Pick Your AI and Build the Workspace

Last lesson was the why. This one is the screwdriver. We're done talking about the racket and we're going to build the thing that lets you run a claim analyst on your couch.

The whole move fits in one breath. You pick one of two AI tools, you spin up a single workspace inside it, you redact your Social Security Number off your documents, and you load everything in once so every conversation you ever start already knows your whole file. That's the engine. Build it and the rest of this course is driving.

The profanity dials back from here and the specificity climbs. Less motor pool, more whiteboard. Let's wire it.

Two serious options, and neither one is the enemy

Two tools are worth your time. Claude, made by Anthropic. ChatGPT, made by OpenAI. Either one runs the claim, and both carry a free tier and a paid tier around twenty bucks a month. I ran Claude for the heavy lifting on my own file, and I'll tell you why, but you won't go wrong with the other.

Claude's strengths are the ones that matter most for claim work. Longer context window, which means it holds more of your documents in memory at once without losing the thread. A projects feature that keeps your documents and custom instructions attached to a dedicated workspace, so you stop re-uploading the same record every session. It follows nuanced instructions better over long conversations. And it's stronger at citing the actual text of a regulation instead of drifting into something it made up.

ChatGPT brings a bigger user base, strong web search for pulling current VA.gov pages, custom GPTs you can build and share with other vets, and a voice mode if you want to talk through your claim while you drive. The cost shows up elsewhere. Most tiers give it shorter context, so it loses the front of a long document set faster, it invents citations more often, and its custom-instructions feature is thinner.

My recommendation is the paid tier at around twenty dollars a month if you can swing it, because the persistent-project workflow is the spine of this whole guide. If twenty bucks is a stretch right now, the free tier still runs the claim. You re-upload more often and you verify harder. That verification habit is the one I set in the last lesson and the one I'll keep grinding on, because the AI is a brilliant intern with a reading addiction. It has swallowed more pages than any human alive and it will still hand you a regulation number that does not exist and smile while it does it. Check everything against ecfr.gov and VA.gov before anything you read turns into something you file.

A workspace beats a chat, and that gap is the whole game

The difference between a casual chatbot session and a claim analyst is one feature: persistence.

A plain chat is amnesia. You explain your service history, you paste in your ratings, you upload an STR, and then you close the tab and the whole thing evaporates. Next session you start from zero, re-typing your life into the box like the machine never met you.

A workspace ends that. You set the instructions once and you attach your documents once. Every new conversation you open inside it inherits the full context, the instructions and every file, already loaded, already read. You ask a question and it answers as something that has read your entire record, because it has. One setup, and from there an unlimited stack of conversations all standing on the same foundation.

Setting it up, click by click

This takes about ten minutes. In Claude you're building a Project. In ChatGPT you're building a custom GPT, or if you're on the free tier, you're using the Custom Instructions feature instead and re-uploading documents each session.

Open your account first. Free tier is fine to start. Upgrade once you hit context limits and the thing starts forgetting the front of your documents.

Create the workspace. In Claude that's the Projects area in the sidebar, then Create Project. Name it something future-you will recognize. I called mine "VA Claim." Plain and obvious beats clever here.

Find the instructions field. Claude calls it project instructions or custom instructions. This is where the system prompt goes, and the system prompt is the entire next lesson. Leave it blank for now. We come back and fill it deliberately, because that prompt is what turns a general chatbot into a strategist that never forgets the plan.

Add your files. Claude has an Add files or Project knowledge section. This is where your claim documents live. We'll cover exactly what goes in below. Upload the ones you have, redacted, and add more as you collect them.

Open a conversation inside the workspace and you're live. Anything you start in there carries the instructions and sees the documents. Your analyst, on call.

Redact before you upload, because it's free insurance

Before you drag your DD-214 into anything, know what's on it. Your DD-214 carries your full Social Security Number. Your STRs often do too. Black it out first.

This costs you two minutes. Any PDF editor with a redaction tool does it. Preview on Mac has one built in, Adobe Reader has one, and free online tools handle it too. Drop a solid black box over the digits and save.

Redact the SSN, every time, no exceptions. Then consider redacting your full date of birth, because month and year are usually fine but full DOB plus your name plus your service branch is enough to uniquely pin you. Consider redacting your home address if it shows up anywhere, because the AI does not need your address to analyze a claim.

Once the identifiers are blacked out, this is all safe to upload: service dates, branch, MOS, rank at separation, your medical conditions, C&P findings, rating decisions with identifiers redacted, STRs with the SSN redacted, lay statements, nexus letters. None of that is sensitive the way a SSN is, and all of it is the raw material the analyst runs on.

What never goes in: anything with an unredacted SSN, anything with bank account information, and photos of yourself unless they bear on a specific claim, like visible scarring you're rating. The AI works fine on redacted records. It can reason about "the veteran's back injury documented in the May 2011 STR entry" without ever seeing a single digit of your SSN. Redaction is free insurance. Pay it.

One policy note, said once. Consumer tiers of these tools generally do not train their models on your chat data by default, but those policies shift like sand, so read the current terms the day you sign up. Business and enterprise tiers carry stronger guarantees. Neither is HIPAA-compliant out of the box for consumer use. Redaction is how you stay ahead of every word of that fine print.

The document list

This is everything you want sitting in the workspace. Start with what you have and add the rest as it lands in your mailbox. You can always feed more in later.

• DD-214, redacted. Confirms service dates, branch, decorations, MOS.
• Service Treatment Records, redacted. These are gold. Every in-service complaint, injury, diagnosis, and treatment is the foundation of your nexus chain.
• Separation exam. Technically part of your STRs, but call it out, because what got noted at separation is load-bearing for service connection.
• All current rating decisions. Every rating letter you've received. The analyst needs your rated conditions and percentages to run the math.
• All C&P exam results. These are what the rater used to decide. Knowing what the examiner wrote tells you where a rating came from and where the weak spot is.
• Private medical records. Current treatment notes, diagnoses, imaging for anything you're claiming or planning to.
• Existing lay statements from spouse, family, battle buddies.
• Any previously denied claim letters, with the rationale.
• SSA disability decision, if you have one. Useful when TDIU comes into play.
• Tax transcripts, if TDIU is on the table, because income against the substantially-gainful-employment threshold becomes a question the analyst has to answer.

If you don't have your STRs, request the full set from the National Personnel Records Center through eVetRecs at archives.gov/veterans. It's free, it's online, and it usually runs two to four weeks. Post-2004 records sometimes come faster through milConnect.

One warning for the older vets. The 1973 NPRC fire ate Army personnel records discharged 1912 to 1960 and Air Force personnel records discharged 1947 to 1964, and if your file was in that warehouse you may have to file a records reconstruction request, which is a slower separate road.

Last thing, a small habit that pays off the whole way through. Keep a running Document_Inventory.md inside the workspace, a plain list of every document, when you uploaded it, and what it holds. When your file gets thick, and it will, that inventory is how you and the analyst both keep track of what's already in the room.

The workspace is live

Once the documents are loaded, you have an analyst on call. Every new conversation inside that workspace starts with your full file already read. Tell it to run the combined-rating math on your current ratings. Tell it to map secondary claim opportunities. Paste in a decision letter and ask where the denial is weak. Sit it down to interview your spouse for a statement.

It sharpens every time you feed it. Add a new C&P result, the picture updates. Drop in a new private record, the analysis deepens. The workspace is the one thing in this whole process that compounds in your favor.

It's still an intern, though. Fast, tireless, read everything, and fully willing to cite a regulation that doesn't exist if you let it off the leash. So we don't. We hand it a job description first.

Mark this lesson done and go build the system prompt. That's where the chatbot stops being a chatbot and turns into a strategist.