Deep Dive Question Helper
4 min
Writing effective Deep Dive Questions
Deep Dive helps you turn a patent set into a structured analysis of a technology. You can use it to understand which approaches companies are developing, what problems their inventions address, and what evidence they provide for the benefits they describe.
The questions you ask determine what Deep Dive extracts and how useful the results are for your research.
💡 Be specific about what you want to learn from each patent and how the answer should be presented.
You do not need to write every question from scratch. The Prompt Helper at the end of this page lets you use Ask Me Anything to turn your research context into suggested questions and a complete prompt for Deep Dive.
How to use the Prompt Helper
- Copy the Prompt Helper below into Ask Me Anything. Under MY RESEARCH CONTEXT, describe what your patent set covers, what you have found so far, and what you want to learn.
- Review the suggested columns. Keep, change, drop, or add columns, then ask for the final prompt.
- Paste the final prompt into Deep Dive. Read the agent’s plan before approving the run.
Example Research Context:
"My set covers water-based pigment inks for production inkjet printing. I have found several binder and fixation approaches, but I do not yet understand how they address adhesion and drying. I want to map the main approaches by company and identify which reported benefits are supported by test results in the patents."
Prompt Helper
You help me write a prompt for a patent analysis tool called Deep Dive. You do not run it: I paste your prompt into Deep Dive.
How Deep Dive works: I have a set of patents I already gathered on one technology. Deep Dive is an AI agent that reads each patent in my set one by one and builds a table: each question is a column, each patent a row. At the end it answers summary questions across the whole table. Before it runs, it shows me its plan, including the instruction it will use for each column, and waits for my approval.
HOW WE WORK
1. Read my research context. Never ask how many patents are in my set. If my context does not say what technology the set covers or what I want to learn, ask me ONE short question about it. After my answer, or if my context is already clear, go straight to step 2. If my context is broad, make reasonable choices for the field and say them in one line.
2. Suggest 5 columns, each already written as a full column prompt (see HOW TO WRITE A COLUMN), using exactly this layout and nothing more:
Column 1: [short label] Prompt: [the full column prompt]
Column 2: [short label] Prompt: [the full column prompt]
(and so on)
If a column builds on others, add one line under it: "Builds on Columns X and Y." Then ask: "Which columns do you want to keep, change, drop or add?" and stop.
3. When I confirm, write the full Deep Dive prompt as one block I can paste, with nothing else inside it (see HOW TO WRITE THE FULL PROMPT). Below it, remind me to read the agent's plan before approving.
4. Keep going with me: more columns, other columns, changes. Each time, give me the full prompt again.
WHAT TO ASK
I already have numbers on how fast this technology is improving, calculated from patent citation data. So do not use columns to measure speed or count activity. Aim the columns at what only the patent text can tell me: which materials, routes or configurations sit inside the technology; which approach each patent takes; what each company is actually building; which problem each invention solves and whether the benefit is shown or only claimed; and what the answers add up to (for example a map of approaches built from other columns). Mix the columns my goal clearly needs with one or two I am probably not thinking to ask, the ones that could surface something unexpected. Keep them grounded in what a patent can actually say.
HOW TO WRITE A COLUMN
Every column prompt is complete on its own, because Deep Dive applies it to each patent separately. It states:
- what to read ("Read this patent's abstract, claims and background section"; for presence questions, "Read the independent claims")
- one thing to answer, in the field's own words
- the exact answer shape and a hard limit:
- for a category: name the categories that fit this field, plus Other ("Answer with exactly one of: chemical (solvolysis), chemical (pyrolysis), enzymatic, mechanical, Other.")
- for yes or no: "Answer Yes or No."
- for an extracted item: what to list and how many ("List at most 3 catalyst materials named, separated by semicolons.")
- for a short answer: "One sentence of at most 25 words."
- the exact text when there is no answer ("If the patent does not state it, write exactly: Not disclosed.")
- "Start with Stated: or Implied:" where the answer could be either
- "Use the patent text only."
Example: Column 1: Problem addressed Prompt: Read this patent's abstract, claims and background section. In one sentence of at most 25 words, state the technical problem or shortcoming of existing solutions that this invention sets out to solve. Describe the problem, not the solution. If the patent does not state or clearly imply a problem, write exactly: Not disclosed. Use the patent text only.
Rules:
- Every column must be answerable from one patent alone. Counting, ranking and comparing patents belong in the summary.
- One thing per column.
- No columns for owner or dates: the table already shows each patent's owner and dates.
- No instructions on slicing or filtering the set: the set is already built and filtered before Deep Dive reads it.
- Keep it technical: no company names of mine, no commercial story, and never word a column to confirm what I hope to find.
- If I give you a reference to compare against, describe it by its technical features; never invent one.
HOW TO WRITE THE FULL PROMPT
1. Goal, at the top: one or two sentences on what the set is, what I want to learn, and what the end product is. Then: "Present all column proposals below together for my approval. Use the column prompts exactly as written. Once I approve them, generate them, then answer the summary questions in order without waiting for further prompts."
2. The columns I agreed, exactly as agreed.
3. Summary questions, numbered. The first is always "Start by stating how many patents were analysed." Ask for counts, splits (by year or company where they matter) and 2 or 3 example publication numbers behind each statement. Ask it to say where evidence is thin. The last question asks for the written answer the goal named.
If a column builds on others, split the prompt into PHASE 1 (columns answered from one patent), PHASE 2 (columns built on phase 1, naming which ones they use) and PHASE 3 (summary). If phase 2 builds categories, ask the agent to show me the categories, each with a one-line definition, plus "Other", and to wait for my approval before labelling.
MY RESEARCH CONTEXT: [describe your research here]