How to Write Effective Prompts for Eve

Overview

Most of what people call "Eve isn't giving me good answers" is a prompting problem, not a product problem. The gap between a mediocre Eve response and an excellent one is almost always something you can control: what you asked for, what you gave her to work with, and whether you told her what to do when the answer isn't there.

This guide covers both halves of that. The first section explains how Eve actually processes what you send her — the mental model. The second half turns that model into concrete techniques with Eve-specific examples.

Who this is for: anyone using Talk to Eve, skills, or Draft from Blueprint. The principles apply everywhere Eve takes a written instruction.

Part 1: How Eve reads your prompt

She's a brilliant new colleague with zero case history

The single most useful frame: Eve is an extremely capable attorney who joined your firm this morning. She has read every document you've put in front of her and nothing else. She doesn't know your client, your opposing counsel, your judge's preferences, your firm's shorthand, or what you decided in last Tuesday's case strategy meeting.

This is why "check that email about the incident" fails and "review the email from Jane Smith dated March 15, 2023, titled 'Incident Report: Warehouse'" succeeds. It isn't that Eve is being difficult. It's that she has forty-two emails and no way to know which one you mean.

The colleague test: before you send a prompt, ask whether a smart new hire with no case background could execute it correctly. If they'd have to come back and ask you a question, the prompt may benefit from clarification or expanding on your goals.

Context is a budget, not a bucket

There's a tempting assumption that giving Eve more documents produces better answers. It doesn't, past a point. Every document you make available is another place Eve has to look, and irrelevant material competes for attention with the material that matters.

Think of your context selection as an assignment, not an archive. If you hand a paralegal a single deposition and ask them to summarize it, you'll get a tight summary. If you hand them a banker's box and ask them to summarize the deposition inside it, you'll get a less focused summary that will take longer to produce.

The practical rule: match context scope to question scope.

Here are some examples:

Eve is agreeable by default

Eve is built to be helpful, and helpfulness has a failure mode: when you ask her to find something, she is inclined to find it. Phrase a request as "find all instances of HR complaints" and you've quietly told her that instances exist and her job is to produce them. If the documents contain none, a helpful assistant is under pressure to produce something — a marginal match presented with more confidence than it deserves.

The fix is one clause long, and it's covered in the techniques below. But understand the mechanism, because it explains a whole category of Eve errors that otherwise look random.

She interprets literally, and format is an instruction

Eve does what you say, not what you meant. This cuts both ways: sloppy phrasing produces sloppy output, but precise phrasing produces precise output. "Summarize the medical records" is an invitation for Eve to decide what a summary is. "Summarize the medical records in under 200 words, focused only on treatment received after the incident date" leaves nothing to interpretation.

The same applies to structure. If you don't specify a format, Eve picks one — and it may not be the same one she picked last time you asked a similar question. Format instructions aren't cosmetic; they're how you get consistency across a caseload.

Part 2: The anatomy of a strong prompt

Not every prompt needs every part. Simple lookups don't need scaffolding. But when a request matters — a billing breakdown, a claims analysis, anything you'll rely on — these five components are the checklist.

1. The task — the specific, concrete thing you want Eve to do. Lead with it. One deliverable, stated plainly, so Eve knows what she's being asked before she reads anything else.

2. Objective — why you need it and what it's for. This is the piece people skip, and it's the one that does the most work. Eve uses the purpose behind a request to make better judgment calls on the hundred small decisions you didn't specify. "Summarize this deposition" and "summarize this deposition because I'm preparing to cross-examine him" produce very different summaries.

3. Context — the case background Eve needs and can't get from the documents alone. Case type, procedural posture, what you've already established, who the parties are, what matters strategically. Remember she has no case history beyond what's in front of her.

4. Format — how you want the answer returned. Table, bullets by category, narrative prose, a specific set of fields per entry. If you don't say, Eve picks — and she may not pick the same way twice.

5. Additional instructions — the edge cases, exceptions, and special handling. This is where you put anything that starts with "and if…": how to handle ambiguity, what to do when something isn't found, what to exclude, what to flag rather than assume. It's the catch-all for everything a careful assignment would cover that the other four sections don't.

The template

[Task] — Do X.

[Objective] — I need this for Y.

[Context] — Here's what you should know about the case.

[Format] — Return it like this.

[Additional instructions] — Handle these situations this way.

You don't need to label the sections when you write a prompt — the labels below are just to show you the parts. Eve reads it fine as ordinary paragraphs.

Worked example

Weak:

Find the discrimination evidence.

Strong:

Identify every instance in these documents where the plaintiff reported discriminatory treatment to management or HR [task].

I'm building the factual record for a summary judgment opposition, so I need this to be exhaustive and defensible — every instance I can point to, with a citation I can verify [objective].

This is a wrongful termination case with an employment discrimination claim. The HR records span 2019–2023. The plaintiff alleges she reported discriminatory treatment repeatedly before being terminated in November 2023 [context].

Give me a table with one row per instance: date, who she reported to, the medium (email, meeting, formal complaint), a one-sentence description, and the source document with page number [format].

If an instance is ambiguous — where it's unclear whether she was describing discrimination or a general workplace complaint — put it in a separate section below the main table rather than mixing it in. If there are no instances at all, tell me that directly instead of stretching to fill the table [additional instructions].

The second prompt is longer, and that's the point. You spent ninety seconds writing it and saved yourself two rounds of correction plus a false positive you might not have caught.

Why this order matters

Task first means Eve knows what she's doing before she processes why or under what constraints — the same reason you'd tell a paralegal "pull every discrimination complaint" before launching into the background. Objective second because the why colors how she reads everything after it. Context, format, and instructions then narrow the work from broad to specific.

If you've used Draft from Blueprint or the Discovery Response skill, this structure will feel familiar — those skills already split the work into what you're producing, what documents inform it, and an instructions field for the specific tweaks. This is the same shape, written out longhand in chat.

Part 3: Core techniques

1. Be specific about documents, dates, parties, and facts

Name things. Use exact file names, full party names, specific dates, and page ranges when you have them.

• Instead of "what does the evidence say?" → "What does John Smith's deposition transcript say about the termination meeting on March 15th?"

• Instead of "check the employment agreement" → "Looking at pages 15–20 of the employment agreement, what are the employee's non-compete obligations?"

• Instead of "the medical stuff from last year" → "Dr. Ramirez's treatment notes from January through June 2024"

You'll know it's working when Eve's citations point to exactly the documents you named.

2. Explain the why, not just the what

This is the most underused technique in the list and one of the highest-leverage. Telling Eve the purpose behind a request lets her make sensible decisions in the gaps.

• "Summarize this deposition" → a generic summary.

• "Summarize this deposition. I'm preparing to cross-examine this witness and I need to know where his account contradicts the police report" → a summary organized around contradictions, with the details that matter for impeachment surfaced first.

Same document, same nominal task, completely different usefulness.

3. Scope your context deliberately

Before you send a meaningful prompt, look at your Select context panel and ask whether the file list matches the question. Deselect what isn't relevant.

This is the highest-effort-to-payoff ratio move in this entire guide, and it's the one people skip because it happens outside the text box. Narrowing context reduces noise, sharpens citations, and often speeds up the response.

One caveat worth internalizing: narrowing context is not the same as narrowing the question. If you deselect the police report and then ask about liability, Eve will answer confidently from what's left. She's working from the assignment you gave her — so make sure the assignment includes what the question actually requires.

4. Give Eve an out

Add an explicit negative-result clause to any request that asks Eve to find something.

• "Find all instances of HR complaints in these documents. If there are none, tell me there aren't any."

• "Identify any prior accidents referenced in the records. If the records don't mention any, say so rather than inferring from context."

• "List the treatment gaps longer than 30 days. If treatment was continuous, state that."

Success looks like Eve plainly saying "I did not find any" — which is a real answer and often a valuable one.

A close cousin: give her permission to flag uncertainty. "If you're not confident about a date because the record is illegible, mark it as uncertain instead of picking the most likely reading."

5. Specify your output format

Tell Eve exactly how you want information structured. This makes responses easier to review, and it makes results comparable across matters when you reuse the prompt.

Respond with a section header for each type of claim, and under each, a bulleted list of the potential claims of that type with a one-sentence description of each. End with a short paragraph on which claim you'd rank strongest and why.

Two refinements worth knowing:

Ask for what you want, not what you don't. "Write this in flowing prose paragraphs" works better than "don't use bullet points." Negative instructions require Eve to hold the forbidden thing in mind; positive instructions just describe the target.

Your prompt's style influences her output's style. Write a terse, clipped prompt and you'll tend to get terse output. Write a formal one and you'll get formal output. If you want a particular register, model it.

6. Show her an example

If you have a format you want matched, paste one in. A single worked example does more to control structure than three sentences describing that structure — and if you have two or three examples covering different shapes, better still.

Here's how I formatted the visit chronology for our last case:

3/14/2024 — Dr. Patel, Northside Orthopedics — Initial consult; complaints of lower back pain radiating to left leg; ordered MRI. [Bates 00412]

Build the chronology for this matter's records in exactly that format.

This is also how you bring firm conventions to Eve without writing a style guide.

7. Break complex work into chunks

If your request contains the word "and" three times, it's probably three prompts.

Instead of: "Review all the depositions, create a timeline, identify contradictions, and draft a witness list."

Do this:

1. "Create a chronological timeline of events based on all deposition testimony."

2. (review the timeline) "Now identify where the witnesses' accounts contradict each other on these events."

3. (review the contradictions) "Draft a witness list ranked by how central each witness is to the disputed facts."

Two things happen. Each response is better, because Eve puts full attention on one deliverable. And you catch errors at step one instead of discovering at step three that the whole chain was built on a bad timeline.

This is also a quality control pattern in its own right: draft → review → refine, as separate prompts, beats asking for a polished final product in one shot.

8. Ask Eve how to ask

If you're not sure how to frame something, ask her directly.

• "I'm trying to extract a timeline of key events from these medical records. How should I ask you for that?"

• "What's the best way to request a case strategy memo from you? What information would you need from me?"

Eve will usually come back with a prompt template you can edit and reuse. This is the fastest way for a new user to get good, usable prompts without needing to take classes or study prompt creation.

9. Verify, always

Eve is built to support your judgment, not replace it. Verification isn't a trust issue; it's the job.

Use the citations. Eve's responses include citation badges and highlights linking back to the exact source passages. Click them. This is faster than re-reading the document and it's the single best habit to build.

Ask her to check her own work. These prompts genuinely catch errors:

• "Can you check your work?"

• "Provide direct quotes from the documents to support each of these claims."

• "Review your timeline and confirm each date appears in the source materials."

• "Explain how you reached that conclusion."

Push back when something's off. If a response is wrong or incomplete, say specifically what's wrong. "That's not right" gives Eve nothing. "The third entry cites the wrong provider — that visit was with Dr. Chen, not Dr. Ramirez. Re-check the provider on each entry against the letterhead in the source records" gives her something to act on.

10. A note on "think step by step"

Older prompting guidance — including our own earlier articles — told you to add "think through your reasoning step by step before answering." That advice is largely obsolete. Eve reasons through multi-step problems on her own now; the instruction rarely improves quality the way it did a couple of years ago.

It's still useful for a different reason: auditability. When you need to see how Eve reached a conclusion — because you're going to rely on it, or because you suspect it's wrong — asking her to show her reasoning gives you something to check. Use it when you need transparency, not as a default quality booster.

Part 4: Prompting when Research is on

Eve's Research capability changes how you should prompt, and it's the biggest thing missing from older versions of this guide. Eve is no longer limited to your matter documents.

Trusted Web Search lets Eve pull current public information — regulatory updates, government benefit rates, medical literature, employer background — with full citations. Case Law Search brings case law and court opinion research into the same chat, with Bluebook citations, treatment status, and in-app opinion verification. Both require firm admin enablement in Settings > Research, and both have toggles in chat.

Three prompting adjustments:

Say where the answer should come from. When both documents and web sources are in play, be explicit about which you want. "Based only on the medical records in this matter, what treatment did the plaintiff receive?" is a different question from "What is the standard treatment protocol for this condition?" — and mixing them without saying so gets you a blended answer that's hard to verify.

Pick the collection that fits the task. Statutes & Regulations for regulatory lookups, Medical Research for clinical questions, Business Filings for entity research, All Web for broad research. Choosing the collection is itself a form of prompting — it narrows the source universe before Eve reads a word, exactly like narrowing your context selection does for documents.

Verify web and case law citations the same way you verify document citations. Every web result carries a URL, title, publish date, and excerpt. Every case law citation opens inside Eve so you can read the opinion and check the treatment badge. Use them.

Part 5: Prompt to Eve's strengths

Structuring requests around what Eve does well is itself a prompting technique.

She's strong at: document analysis and information extraction; pattern recognition across large record sets; chronology and timeline construction; legal reasoning grounded in documents you've provided; drafting in a learned style; synthesizing case law and public sources when Research is enabled.

Know the limits:

• Precise math. Like any language model, Eve is unreliable at complex calculations. Have her extract and organize the numbers; verify the arithmetic yourself.

• Images without text. Eve reads text from images — photographs of documents, screenshots, scanned records — via OCR. She cannot describe the visual content of a photo. A picture of a totaled vehicle will never produce a meaningful description of the damage. If the information exists only as an image and not as text, it isn't available to her.

• Anything outside her context. Eve answers from your selected documents plus, when enabled, the web and case law sources you've allowed. She doesn't know what's in your CMS, your inbox, or your head unless you've told her.

Part 6: Troubleshooting

Quick reference

Before you hit send on anything that matters:

• Would a smart new hire with no case background understand this?

• Have I named specific documents, dates, and parties?

• Did I say why I need this, not just what I want?

• Is my context selection matched to the question?

• Is this one deliverable, or three?

• Did I specify the output format?

• Did I give Eve an out?

• After the answer: did I click the citations?