๐๐จ๐ฆ๐ฉ๐๐ง๐ฒ ๐๐๐๐ญ๐ข๐ง๐ ๐๐ซ๐๐ฉ. ๐-๐๐ฉ๐ฆ
One of the most intensive tasks - prep for a meeting with company management. This works for both the sell side (covered company) or the buyside (portfolio company).
The prompt was just this. โ๐ ๐ฐ๐ถ ๐ข๐ณ๐ฆ ๐ฉ๐ฆ๐ญ๐ฑ๐ช๐ฏ๐จ ๐ข๐ฏ ๐ฆ๐ฒ๐ถ๐ช๐ต๐บ ๐ณ๐ฆ๐ด๐ฆ๐ข๐ณ๐ค๐ฉ ๐ข๐ฏ๐ข๐ญ๐บ๐ด๐ต ๐ฑ๐ณ๐ฆ๐ฑ๐ข๐ณ๐ฆ ๐ง๐ฐ๐ณ ๐ข ๐ฎ๐ฆ๐ฆ๐ต๐ช๐ฏ๐จ ๐ธ๐ช๐ต๐ฉ ๐ต๐ฉ๐ฆ ๐๐๐ ๐ข๐ฏ๐ฅ ๐๐๐ ๐ฐ๐ง ๐๐๐. ๐ค๐ณ๐ฆ๐ข๐ต๐ฆ ๐ข ๐ฃ๐ณ๐ช๐ฆ๐ง๐ช๐ฏ๐จ ๐ฏ๐ฐ๐ต๐ฆ ๐ง๐ฐ๐ณ ๐ต๐ฉ๐ฆ ๐ฎ๐ฆ๐ฆ๐ต๐ช๐ฏ๐จ, ๐ถ๐ด๐ช๐ฏ๐จ ๐ฅ๐ข๐ต๐ข ๐ง๐ณ๐ฐ๐ฎ ๐ต๐ณ๐ข๐ฏ๐ด๐ค๐ณ๐ช๐ฑ๐ต๐ด, ๐ฏ๐ฆ๐ธ๐ด, ๐ง๐ช๐ญ๐ช๐ฏ๐จ๐ด ๐ข๐ฏ๐ฅ ๐ณ๐ฆ๐ค๐ฆ๐ฏ๐ต ๐ฑ๐ณ๐ช๐ค๐ฆ ๐ฉ๐ช๐ด๐ต๐ฐ๐ณ๐บ. ๐๐ฏ๐ค๐ญ๐ถ๐ฅ๐ฆ ๐ง๐ช๐ฏ๐ข๐ฏ๐ค๐ช๐ข๐ญ๐ด ๐ข๐ฏ๐ฅ ๐ณ๐ข๐ต๐ช๐ฐ๐ด. ๐ช๐ฅ๐ฆ๐ฏ๐ต๐ช๐ง๐บ ๐ด๐ฐ๐ฎ๐ฆ ๐ต๐ฐ๐ถ๐จ๐ฉ ๐ฒ๐ถ๐ฆ๐ด๐ต๐ช๐ฐ๐ฏ๐ด ๐ง๐ณ๐ฐ๐ฎ ๐ต๐ฉ๐ฆ ๐ญ๐ข๐ด๐ต 2 ๐ฆ๐ข๐ณ๐ฏ๐ช๐ฏ๐จ๐ด ๐ค๐ข๐ญ๐ญ๐ด".
What came back in 5 minutes is the briefing attached below.
๐ ๐๐ฎ๐ ๐ฏ๐น๐ผ๐๐ป ๐ฎ๐๐ฎ๐. But the interesting part is how it was built. I asked Claude and her was the response:
"viaNexus MCP Tools The foundational data layer. I used the viaNexus:fetch tool across eight separate dataset calls โ QUOTE, STOCK_STATS_US, COMPANY_DESCRIPTIONS, SUMMARY_NORMALIZED_FINANCIALS (quarterly and annual), NORMALIZED_FUNDAMENTALS, TRUEBEATS_EPS_REVENUE_FORECASTS, EVENTS_CALENDAR, and HISTORICAL_EVENTS_TRANSCRIPTS โ plus MT Newswires for analyst coverage. All fetched in sequence to populate every data point in the report with live, sourced figures rather than recalled knowledge.The Equity Research Report Skill (/mnt/skills/user/equity-research-report/SKILL.md) A custom skill you built in a prior session. It defined the exact data fetch sequence, the report structure (12 sections), formatting rules (human-readable numbers, no invented figures), and quality checks. Without reading that skill first, the report would have been structurally different and likely missed several datasets.Transcript Parsing via Bash The AIERA transcript files for Q4 2025 (767 entries, ~1MB) and Q3 2025 (719 entries) were too large for direct context. I used bash_tool with Python to decompress and navigate the JSON structure, extract the full transcript text, locate the Q&A sections by position, and read specific 8,000-character windows to surface the key management exchanges โ the Jeff Sprague footprint question, Steve Tusa on litigation costs, Julian Mitchell on macro vs. self-help, Scott Davis on NPI quality.The Frontend Design Skill (/mnt/skills/public/frontend-design/SKILL.md) Governed the visual output. Rather than producing a plain markdown document, I read the design skill and applied it to produce a production-grade HTML briefing with a considered aesthetic โ Playfair Display / IBM Plex Sans type pairing, a dark masthead with red accent, structured color-coded sections, bull/bear boxes, and a distinct question format with source tagging.Synthesis and Judgment The data and transcripts provided the raw material, but constructing the 11 questions, the bull/bear cases, and the analyst summary required integrating across sources โ connecting a pricing comment in the Q3 call to a margin guidance assumption in Q4, or linking the Wolfe Research target raise to the RBC bear case in the same week. That analytical layer sits outside any single tool.
In short: the skill files told me what to build and how to structure it, the MCP tools provided the data, bash gave me access to the transcripts, and the synthesis layer turned it into something useful for an actual meeting."
The "reasoning" is getting really powerful, and exceeds what most analysts would do coming to this company raw. A seasoned analyst covering the company might say otherwise. Connecting guidance shifts, analyst positioning, and management commentary into questions youโd actually ask in the room. ie
โถ๏ธ Can NPI really carry growth in 2026 given the shift from commercial execution?
โถ๏ธ When do litigation costs realistically inflect lowerโif at all?
โถ๏ธ What does footprint consolidation actually mean for the P&L?
Now hereโs the part that should make people pause - the estimated cost of this report (excluding the source content): Token Usage as estimated by Claude is estimated as per below:
| Component | Approx. Tokens |
|---|---|
| System prompt + memory context | ~8,000 |
| Skill files read (equity research + frontend design) | ~4,000 |
| MCP data responses (8 dataset fetches) | ~15,000 |
| Transcript parsing output (bash tool reads) | ~25,000 |
| Report generation (HTML output) | ~12,000 |
| Tool calls, conversation overhead | ~5,000 |
| Total estimate | ~70,000 tokens |
The approximate Cost Using Claude Sonnet 4.6 API pricing as a reference point:
| Tokens | Rate | Cost | |
|---|---|---|---|
| Input | ~55,000 | $3 / 1M | ~$0.17 |
| Output | ~15,000 | $15 / 1M | ~$0.23 |
| Total | ~$0.40 |
The data behind it came from real partnersโnormalized, permissioned, reliable: SavaNet for fundamentals, Aiera for transcripts & events, ExtractAlpha TrueBeats for forward estimates and MT Newswires for news - all flowing through and normalized by viaNexus. All accessible from a single prompt.
The takeaway isnโt โAI is impressive.โ Itโs that the architecture matters.
No specialized copilot. No heavy desktop. No brittle workflows. No unreliable, unlicensed data scraped from the internet!
Just Claude + Skills + permissioned dataโconfigured once, usable every time.
Try it for your self - with a free trial, or your subscription, connect the viaNexus MCP: https://vast-mcp.blueskyapi.com/mcp