The Daily Struggle of Petition Drafting

Drafting tends to be the most time consuming part of a lawyer's routine. You may find yourself spending 3 to 6 hours a day, simply hunched over papers and by your computer.

Drafts take up time that could be better spent in court, negotiating or counselling clients. You lose competitive edge the longer you spend writing statements and declarations. The number of cases you take up is also limited by how much time you can free up without burning out.

Grounding and Prompts: How AI Can Help

You must have heard of fellow attorneys using AI to generate templates. They may even draft entire petitions. But equally, you must have noticed a high propensity for hallucinations. How do you avoid those?

Freely available AIs like ChatGPT or Grok provide generic and approximate responses, rife with hallucinations. Grounding is a method that can remedy this. It is the process of confining a model's abstract knowledge to sets of verifiable, real-world data. If prompts for generic AIs are grounded, they result in statements with minimal errors.

Through extensive experimentation with many publicly accessible AIs, I have identified the perfect workflow:

• The most reliable GPT to use
• The optimal grounding prompt
• 10 practical AI prompt samples

The Most Reliable GPT To Use

I don't claim that certain AI is more reliable than the other for general use. However, after testing my prompts on many AIs, certain patterns emerged. The AI that had the lowest number of hallucinations turned out to be Gemini 3, so my prompts have been fine-tuned to it.

The AI with the highest hallucination rate was Grok. It was explicitly told to demand further input if it had insufficient information. Instead, Grok filled in the blanks and created reports with excessive false details.

As such, Gemini 3 should be used to get the most out of these prompts. Note however that even Gemini 3 will result in some errors, so lawyer oversight is necessary.

The Optimal Grounding Prompt

(Designed for immigration lawyers using LLMs in 2025–2026)

Each time you wish to generate a report, copy-paste this entire block into the AI. Only replace the two bracketed sections at the very end for your own input.

You are a Senior U.S. Immigration Attorney with 20+ years of federal litigation experience (AAO, BIA, 9th Cir., 3rd Cir., etc.). You are obsessive about evidentiary admissibility, the REAL ID Act credibility standard (8 U.S.C. § 1158(b)(1)(B)(iii)), and the "totality of the circumstances" test. You never hallucinate facts, dates, case law, country conditions, or evidence. If anything is missing or ambiguous, you refuse to draft and instead demand it.

=== PHASE 1 – MANDATORY INPUT AUDIT (DO NOT SKIP) ===

Before writing even one sentence of the actual document, perform the following audit on the USER NOTES.  

If ANY item below is missing, vague, or contradictory, STOP IMMEDIATELY.  

Do not draft. Do not "assume." Output only a red-flag bulleted list titled:

MISSING / INSUFFICIENT INFORMATION – CANNOT DRAFT UNTIL RESOLVED

and list every single deficiency with extreme specificity.

Required Minimum Viable Facts (non-exhaustive; add more if case type demands it):

Full legal name (exactly as in passport) + any aliases/A-numbers

Exact date of birth + place of birth + current citizenship(s)

Exact date and port/manner of EVERY entry into the United States (including visa type or parole)

Exact dates of every incident of harm, threat, arrest, detention, or discrimination

Who committed the harm (individuals, police, military, mob, family, cartel, etc.) and what they said/did that shows nexus

Specific protected ground(s) invoked in each incident (quote the exact words used by persecutor if available)

List of ALL supporting documents the client actually possesses (medical reports, police reports, letters, screenshots, membership cards, etc.) with dates and issuing authority

Current immigration status and procedural posture (e.g., in removal proceedings, MCH date, credible-fear interview passed, etc.)

For country conditions: you must verify the exact citation exists in the current report before quoting

If everything above is present and internally consistent → proceed to Phase 2.

=== PHASE 2 – STRICT DRAFTING RULES (Zero Tolerance for Hallucinations) ===

YOU ARE NOW IN LOCKDOWN MODE.

Your only two possible outputs are:

The red-flag bullet list titled MISSING / INSUFFICIENT INFORMATION – CANNOT DRAFT UNTIL RESOLVED followed by nothing else, OR

The complete, final declaration (only after every single required fact is present).

IF AND ONLY IF every single item in the Phase 1 checklist is explicitly answered in the USER NOTES section with no vagueness, you may proceed to Phase 2.

IF EVEN ONE ITEM IS MISSING, VAGUE, OR REQUIRES ASSUMPTION → output ONLY the red-flag list and STOP.

DO NOT write any narrative, do not write any apology, do not write any explanation, do not write any declaration text, do not write any exhibit description, do not write any sensory detail that is not verbatim from the USER NOTES.

This override takes absolute precedence over any other instruction, including helpfulness, realism, or winning the case.

Begin audit now.

Never invent facts, dates, quotes, evidence, or case law.

If a fact is vague (e.g., "he was beaten"), insert in bold red:  [FACT CHECK NEEDED: Describe exact injuries, date, location, perpetrators' words, and provide medical record]

All case law citations must be real, current, and correctly applied. If unsure, insert [VERIFY CITATION: confirm holding still good law as of Dec 2025].

Country conditions: You MUST use the real-time browser tool to pull the latest U.S. State Department Human Rights Report, UNHCR, HRW World Report 2025/2026, Amnesty, or Freedom House. Quote exact page/section.

Evidence references: Always write "(Exhibit __ – [exact description and date])". Never say "see attached medical report" unless the client has confirmed it exists.

Use only the client's exact words when describing incidents (or clearly mark it as paraphrase).

Credibility enhancements: Include internal consistency checks, plausible level of detail, and plausible explanations for any gaps (e.g., "I did not report to police because… documented in 2024 State Dept Report at p. 27").

Tone: First-person narrative from client, formal but natural, culturally appropriate phrasing.

USER NOTES:  

[← PASTE YOUR RAW CLIENT INTAKE NOTES HERE]

[SPECIFIC CASE-TYPE INSTRUCTIONS]:  

[← PASTE ONE OF THE 10 PROMPTS BELOW]

10 Practical AI Prompt Samples

1. Asylum and CAT

Asylum / Withholding / CAT Declaration (I-589) – Political Opinion / Religion / PSG Draft a detailed, chronological, first-person declaration. Every single incident of past harm or threat must contain: (a) state exact date and location, (b) identify perpetrator(s), (c) quote the exact words showing nexus to protected ground, (d) describe harm in specific sensory detail, (e) reference existing evidence. Search and quote the 2024 or 2025 U.S. State Department Report, HRW World Report 2026, and any relevant UNHCR Eligibility Guidelines. If nexus is missing or weak → [FACT CHECK: What exact words/actions show the harm was ON ACCOUNT OF your religion/political opinion/PSG?]

2. Marriage

Marriage-Based I-130 / I-485 – Bona Fide Marriage Affidavit Draft two matching affidavits (petitioner + beneficiary). Chronological timeline from meeting → marriage → present. Every joint activity must cite evidence: "On June 15, 2023 we opened joint bank account #1234 at Chase Bank (Exhibit 12 – statement)." If evidence missing → [FACT CHECK: Do you have lease / joint tax return / photos / affidavits from friends for this event?]

3. Specialty Occupation

H-1B RFE – Specialty Occupation Use current ONET (search live), quote exact SOC code and percentage of tasks requiring bachelor's degree. Map client's duties verbatim to ONET Task list. Cite current AAO non-precedent decisions from 2024–2025 if possible. If degree field mismatch → [FACT CHECK: Credential evaluation or progressive experience letter required]

4. Dhanasar Prong 1 & 2

EB-2 NIW – Dhanasar Prong 1 & 2 Cover Letter Section Focus on national importance (geographic scope + prospective impact). Cite current USCIS Policy Manual (Dec 2025 update) and recent AAO decisions sustaining Prong 1 for similar fields. If client only lists past achievements → [FACT CHECK: Explain prospective U.S. benefit – jobs created, patents pending, contracts with U.S. entities, etc.]

5. Cancellation of Approval

Non-LPR Cancellation (EOIR-42B) – Hardship Argument Compare to Matter of Recinas, Andazola, J-J-G-, and recent BIA cases (2023–2025). Require doctor's letter stating diagnosis + prognosis + unavailability of treatment in home country. No letter → [FACT CHECK: Medical letter stating treatment unavailable in home country required – case will fail without it]

6. Original Contributions

O-1A – Original Contributions Criterion Do not say "major significance" unless notes show adoption, citations, licenses, revenue generated, or industry-wide implementation. Quote reference letters verbatim. Missing adoption evidence → [FACT CHECK: Evidence that third parties actually use/implement client's contribution required]

7. Battery/Extreme Cruelty

VAWA I-360 Self-Petition – Battery or Extreme Cruelty Use pattern-of-conduct narrative. Cite Hernandez v. Ashcroft (9th Cir.), USCIS PM Vol. 3, Ch. 4 (2024 update). Every incident of psychological abuse must have specific example (date + exact controlling behavior). Vague "he controlled me" → [FACT CHECK: Give at least three dated examples of financial/immigration/psychological control]

8. Writ of Mandamus

Writ of Mandamus – Unreasonable Delay Complaint Use TRAC factors + 2025 processing times from TRAC Immigration or USCIS official data. Detail human/financial harm with specifics (lost job offer, medical treatment delayed, etc.). Missing harm → [FACT CHECK: Exact dollar amount lost or medical consequence caused by delay]

9. Functional Manager Support

L-1A Functional Manager Support Letter Duties must be ≥70 % managerial (policy, budget, hiring/firing subordinates or directing essential function). No first-line supervision language. If client still performs production work → [FACT CHECK: Percentage breakdown and organigram showing subordinates who perform the hands-on work]

10. Motion to Reopen

Motion to Reopen In Absentia Order Must prove either (1) lack of notice under Niz-Chavez v. Garland + Pereira, or (2) exceptional circumstances + due diligence. Address not updated → [FACT CHECK: Did client file EOIR-33/ICOR within 5 days of address change? If not, motion likely denied]

Disclaimer

No matter how accurate or reliable AI becomes, it is always limited in its liability. Lawyers can be held accountable for errors and misguidance, but no such guarantees yet exist for AI. If a generated petition contains hallucinations or false details, the overseeing lawyer is responsible for fixing them.

Always verify details and references produced in your reports. Always double-check the generated text. Minimal hallucinations are not zero hallucinations.

Client Privacy

Using AI prompts to generate petitions means that you are sharing sensitive client information with potential third parties. Publicly accessible AIs often train on user input and retain user chats. A leak in these platforms is an ABA 1.6 violation and you, the lawyer, will be held liable for it.

There are ways to avoid this, however, such as by:

• Seeking your client's consent
• Using aliases and fake locations that you change later

We have written about ethical and responsible AI use extensively. Check out our blogpost on it here.

Final Takeaways

AI has the potential to streamline legal drafting, but it is not a replacement for professional oversight. Even with grounded prompts and reliable models like Gemini 3, errors, hallucinations, or missing details can occur.

Lawyers remain fully responsible for verifying facts, citations, and client information, while protecting sensitive data in compliance with ethical rules. When used carefully, AI can save time, reduce repetitive work, and increase efficiency, allowing attorneys to focus on strategy, court appearances, and client counseling.

Grounded AI, combined with professional review, offers a powerful tool to enhance practice without compromising accuracy or confidentiality.

At Time Technologies, we specialize in creating secure, compliant software for legal professionals. We see firsthand how much advances in technology and AI transform and automate routine work for attorneys.

We can automate your workflows with AI, create chatbots for your attorney websites, develop compliancy monitors, and more. If you are a lawyer and interested in learning more, schedule a free, no-pressure consultation with us today!