Claude Projects for EB-1A Petitions: Attorney Setup Guide

The pattern is consistent enough to be recognizable. An attorney opens Claude, types something like "write a Criterion 5 argument for my client who has 847 citations," gets back two paragraphs of vague language about "significant contributions to the field," and concludes that AI does not work for EB-1A drafting.

The conclusion is wrong. The setup is the problem.

Claude without a configured project has no legal context, no knowledge of the specific client, and no instruction about how you want the output formatted. You are asking a skilled writer to produce a legal brief without telling them what country, what legal standard, what case, or what form the output should take. Of course it fails. Configured correctly, the same model produces first drafts that need attorney editing, not rewriting.

That gap (from generic output to editable first draft) is almost entirely a configuration problem. This guide fixes it.

What Claude Projects Is and Why It Changes the Workflow

A Claude Project is a persistent workspace that lives in your Claude account. It has three components: a knowledge base (files you upload), project instructions (a system prompt you write), and the conversation history from all sessions inside that project.

The key behavior: every conversation you open inside the project starts with the knowledge base and project instructions already loaded. You do not re-explain the legal standard. You do not re-paste the client evidence list. You do not re-specify the output format. All of that loads automatically.

This matters for EB-1A work because the same legal context applies across dozens of tasks for the same client. The Kazarian v. USCIS, 596 F.3d 1115, 9th Cir. 2010 two-step framework, the client's selected criteria, the exhibit numbering system: these should be established once, not reconstructed at the start of every session. Without a project, every conversation starts cold. You lose the previous session's context completely. The model that spent twenty minutes with you working through the Criterion 8 argument forgets everything when you open a new tab.

With a project, that prior context persists. Not through conversation memory (Claude does not have persistent memory across separate conversations) but through the knowledge base and instructions you have configured. The model does not "remember" your last session, but it starts each new one already knowing everything you put in the project setup.

For a single EB-1A petition with ten criteria, five expert letters, and a final merits argument, that configuration investment pays back in the first two sessions.

Which Plan You Need

Claude Projects is available on all paid plans: Pro ($20/month), Team ($25-30/seat), and Enterprise (approximately $60/seat). The technical features of Projects are similar across plans. The difference that matters for attorneys is data handling.

Pro and Team plans: no BAA. You can use these for testing your setup with anonymized data, for building template argument structures, and for any task that does not involve client-identifying information. A lot of useful EB-1A work falls into this category: drafting generic criterion frameworks, analyzing publicly available AAO decisions, building your system prompt template.

Enterprise plan: BAA available. This is the only tier appropriate for uploading actual client documents with names, A-numbers, employer details, or other identifying information. The BAA is not automatic. An administrator needs to activate HIPAA compliance in the Enterprise settings. Once activated, your inputs are not used for model training and the data handling meets the contractual standard that ABA Model Rule 1.6 compliance requires.

If you are building your first EB-1A project with anonymized test data to learn the system, Pro works fine. If you are setting up a client-specific project with real exhibits, you need Enterprise.

Steps 1–2: Create the Project and Write Instructions

Log into claude.ai. In the left sidebar, look for the "Projects" section (it appears below your recent conversations). Click "New Project."

Give the project a name. For client-specific projects, a naming convention like "EB1A / [Client Last Name] / [Year]" keeps things organized when you have multiple active matters. For a general EB-1A practice project with templates and reference materials, something like "EB-1A Templates" works.

After creating the project, you land on the project configuration screen. Two sections: "Project Knowledge" (the file upload area) and "Project Instructions" (the system prompt text field). Complete the instructions first, then upload files. Getting the instructions right before you upload is easier because you will reference the uploaded materials in the instructions.

Writing the project instructions

The project instructions field is the highest-value configuration in the entire setup. This is the system prompt that loads with every conversation. For EB-1A work, it needs four things: the legal standard, the client context, the output format, and the evidence constraint.

Below is a template you can copy and adapt. Replace the bracketed sections with client-specific information.

A few notes on each section.

The legal standard block includes the exact statutory language from 8 CFR 204.5(h) and names the Kazarian framework. Without this, the model treats EB-1A as a general immigration topic and misses the two-step structure. With it, every criterion argument the model drafts is framed around what Step 1 actually requires.

The client context block is what makes this a case-specific project rather than a generic template. Paste in the beneficiary's name, their field (specific: "computational materials science" not "science"), the criteria you have selected for this petition, and any context that helps frame the evidence: an exceptional achievement the beneficiary is known for, the key exhibits, the field's relevant community. One or two short paragraphs is enough.

The output format block prevents the two most common failure modes: output that does not cite exhibits, and output that invents facts that sound plausible. The "one section at a time" instruction prevents the model from trying to draft the entire petition in one response, which produces surface-level coverage of every criterion instead of deep analysis of one.

The constraint block is the most important line in the template. It tells the model to flag rather than fill gaps. Without it, models will produce confident-sounding arguments that cite general knowledge about the field rather than the actual exhibits. With it, you get honest output that says "I cannot find support for this claim in the provided materials." That is far more useful.

Step 3: Upload Your Knowledge Base

The knowledge base holds documents that load as context in every conversation. Not every possible document. The right documents. More files is not better if most of them are irrelevant to the current task.

Upload these:

An anonymized client profile. One document, two to three pages: beneficiary's field, current role and institution, evidence list with exhibit numbers, selected criteria and the primary evidence supporting each one. This is the reference document the model will cite in almost every conversation.

The criteria checklist with exhibit mapping. A simple table: criterion number, criterion name, exhibits assigned, brief description of what each exhibit shows. This lets the model anchor arguments to specific evidence rather than reasoning in the abstract.

Relevant AAO non-precedent decisions for the client's field. Two or three decisions from USCIS's USCIS AAO decisions database that illustrate how the AAO has evaluated similar evidence in this field. These are public documents, not client materials, so they are appropriate for any plan tier. They help the model understand how adjudicators frame comparable evidence. The AAO EB-1A decisions from 2024-2025 article covers the most relevant recent decisions by field.

Your firm's preferred argument templates (optional). If you have an existing brief structure or argument framework you have used in successful petitions, upload it. The model will use it as a style reference.

Do not upload these:

Full exhibit PDFs unless you are on Enterprise with BAA activated. For the purpose of this project, a one-page summary of each exhibit covers what the model needs: what the document is, what facts it contains, what criteria it supports. Full PDFs consume context space and make it harder for the model to retrieve the specific facts you need.

Confidential financial records, salary data, or documents that contain Social Security numbers or A-numbers, unless you have verified your plan's data handling terms.

Multiple overlapping reference documents on the same topic. Pick the most relevant AAO decision per criterion, not five decisions that say the same thing.

Step 4: Structure Your Conversations

Once the project is configured, how you use it matters as much as how you set it up.

One conversation per petition section works better than a single long conversation. Open a new conversation for the Criterion 5 argument, complete that section, close the conversation. Open a new one for Criterion 8. The context window fills up quickly with long back-and-forth exchanges, and older parts of the conversation start dropping out of context as the session grows. Separate conversations per section keeps each one tight.

Name your conversations. Claude lets you rename conversations. Use names like "C5: publication impact argument" or "Expert letter brief: Dr. Singh." When you return to a conversation to revise a section, a clear name is faster than scanning a list of untitled sessions.

Use the knowledge base to anchor, not to replace the conversation. The project instructions and knowledge base load at the start, but they are background context. For each conversation, provide the specific facts you want the model to use: the exact exhibit numbers, the specific data points, the name of the expert. "Use the evidence list from the knowledge base" is less precise than "Draft the Criterion 5 argument using Exhibit 12 (citation data), Exhibit 13 (h-index report), and Exhibit 14 (Google Scholar screenshot)."

This specificity is not extra work. It is what separates editable output from generic output. The model cannot make good choices about which exhibit to cite most prominently without you telling it.

Five Prompts to Test Your Setup

Once the project is configured, run these five prompts in order. They test different parts of the setup and give you a realistic sense of what the model can do before you rely on it for a real section.

Prompt 1: Evidence strength assessment

This tests whether the model has correctly absorbed the client profile and can reason about relative evidence strength. If it returns a generic answer about what strong evidence looks like rather than specific analysis of the uploaded client materials, the knowledge base is not loading correctly or the client profile document needs more specificity.

Prompt 2: Single criterion argument

This tests the core petition drafting task. The output should cite specific exhibit numbers, reference facts from the client profile (not general knowledge about the field), and stay within the word count. If it produces output without exhibit citations, check that the evidence constraint in your project instructions is correctly formatted.

Prompt 3: Weakness identification

This tests the model's ability to apply the Kazarian framework critically. An attorney reviewing this output before a USCIS filing needs honest analysis, not just argument construction. A well-configured project produces specific concerns ("Criterion 2 relies on one professional organization with unclear membership standards") rather than generic cautions.

Prompt 4: Expert letter brief opening

Expert letter briefings are where most attorneys report the largest time savings. A good briefing takes 45-90 minutes to write from scratch. A project-generated first draft takes two minutes and requires 15-20 minutes of attorney editing. The output should reference the expert's relationship to the beneficiary and frame the specific evidence the expert will address.

Prompt 5: Step 2 totality argument

This is the hardest task and the best test of whether the project instructions have correctly set up the two-step framework. Step 2 requires arguing that the totality of evidence rises to the level of "that small percentage at the very top," not just that criteria are met. If the model conflates Step 1 and Step 2, the legal standard block in your project instructions needs more specificity about the distinction.

Managing Multiple Client Projects

Once you have one working project, the setup for a second client is faster. The project instructions template stays the same across all EB-1A matters. Only the client context block changes. The knowledge base structure stays the same: client profile, criteria checklist, AAO decisions, optional firm templates. The AAO decisions may even stay the same if your clients work in related fields.

A naming convention for the project list makes the Claude interface manageable. "EB1A / [Last Name] / [Year Filed]" works well. Archive projects for completed matters rather than leaving them active. The Claude interface shows all projects in the sidebar and a long list of old matters clutters the workspace.

For practice management, keep two types of projects separate: client-specific projects (one per active matter, Enterprise tier if the client has uploaded files) and a practice reference project (generic, Pro or Team tier acceptable) that holds your firm's template arguments, research on common RFE issues, and a library of useful AAO decisions. The practice reference project is the one you build over time and share across matters. It does not contain client data, so it does not require Enterprise.

The practice reference project structure:

• A master criteria argument template for each of the ten EB-1A criteria
• Your annotated notes from the 2024-2025 AAO decision review for each relevant field
• A one-page RFE response framework for the criteria that generate the most officer challenges (Criterion 5 and Criterion 2 account for the majority)
• Any publicly available AAO precedent decisions relevant to fields you handle regularly

Practical Limitations

Claude Projects does not eliminate attorney judgment. It replaces blank-page time, not review time.

Where Claude Fabricates

The model invents facts. Not randomly, but in specific, predictable patterns. Knowing the patterns makes review faster.

Specificity inflation. When you ask Claude to strengthen a vague claim, it adds precision the evidence does not support. "Published in a well-regarded journal" becomes "ranked in the top 5% of journals in the field." If the exhibit does not contain that ranking, the sentence is fabricated. The more specific the output sounds, the more carefully it needs verification.

Citation count extrapolation. Claude will produce a sentence like "the petitioner's work has generated over 2,400 citations across three major databases." If your exhibit shows 2,400 Google Scholar citations, that may be accurate. If the exhibit shows 2,200, Claude rounded up. If the exhibit covers a single platform, Claude inferred the rest. Both are common failure modes. Every number needs to trace back to its source exhibit.

Expert credential bleed. In expert letter drafts, Claude sometimes attributes credentials from one expert to another, particularly when multiple letters have been briefed in the same project session. An expert described as having "35 years of experience at Stanford" may have been briefed as having 25 years, or that detail may belong to a different expert in the same conversation. Verify every expert's stated credentials against the briefing document for that specific expert.

Policy claims without dates. Claude will state that "USCIS officers are required to consider..." or "the Policy Manual provides..." These claims may have been accurate as of the model's training data. The model has a cutoff and does not know about memos issued after it. Any policy claim needs verification against the current USCIS Policy Manual at uscis.gov before filing. CFR quotations need the same check: a paraphrased version of 8 CFR 204.5(h)(3) is not acceptable in a USCIS filing, even if it is close. Verify every regulatory quotation verbatim against e-CFR.

Context Window Behavior

A Claude conversation has a token limit. When a conversation approaches that limit, the model drops the oldest content silently. It does not warn you. It does not say "I no longer have access to the facts from the beginning of this conversation." It continues producing text confidently, but it is no longer anchored to the facts you established earlier.

For EB-1A work, this creates a specific risk. The exhibit facts you specified at the start of a session ("Exhibit 12 contains the citation report showing 2,200 Google Scholar citations") can fall out of context by the fourth revision round. The model fills those gaps by inference, not by re-reading the exhibit.

One conversation per section solves this structurally. When you finish the Criterion 5 argument, close that conversation and open a new one for Criterion 8. Specify the relevant exhibit facts again at the top of the new session. Keep each session short enough that early context stays in the window throughout.

If you need to revise a section later, paste the relevant draft text and the supporting exhibit facts into a fresh conversation rather than scrolling back into an old one. Prior sessions do not retain context once closed.

Document Retrieval Degradation

More files in the knowledge base does not mean better output. Past roughly 15 substantial documents, the model's ability to retrieve specific facts from specific files degrades. It starts blending information across documents.

For EB-1A work, exhibit attribution is filing-critical. "The petitioner's h-index of 23, as documented in Exhibit 14" needs to come from Exhibit 14. If you have 30 files in the knowledge base and the h-index appears in three of them with slightly different values (the Scopus report, the Google Scholar screenshot, and the CV), Claude may produce a composite that does not match any of them exactly.

Keep the knowledge base lean: the client profile, the criteria checklist, two or three AAO decisions, and a firm template or two. For each drafting session, paste the specific exhibit summaries for the section you are working on into the conversation directly. Precise retrieval from a short context beats fuzzy retrieval from a large one.

A project configured for one client produces poor output for another. Opening a new client matter in an existing client project is worth avoiding even briefly. The model blends context from both matters if you do not keep them in separate projects.

The full prompt library for EB-1A and O-1A work covers thirty task-specific prompts across five categories, all designed to work with the project setup described here. For a broader comparison of Claude against other tools available in 2026, the AI tools comparison for immigration attorneys covers pricing, BAA status, and workflow fit across the main options. For the broader picture of AI tools available for immigration practice, the AI in legal practice hub covers ethics guidance, BAA requirements, and tool selection across all visa categories.

If you want all of this without the setup work (document upload, automatic evidence-to-criteria mapping, petition generation grounded in your specific exhibits), that is what Immigration Copilot is built to do.