Claude Code for Finance: From Information Gathering to Judgment — Part 1 of 3

The junior analyst still needs a decision-maker

A junior analyst gathers a mountain of information. Someone still has to turn it into a good decision. That is Doug O’Laughlin’s opening model for AI in this Latent Space conversation: useful for tedious research, frequently mistaken, and powerful under expert supervision. Information gathering and investment judgment are different jobs.

Repeated analysis teaches a human analyst where their judgment is reliable: which problems they understand and where their methods consistently work. O’Laughlin does not think AI yet has that same ability to learn about its own expertise. His enthusiasm rests on a narrower claim: AI substantially amplifies existing experts, provided they still synthesize the work and decide what matters.

ASML and the end of the old semiconductor playbook

O’Laughlin’s own apprenticeship began with investing. Swyx knew him through the anonymous finance account Value Mule, with their early Twitter interactions dating to roughly 2017–2018. His interest in semiconductors came independently, while searching for quality companies that could compound over time. Discovering ASML in 2018 sent him into semiconductor research: textbooks, manufacturing complexity and the question of which companies could build these technologies.

He describes semiconductor manufacturing as science fiction made real and recommends Asianometry’s extensive ASML coverage. But technical fascination leaves an investor with another problem. Bank-arranged visits and investor-relations presentations can filter out the understanding needed to judge a technology. Even after deeper study, the question remains: what changes this quarter or this year, and what is already reflected in the price?

The semiconductor thesis challenges the old description of a consolidated, mature industry with limited growth. Its premise is that Moore’s Law was ending: shrinking chips could no longer deliver the routine performance gains customers had taken for granted. Further improvement would depend more on difficult engineering across chips and systems, increasing the value and potential pricing power of companies capable of doing it. Nvidia’s advantage is framed as system-wide expertise in chips, networking, design and scale-up.

Recalling a thesis developed around 2020, the researcher connects two forces: - Growing compute demand: AI scaling laws pointed toward increasing demand for computation. - Slowing performance supply: weaker gains from chip miniaturization made additional performance harder to obtain. Together, these suggested stronger semiconductor economics and a larger role for parallel computing. The researcher’s conclusion at the time was unusually concentrated: Nvidia would be effectively the sole beneficiary.

That conviction also led to online semiconductor debates with Dylan, followed by meetings and conferences. The researcher describes their perspectives as complementary: Dylan approached the industry through technology, while their own background was more financial. Even so, Nvidia’s eventual success exceeded their expectations.

Another participant challenges the retrospective success story: Nvidia was one thesis among many, so what happened to the ideas that failed? No accounting follows. The career principle offered is commit when a large trend meets strong conviction. The researcher recalls an early fascination with TikTok in 2019, but describes the semiconductor thesis as the trend around which they reoriented their life.

A hiring case study becomes an agent test

The move toward Claude Code for financial research starts from the same objective: find an information advantage and combine facts in ways others have missed. A SemiAnalysis researcher uses the firm’s financial-analyst hiring case study as an informal agent benchmark. It asks for company analysis in a prescribed format—difficult, multistep work the researcher describes as taking a human 24 hours. Can an agent complete it in one attempt, and how does its output compare with human submissions?

The experiments included ChatGPT agent mode, Gemini CLI and Claude Code. Earlier projects with Opus 4 needed substantial feedback and suffered mistakes; a Codex experiment also fell short of the desired seamless execution. During holiday experimentation around December 26–27, the setup described as “Claude Code 4.5” began producing usable initial projects in one attempt and responding productively to requests for improvements. The threshold was output worth building on. This was a personal assessment, without controlled scores or fully specified model configurations; the year is not stated.

The work mixed dashboards and Excel. The researcher found the agent moderately capable of producing a basic financial model or moving information between places, though still below human level at spreadsheet work. Its usefulness did not depend on perfecting every estimate: in the researcher’s view, small improvements in earnings-per-share precision do not themselves produce good investment decisions.

That distinction underlies a criticism of sell-side research—the bank-associated research business—as aging and insufficiently differentiated. SemiAnalysis’s research approach emphasizes the timing and magnitude of technological change. The researcher argues that understanding these inflection points matters more than mechanically maintaining slightly different earnings estimates.

Product readiness can dominate spreadsheet precision. The hypothetical example is AMD’s Helios rack: if it arrived on time and could begin generating tokens immediately, the researcher argues, that could change AMD’s revenue by billions of dollars. The example concerns the business consequences of readiness and timing; it does not report that those milestones have been achieved.

The research task is to find the one to three factors that actually matter. A better component matters financially when its advantage creates a substantial opportunity in its application. Recognizing that requires depth in specialized components and breadth across the supply chain: a small upstream change can affect a much larger downstream business. Learning enough to make that connection is expensive.

From portfolio notes to a daily adoption monitor

One personal workflow begins with stock positions and scattered investment notes. Claude Code consolidates them, assembles a portfolio and performs basic risk analysis. The next request makes the researcher’s judgment more explicit: build an investment framework matching their style, grade the holdings and challenge them. The reported benefit is turning an implicit investment process into something that can be inspected and revised.

Another project tests whether enthusiasm for Claude Code corresponds to observable adoption. It turns a recognizable commit signature into a recurring analysis: 1. Collect matching commits: supply the signature and query GitHub activity through its API. 2. Examine volume and share: chart matching commits and their proportion of GitHub activity. 3. Refresh daily: schedule the analysis with a cron job.

The researcher describes unusually rapid growth. Another participant contrasts this with PR Arena’s emphasis on merge rates: adoption share and acceptance rate answer different questions. The GitHub query scope, denominator and exact dates are unspecified, so the approximate percentages discussed do not establish a precise growth rate or share of all generated code.

Chart production becomes reusable as well. The researcher gives an Opus model SemiAnalysis’s color scheme, style advice and summaries of visualization books, then compresses the guidance into a skill—a reusable set of instructions for producing charts from different datasets. The participants agree that consulting roughly 70 books may be excessive. The rationale is that broad synthesis feels cheap enough to try, while the resulting instructions should stay short enough for repeated use.

The model can also help select a representation: given a collection of dates and a request for visual alternatives, it proposes a timeline. This extends the argument beyond programming. Financial models also construct abstractions and relationships, so the researcher sees better coding agents as a reason to explore automation across information work.

Separate evaluation; preserve coherent project context

To make model judgments more consistent, the researcher uses a rubric: explicit criteria with scores, such as assessing X, Y and Z out of ten. They report less variation between responses, while errors remain. Two placements serve different purposes: - Task and rubric together: the model receives the criteria while producing its answer, allowing it to evaluate and revise its own work. - Rubric afterward: gathered material receives a separate assessment, which another participant suggests may also attract more human attention to the evaluation.

The researcher prefers a fresh context for evaluation because generation and assessment can become entangled. In their experience, Opus 4.6 is overly agreeable, and separate evaluation seems to reduce drift toward endorsing existing work. That is a personal workflow judgment, not evidence that separation eliminates bias.

For execution, the preference is to keep a project within one context window—the material available to the model as it works—when the project is small enough to fit. In a million-token window, skills and CLAUDE.md project instructions occupy a smaller fraction of capacity than in a 200,000-token window, leaving more room before a reset. Sub-agents can work in separate contexts and return material to the main one. The researcher views repeated compaction, or compression of the accumulated context, as potentially preserving noise in compressed form rather than restoring a clean project state.

Code does the work; human formats deliver the result

One participant proposes a hierarchy from chips to AI to software engineering, with code as the interface for other information work. Under this view, data analysis can run through code even when its final product is a chart or financial model.

The concrete comparison is Claude for Excel versus Claude Code producing Excel output. The researcher strongly prefers Claude Code using Python and Excel skills in their own work. Their proposed division of labor is machine-oriented processing followed by human-oriented output: let agents assemble and synthesize information through code, then deliver the format the reader needs. A Matplotlib chart delivered as an image can serve the purpose without manual construction in Excel or PowerPoint.

Another participant raises a production tradeoff: new charts may look inconsistent with older ones. The researcher accepts some inconsistency, retaining the watermark while prioritizing faster delivery and what they consider better formatting. Visual continuity can give way when useful information arrives faster.

Bounded sub-agents versus coordinating teams

The agent-team trial involved analyzing many companies with different key performance indicators and combining the results into one dashboard. The researcher asked Claude to divide the work among teams and found the outcome disappointing. Their qualitative comparison distinguishes three approaches: - Ordinary sub-agents: preferred for bounded assignments with project instructions, followed by a return to the main agent. - Claude agent teams: reported to worsen results in these experiments; another participant emphasizes their experimental status. - Kimi 2.5 Agent Swarm: reported to improve performance meaningfully in the researcher’s experiments.

The researcher also describes an internal benchmarking workflow: give models a set of problems, run them repeatedly—20 times in the example—and examine performance and qualitative differences. No scores, problem set or evaluation rubric accompany the comparison. The suggestion that training differences explain the team results remains speculation; the exchange does not establish the systems’ training procedures.

The larger opportunity is verified intermediate outputs feeding subsequent work. The researcher imagines linking agents into a larger process, with each producing material the next stage can use. Coherent context windows and efficient token use would be necessary to keep that process productive.

A brief discussion of product strategy follows. One participant sees Kimi as more focused on turning models into products than Qwen or DeepSeek. Another argues that Qwen has a route through Alibaba, but also faces competing priorities across its parent company’s businesses.

Agents lower setup effort but still need supervision

The enthusiasm becomes more guarded around OpenClaw’s access to personal information. One participant initially enjoyed letting it read email and calendar material, then withdrew over concern that those sensitive inputs were exposed to prompt injection. No exploit is reported. The other uses a staged approach: start with a separate, lower-stakes email account and expand access only after the bot proves useful. This bot had not earned that expansion.

Even a focused terminal workflow needs supervision. The researcher describes reminding Claude Code to use an API already configured in .env and specified in CLAUDE.md. Tool availability does not ensure tool use. The agent harness—the surrounding instructions, tools and controls—still needs work, and the researcher still has to redirect the agent.

The automation discussion separates two practical considerations: - Conventional workflows can offer tighter control: one participant argues that many bot use cases could have been implemented more securely in Zapier. - Agents reduce the effort to finish: the other values eliminating the clicking, copying and configuration that can keep a workflow from reaching completion. Their dramatic setup-time comparison is illustrative rather than measured. The product argument is that reducing setup effort changes which automations people actually complete.

One participant frames Zapier’s challenge as an innovator’s dilemma: a successful business and large user base organized around strict workflows can constrain a shift toward flexible agents. Earlier assumptions become organizational constraints, echoing the semiconductor discussion’s challenge to a mature-industry playbook.

The unfinished setup: shared context and underused hooks

Part 1 closes with a setup still under construction: a few base skills, many APIs and a SemiAnalysis internal server accessed through an API. Its purpose is to give researchers a shared baseline of research context. Making the relevant information available across the firm is the central priority.

The final exchange turns to hooks. One participant argues that they are underrated and says a Ralph Loop can be implemented with a hook; the researcher acknowledges underusing them. The terms are introduced but not explained before this part ends, leaving the implementation for a fuller discussion.