Folder IT’s engineering team continues publishing real technical work in the open. That’s the same pattern behind a recent Agentforce Revenue Management configurator project and a ServiceNow-to-Teams integration. This time, though, the subject sits closer to the core of what Folder IT sells: evidence-traced AI planning. The team built it as a real, open-source Claude Code tool, not a capability described in a sales deck.
Folder IT calls the tool Discovery Workbench. It turns a business problem plus supporting evidence into a cited, Obsidian-ready plan. In addition, it enforces a rule most AI tools skip entirely: every material claim in the output carries either a source citation or an explicit [assumption], [hypothesis], or [open question] tag. The tool never asserts a fact without one of the two. The full build is public on GitHub.
The Problem: AI Agents That Assert Instead of Cite
Most AI planning tools produce confident-sounding prose. There’s often no way to tell which parts came from real evidence and which parts the model filled in. That gap is a large part of why AI initiatives stall inside enterprises. In fact, Gartner has flagged governance — not model quality — as the top reason agentic AI projects risk cancellation. More than 40% are at risk by 2027 for exactly that reason.
A business plan is a worse place for that gap than most. A recommendation can sound authoritative but trace back to nothing verifiable, and that can send a real budget decision in the wrong direction. Claim discipline — citing every material claim or flagging it plainly as unverified — is a structural fix, not a prompting trick.
What Discovery Workbench Actually Does
The tool runs a fixed pipeline: evidence, then map, then plan, then review. A user drops source documents — PDFs, spreadsheets, interview notes — into a case folder. Deterministic Python scripts convert them to text. Claude Code then builds an append-only evidence map. From there, it generates a business plan from numbered templates, and finally reviews that plan against the evidence map before final export.
Append-only evidence IDs matter here more than they sound like they should. Once an evidence entry gets an ID, later steps can only add new entries. They can’t quietly edit or delete an earlier one. As a result, every claim in the final plan traces back to a specific, unaltered piece of source evidence.
Why It Runs Local, Not as a Hosted AI Agent
Discovery Workbench makes no Anthropic SDK calls, sends no API requests, and touches no network at runtime. Claude Code is the reasoning engine for every step. The Python scripts around it, on the other hand, are deterministic glue — document conversion, CSV summaries, report assembly. Business data and generated plans stay on the user’s machine, git-ignored by default.
That architecture decision tracks a broader pattern in enterprise AI evaluation. Buyers increasingly check where their data lives during an AI process before they check anything else about the tool itself. The same filter shows up across evaluations of custom AI development company proposals more broadly. A local-first, network-free design answers that question before anyone even asks it.
The Five-Role Architecture
Five roles make up the system, each with a fixed job. CLAUDE.md holds the project’s load-bearing rules, auto-loaded every session. Commands are the /workflow steps a user runs by hand — /create-case, /process-evidence, /generate-business-plan, and so on. Skills are reusable procedures Claude invokes on its own, like evidence extraction or Obsidian export. Templates define the output structure for each plan section. Scripts handle the deterministic, non-LLM work — no reasoning, no network, just glue code.
A full file-by-file breakdown of every command, skill, script, and template lives in docs/ARCHITECTURE.md. A complete worked example — a fictional case run start to finish — lives in docs/EXAMPLE.md.
This Is What an AI Pod Actually Builds
Folder IT’s AI Pod model pairs a senior Forward Deployed Engineer with AI coding agents. Those agents form a core part of the execution stack. An internal AI Lab backs that engineer up for architecture and technical review. That’s a specific claim about how the work gets structured, and Discovery Workbench is one direct example of it. A Forward Deployed Engineer used Claude Code as the reasoning runtime, with structural guardrails built in: the evidence firewall, the claim discipline, the append-only IDs. In practice, those guardrails do the work a project manager coordinating a team of developers would normally have to enforce by hand.
The related AI Pod reference architecture documents the broader pattern. Discovery Workbench is a working instance of it, not a diagram.
Frequently Asked Questions
What is Discovery Workbench? It’s a Claude Code–native, local-first tool. It turns a business problem plus supporting evidence into an evidence-traced, Obsidian-ready plan, using a fixed evidence-map-plan-review pipeline.
Does Discovery Workbench call an external AI API or send data over the network? No. Claude Code runs the reasoning steps locally. The Python helper scripts make no Anthropic SDK calls, no API requests, and no network calls of any kind. Business data and generated plans stay on the user’s machine.
What is claim discipline in an AI-generated business plan? It’s a rule requiring every material claim in the output to carry a citation to a real source. Otherwise, it needs an explicit [assumption], [hypothesis], or [open question] tag. That way, a reader can immediately tell what’s evidenced and what isn’t.
What is a Folder IT AI Pod? It’s a delivery model built around one senior engineer — a Forward Deployed Engineer — who owns a problem end to end. That engineer works with AI coding agents as a core part of execution, and an internal AI Lab backs the work up for review. It’s not a project manager coordinating a team of developers.
Can this tool be used commercially? Yes, under the MIT license specified in the repository.
The Method Is the Product
Discovery Workbench is a small tool solving a specific problem. It turns scattered evidence into a plan a reader can actually verify, claim by claim. What it demonstrates matters more than what it does, though: evidence-traced AI planning and structural claim discipline are buildable today. They work in the open, too, without a hosted API or a black box in between.
Folder IT builds AI Pods and custom AI systems for enterprises that need the same discipline applied to their own problems. Get in touch to scope what an AI Pod engagement would look like for a specific use case.
