensus EPS ($1.12) and revenue ($4.8B). The edge for retail investors may be less about better analysis than never missing a catalyst, a task institutions staff with analysts.