Brainstorm, Process, Implement

Brainstorm, Process, Implement is a simple, repeatable workflow for working with AI agents and large language models.

It's designed for advisors and professionals who are relatively new to LLMs — and want to skip the trial-and-error phase.

Why this matters

Most people new to LLMs jump straight to asking the model to produce a finished output — and then get frustrated when the result is inconsistent, incomplete, or wrong. The issue isn't the model. It's the missing prep work.

BPI separates the thinking from the doing. Brainstorm the approach first. Process your data and set up your workspace. Then implement — with clean inputs and a clear plan. The framework works for anything from a quick research memo to a multi-week data pipeline.

The Framework

Three steps. In this order. Every time.

Brainstorm

Think out loud with the model before you build anything

The single most underrated step when working with LLMs is the brainstorming phase. Before writing a single line of code or structuring a single file, talk through the problem with the model.

Describe what you're trying to accomplish. Ask the model to opine on approaches, surface tradeoffs, and suggest architectures you might not have considered. Treat it like a thought partner, not a command line. Refine the plan iteratively — the back-and-forth is where the best ideas emerge.

This is especially critical for multi-step, long-form research projects where the wrong architecture at the start can cost you hours later. Ten minutes of brainstorming can save ten hours of rework.

Process

Set up your workspace and preprocess your data before implementation

Once you have a clear plan, the processing phase is about making the implementation as frictionless as possible. This is where you do the unglamorous but essential prep work that separates clean, reliable outputs from frustrating, error-prone ones.

Create a dedicated folder for the project. Add a date-stamped Markdown file with your holdings, assumptions, or inputs. If you're working with CSV or PDF data, preprocess it — clean the column names, handle missing values, standardize date formats. Set up a SQLite database if the project calls for it, and consider a cron job to keep data fresh on a regular cadence.

The rule of thumb: the easier you make it for the model to consume your data, the better your results will be. A well-structured input file is worth more than a perfectly crafted prompt.

Implement

Execute the plan with clean inputs and a clear architecture

With a solid plan and prepped data, the implementation phase is where the model actually builds. This could mean generating a Python script for portfolio attribution, creating a research report, or setting up an automated dashboard.

Because you invested time in brainstorming and processing, the implementation is tight and focused. The model already understands the architecture, the data is clean and accessible, and the output format has been agreed on. You're not debugging structure and data at the same time.

Run the output, review it, and iterate. The feedback loop at this stage is fast because the foundation is solid. What used to take days of trial and error now takes minutes of refinement.

Real-World Example

See us use the BPI method to build out a robust attribution engine

Here's what a portfolio attribution exercise looks like when you follow the BPI framework, step by step.

Brainstorm

Tell the main agent what you want to build. Here: daily portfolio attribution tracking across multiple portfolios with constituent securities.
Let it opine on the architecture — SQLite vs Markdown, cron cadence, look-ahead bias, benchmark mapping — and flag what you missed.
Go back and forth. Refine until you have a clear plan you're confident in.

Process

Upload the source data. Here, a PDF with portfolio weights by date.
Spin up a dedicated PDF extractor agent to pull the data into clean, normalized Markdown columns — ready for SQL ingestion.
QA the extraction: have the main agent review the output against its own read of the PDF to catch discrepancies.

Implement

The main agent writes a deterministic Python script for the attribution engine — written once, reused forever.
Set up a cron job to pull daily pricing and run the report on a regular cadence.
Hand off artifact generation to a specialist agent to produce a polished, interactive HTML memo for circulation.

The same BPI pattern applies to any project — research reports, model dashboards, compliance workflows, and beyond.

Three things to remember

01 · Don't skip brainstorming

Ten minutes of thinking out loud with the model saves hours of rework. The model sees patterns you don't — let it.

02 · Clean data beats clever prompts

A well-structured input file is worth more than any prompt engineering trick. Garbage in, garbage out still applies.

03 · Save what works

Once a BPI workflow produces great results, save it as a reusable template. The goal is to do the thinking once and execute forever.

Ready to put BPI into practice?

AdvisorClaw gives you the agents, the workspace, and the tooling to run the BPI framework on real advisory workflows.

Request early access →Explore resources