Claude AI vs Claude Code
A Tale of Two ETF Flow Reports
Analysis Date: 2026-02-19
J’apporte la comparaison de traitement entre claude code et claude ai sur un meme dataset ETF :
claude code est mon outil quotidien pour « coder », claude AI est l’interface de chat que tout le monde connait sur son navigateur.
Les différences essentielles :
claude AI :
web search, génération de fichiers, artifacts, image search, tools météo/sports etc.
Fonctionnement conversationnel, il attend les instructions tour par tour.
Il maintient automatiquement une mémoire des conversations fragmentaire et asynchrone et a un systeme de registre, il monte une VM qui ne persiste pas.
Claude Code :
tools orientés dev : lecture/écriture fichiers, exécution de code, appels bash, navigation dans l’arborescence. Pas de web search natif.
Branchement sur serveurs MCP, fonctionnement agentique : boucle en autonomie sur tâche complexe.
Acces au filesystem, aux settings, au claude.md.. on spécifie nous mêmes ce qu’il utilise.
Claude desktop :
je l’ai ajouté car ca a l’air d »un mix entre les deux, mais pour moi il n’est pas mature : par ex les interactions avec le filesystem windows sont problematiques, en fait dès qu’on sort de la wsl2/linux ca se passe mal. Il est néanmoins extremement prometteur avec cowork les connecteurs et les plugins.
Bref je vous laisse découvrir.
Introduction
On February 19, 2026, the same ETF flow dataset from ETFdb (4,396 ETFs) was processed through two fundamentally different approaches: programmatic code generation and Claude AI-assisted analysis. The results reveal striking differences in philosophy, presentation, and utility.
This article examines both outputs to understand when each approach excels and what trade-offs are involved.
At a Glance
| Metric | Code-Generated | Claude AI |
|---|---|---|
| File Size | ~1,000+ lines | 313 lines |
| Language | English | French |
| Tone | Neutral, data-driven | Editorial, opinionated |
| ETFs Listed | 15+ per section, 30+ commodities | 10-15 curated picks |
| Commentary | Generic labels | Personalized analysis |
| Asset Classes | 9 (including Volatility, Preferred Stock) | 5 (focused on major classes) |
Philosophy: Completeness vs. Curation
Code-Generated Report
Goal: Capture everything, decide nothing.
- All 9 asset classes displayed (Equity, Bond, Commodity, Currency, Real Estate, Multi-Asset, Preferred Stock, Alternatives, Volatility)
- Top 15 inflows and outflows with full metadata
- 30+ commodities listed comprehensively
- Machine-readable structure for downstream processing
Claude AI Report
Goal: Synthesize signal from noise.
- 5 key asset classes with narrative context
- Curated « Watchlist » organized by investment themes
- Direct portfolio implications (« ton GDX.L beneficie »)
- 7 numbered actionable takeaways
Language & Editorial Voice
The most immediately obvious difference is language and tone:
Code-Generated Commentary
Generic, templated descriptions that apply the same label to any ETF matching the divergence criteria.
Claude AI Commentary
Contextual analysis that connects multiple data points into a coherent market narrative, written in French for a specific audience.
Data Presentation
Code-Generated Table Row
<td class=« ticker »>VOO</td>
<td>Vanguard S&P 500 ETF</td>
<td class=« cat »>Large Cap Growth Equities</td>
<td>$858.53B</td>
<td class=« pos »>+$9.96B</td>
<td class=« pos »>+$3.25B</td>
<td class=« pos »>+1.2%</td>
<td class=« neg »>-0.8%</td>
<td><span class=« »></span></td>
</tr>
Verbose, includes AUM, Flow/AUM ratio, full ETF name. Empty signal span when no divergence detected.
Claude AI Table Row
<td class=« ticker »>VOO</td>
<td class=« pos »>+$9,960M</td>
<td class=« pos »>+$26,010M</td>
<td>-0.8%</td>
<td class=« cat »>Large Cap</td>
<td>SPY-to-VOO rotation continue</td>
</tr>
Compact, omits AUM, adds human-written signal interpretation in the last column.
Unique Features
Code-Generated Only
- Multi-Asset, Preferred Stock, Alternatives, Volatility macro cards
- IQLT distribution signal detection
- Full ETF names (e.g., « Vanguard S&P 500 ETF »)
- AUM and Flow/AUM columns
- 30+ commodities in deep dive
- SPYM, BND, SGOV, AGG, LQD, JAAA in top flows
Claude AI Only
- Watchlist Portfolio section with themed groupings
- Direct advice: « valide COPX, FCX, COPA »
- Takeaways with 7 numbered actionable insights
- Crypto « Oversold » tags (tag-os)
- Flow Acceleration analysis (1W vs 4W avg)
- Personalized commentary: « ton GDX.L beneficie »
Signal Detection
Both reports use similar signal classification tags:
| Signal | Code-Generated | Claude AI |
|---|---|---|
| CONVICTION | Price UP + Strong Inflows | Same + contextual narrative |
| ACCUMULATION | Price DOWN + Inflows | Same + « smart money buying » |
| DISTRIBUTION | Price UP + Outflows | Same + inline « PROFIT TAKING » |
| OVERSOLD | Not detected | RSI < 35 (crypto focus) |
Claude AI adds the tag-os (Oversold) classification and uses inline custom labels like « PROFIT TAKING » for more granular signals.
Trade-offs
| Consideration | Code | AI |
|---|---|---|
| Reproducibility | Deterministic, same input = same output | Variable, depends on prompt/context |
| Scalability | Runs in milliseconds | Requires AI processing time |
| Data Completeness | All ETFs, all metrics | Curated subset |
| Actionability | Requires user interpretation | Direct recommendations |
| Personalization | One-size-fits-all | Portfolio-aware |
| Narrative Context | Isolated data points | Connected market story |
When to Use Which
Use Code-Generated When…
- Building a screening pipeline
- Archiving historical data
- Feeding downstream analytics
- Needing audit trails
- Processing multiple dates automatically
- Sharing objective data with teams
Use Claude AI When…
- Preparing morning briefings
- Making portfolio decisions
- Communicating insights to clients
- Identifying non-obvious patterns
- Connecting flows to macro themes
- Getting second-opinion analysis
Conclusion
The Verdict: Complementary, Not Competing
The code-generated report serves as the source of truth: comprehensive, reproducible, and machine-readable. It’s the foundation for any serious data analysis workflow.
The Claude AI report serves as the interpretation layer: contextual, actionable, and personalized. It transforms raw data into investment intelligence tailored to a specific portfolio and user.
The optimal workflow combines both: code for data integrity, AI for insight extraction. The code ensures you never miss a signal; the AI ensures you understand what it means.