Prompt Chain in Claude: Reading a Published Policy Update and Listing the Matter Types to Review

What This Builds

A long policy update lands, and you need two answers before you can plan anything: what does it say, and which kinds of open matters might it touch? This chain splits that reading into four steps in one Claude chat. Claude first summarizes the update in the update's own words with paragraph references, then lists the questions the update raises for your own research, then compares a list of matter type codes with the update's text, then drafts a one-paragraph note to staff. You check the output after every step against the source. The outcome it serves is a first read that is anchored to paragraphs you have verified, so your own reading and research start from a map and not from a blank page.

The chain never tells you what the update means. You read the source and decide what it means. The legal research happens in your own legal research service.

Prerequisites

• A Claude account. The paid plan is Pro ($20/month), and that is the whole ongoing cost of this build. Check claude.ai for what the free tier covers.
• A published policy update saved from the agency's own website, as text you can paste. Use the update and nothing else. The guide names no update and quotes none.
• A short list of your open matter type codes, such as FB-ADJ, EB-PET, NATZ and REMOVAL. Those four are invented examples. Use your own codes.

The Concept

Think of a research assistant who reads a document you hand over, then highlights passages and labels them with page numbers, but is not allowed to tell you what any of it means. You still read every highlighted passage. The assistant's value is the map: where to look first.

Claude plays that assistant, and every step ends at a checkpoint where you act as the reviewer. The model is not a source of law. It does not know the current rules, and it cannot tell you what an agency does in practice. Anything it adds beyond the pasted text is unverified until you read it on the primary source, and an AI-produced citation can be invented. Lawyers have been sanctioned for filing citations that an AI made up, so treat every reference outside the pasted update as a lead to check in your research service and never as authority.

Data caution. The chain works on a published document and on matter type codes only. Client names, A-numbers, receipt numbers, matter facts and any individual's immigration history stay in your firm's approved systems and never go into this chat. Before you paste anything non-public, check Claude's data and privacy settings for whether chats are used to train models and how long they are kept, and check whether your firm approved the tool for client data under your confidentiality duties and your state bar's rules on AI use. A published update and a list of type codes do not need that approval, but your firm may still have a policy on which AI tools you use.

Build It Step by Step

Part 1: Prepare the two inputs

1. Open the agency's own page for the update and copy the text into a plain document. Keep the paragraph or section numbering the update uses. If the update has no numbering, add your own markers such as [P1], [P2] and so on before you paste, so the references mean something.
2. Make a second plain list of your matter type codes, one per line. Use type codes only, never matter numbers.
3. Open a new chat in Claude.

Part 2: Prompt 1, a plain summary in the update's own words

Paste this prompt. Then replace the bracketed line with the update text.

Below is a published policy update. Write a plain summary of it as a numbered list. Each item must use the update's own wording as closely as possible and must end with the section or paragraph reference where that wording appears, for example (P3). Do not add any rule, date, fee, form, requirement or reason that is not in the pasted text. Do not say what the update means for any person or case. If the update does not say something, write "not stated". Begin the answer with the label "Draft for attorney review".

Update text:
[PASTE THE UPDATE TEXT]

Checkpoint 1. Open the update next to the answer. For each item, find the cited paragraph and read it. Confirm the wording matches, the reference points to the right place and nothing was added. Delete any item with a reference you cannot find. Look for sentences that sound like a rule but have no matching paragraph. Anything the answer states with no paragraph behind it is out. The update is the only source in this chat.

Part 3: Prompt 2, questions for your own research

In the same chat, send:

Using only the numbered summary above and the pasted update, write a list of questions this update raises for the attorney's own research. Word every item as a question and cite the paragraph that prompts it. Do not answer any question. Do not suggest what the answer is, and do not name any statute, regulation, case or policy that is not in the pasted text. Keep the label "Draft for attorney review".

Checkpoint 2. Read the questions against the update. Each should be a question, not a statement in disguise. Remove any question that hides an answer, such as one that begins "Given that the update now requires". Take the questions that remain to your legal research service and the agency's current instructions, and answer them there. Write down where each answer came from. Claude has not answered anything, and it should not.

Part 4: Prompt 3, which matter type codes the text mentions or might touch

Paste your list of codes in place of the bracketed line.

Here is a list of matter type codes. For each code, say whether the pasted update's text mentions it, might relate to it, or does not appear to mention it, and quote the passage with its paragraph reference for every code you place in the first two groups. If you cannot quote a passage, put the code in the third group. Do not explain what a code means beyond the code itself, do not say what the update means for any case, and do not state any rule. Present the answer as three lists for the attorney to review. Keep the label "Draft for attorney review".

Matter type codes:
[PASTE THE LIST OF CODES]

Checkpoint 3. Check each quoted passage in the update. A code in the first two lists with a quote that is not in the update, or a quote that does not mention anything related, belongs in the third list. Then check the third list the other way. Search the update text for each code's subject words yourself, because the model can miss a mention. The result is a list for you to review, not a conclusion about any client. Open the files for the matter types you decide to look at, using your own systems.

Part 5: Prompt 4, a note to staff

Write one short paragraph for an internal note to the firm's staff. It must say that the attorney is reviewing a published policy update, list the matter type codes the attorney has chosen to review (listed below), and say that the note is not an instruction about any matter and that the attorney will explain if anything changes. Do not describe the update's content, do not give any instruction about any client's case and do not give legal advice. Label it "Draft for attorney review".

Codes the attorney is reviewing:
[LIST THE CODES YOU CHOSE]

Checkpoint 4. Read the paragraph. It should name the update only as "a published policy update" or by the title you give. It should list your chosen codes and nothing more. Remove any sentence that tells staff what the update means or what to do on a matter. Send it yourself, in your own words if you prefer. Staff do not interpret the update for clients.

Real Example: A Sample Run With Invented Codes

Setup: A saved update (not named here), numbered by paragraph, and the invented codes FB-ADJ, EB-PET, NATZ and REMOVAL. Input: Prompts 1 to 4 in one chat, with a checkpoint after each. Output, step by step (an illustration of the shape of a reply, not a guaranteed result):

• Prompt 1 returns items such as "[Quoted wording from the update] (P2)" and "not stated" for anything the update leaves out.
• Prompt 2 returns questions such as "What does the wording in P4 cover, and is the current form instruction consistent with it? (P4)". You take each to your research service.
• Prompt 3 returns three lists. The first two carry a quoted passage and a reference for each code, for example "NATZ: [quoted passage] (P6)". The third holds the codes the update does not appear to mention.
• Prompt 4 returns a paragraph that says the attorney is reviewing a published policy update, lists the codes you chose, and says it is not an instruction about any matter.

What you do next: You read the cited paragraphs yourself. You run the research questions in your service, open the matter types you flagged and decide what, if anything, to tell anyone. The model does none of that. The wording in the illustration is a placeholder, and the real answers depend on the real update.

What to Do When It Breaks

• A quote does not appear in the update → Delete it and rerun the prompt with the sentence "Every quotation must be an exact copy of text I pasted." The pasted text is the only source.
• Claude states a rule, a date, a fee or a case name that is not in the update → Delete it. It is unverified until you read it on a primary source in your research service.
• A paragraph reference points to the wrong place → Check whether the update's numbering was copied correctly. Add your own [P1] markers and run Prompt 1 again.
• The summary drifts into what the update means for clients → Remove that text and repeat Prompt 1 with a reminder that no meaning or effect is to be stated.
• A matter type code is missing from the lists → Add it by hand after checking the update yourself. The model can drop an item without saying so.
• A long chat starts to drift → Start a new chat and paste the update and the Prompt 1 summary you verified. Then continue from Prompt 2.
• The update changes after you saved it → The chat has the old text. Save the new version and begin again from Prompt 1.

Variations

• Simpler version: Run Prompt 1 only and read the cited paragraphs. Then do the matter review yourself.
• Extended version: Raise the aim from a reading to a review that is tracked. Keep the verified summary and the questions with the matter type list as a short file note, and revisit it when your research service shows new material. Keep to the rule about published documents and codes only.

What to Do Next

• This week: Run the chain once on a published update you have already read, and compare the map with your own notes.
• This month: Save Prompts 1 to 3 in a Claude Project so each new update starts from the same instructions.
• Advanced: Pair it with the Daily Notice Response Watch so that any matter you flag gets a tracked date.

Advanced guide for immigration attorney professionals. These techniques may need a paid Claude plan. Menus and settings change, so check Claude's current screens and help pages.