Detailed Analysis
A Reddit user has documented a workflow that uses Claude alongside the Model Context Protocol (MCP) to bridge TradingView's Pine Script language with Python-based backtesting infrastructure. The process involves several discrete steps: taking an existing TradingView indicator written in Pine Script, converting it into a strategy-format Pine Script, then transferring that logic into a separate chat session where it gets translated into Python for parameter sweep backtesting across chosen assets and timeframes. While functional, the user notes that the workflow breaks down when started fresh in a new conversation, suggesting the multi-step conversion process isn't reliably reproducible without some persistent structure to anchor it.
This scenario illustrates a common friction point in "vibe coding" — a term that has emerged to describe iterative, conversational software development with LLMs where users describe intent in natural language and let the model handle implementation details. The core challenge here isn't Claude's technical capability (it clearly can perform each individual translation step), but rather context continuity: without a mechanism to preserve the exact sequence, conventions, and edge-case handling established in one session, each new chat effectively starts from scratch, forcing the user to re-derive institutional knowledge every time.
This is precisely the gap that Anthropic's "Projects" and "Skills" features are designed to address. Projects allow users to persist context, custom instructions, and reference files across an entire body of related work, while Skills (a newer Anthropic capability) let users package reusable, structured procedures — essentially turning a one-off successful prompt sequence into a repeatable, documented capability the model can invoke consistently. For a workflow like this one, which chains together Pine Script-to-strategy conversion, cross-language translation to Python, and parameter sweep logic, formalizing the steps into a Skill would let Claude execute the pipeline deterministically rather than relying on the model "remembering" undocumented tribal knowledge from a prior chat.
More broadly, this case is a small but telling example of how technical and quantitative traders are adopting LLM-driven coding agents to automate financial engineering tasks that traditionally required dedicated quant developers — writing custom indicators, converting between proprietary scripting languages, and running systematic backtests. It also reflects a broader pattern across Claude's user base: as people move from single-shot prompting toward building semi-autonomous, multi-tool pipelines (chaining MCP servers, code execution, and file-based memory), the demand for durable, shareable "recipes" — rather than ephemeral chat instructions — is becoming a central design consideration. The user's instinct to formalize this into a Skill or Project, with an eye toward eventually packaging it as a standalone app, mirrors a broader trajectory in which power users are increasingly treating LLM agents as programmable infrastructure rather than one-off assistants.
Read original article →