How is a prompt written?

The task is fixed: a draft statement of defence for the defendant client in a goods-price claim. Facts and figures are fictional. The full text of all three versions is on the page. The numbered marks under each output show which sentence must be verified and why. The number of marks does not fall; their kind changes: with a weak prompt they are errors and inventions, with a strong prompt they are placeholders left on purpose to fill in or verify.

• Error or invention
• To fill in or verify

Outputs are written as examples; they are not the actual response of an AI model.

Write me a statement of defence. Our client is the defendant, the claimant filed a debt claim, and we say we paid it.

You are a lawyer's assistant. A goods-price claim of TRY 250,000 was filed against our client (the defendant). The client says it paid TRY 150,000 on 12.03.2026 and has the receipt. Write a statement of defence in a formal register.

ROLE: You are a drafting assistant for Turkish law; you give no final conclusions and mark unknowns with [ ].FACTS: Claimant [CLAIMANT] sued our client [DEFENDANT] for a TRY 250,000 goods-price claim. The client wired TRY 150,000 on 12.03.2026 (receipt available). The client says it reported a delivery defect for the rest; we hold the notice. Date of service of the statement of claim: [SERVICE DATE].FRAMEWORK: HMK arts. 127, 128, 129. Cite no other provision or decision; leave [Article] if needed.FORMAT: court and file details, parties, answers to the facts (numbered), evidence, legal grounds, conclusion and prayer.TONE: Calm, formal, short sentences.CONSTRAINTS: Do not write dates, amounts, courts, file numbers or decisions absent from the input; mark gaps with [ ].

Observation: a strong prompt does not make the model smarter, it makes it honest. It says where it does not know; your job is to close those places with documents and sources. Do not leave computable data such as deadlines to the model; the legal deadline calculator does the same job by fixed rules.

In a Turkish filing, AI is a drafting assistant, not the author, reviewer or signatory. Under Article 35 of the Attorneys' Act (Law No. 1136), giving opinions on legal matters, pursuing and defending claims before judicial bodies, following up court procedures and preparing all documents for those tasks belong exclusively to lawyers registered with a bar. Article 34 of the same Act requires a lawyer to perform the duty with care, accuracy and honour and to follow the professional rules set by the Union of Turkish Bar Associations. Read together, the result is plain: whatever the tool, the lawyer answers for what the document says and for what reaches the court.

That is why AI output is not a legal opinion. A model does not know which provision is in force, whether a decision really exists, or the facts of your file; it produces plausible text. Use it to build an outline, write a first version, polish language and summarise input, and keep legal judgement, citations and the prayer for relief to yourself. Courts in Turkey work in Turkish, so ask the model to draft in Turkish and review in Turkish; the English samples below are for readability. Case management keeps the file in order and document management keeps draft versions.

An effective drafting prompt is not a one-line request. It is a briefing in six small parts. Each part closes a gap the model would otherwise guess at, and every guess becomes a sentence you have to verify in the output.

1. 01/06

   Role

   Tell the model who it is and what it will not do. A role adds no knowledge; it only sets the form of the answer and the level of caution. The sentence “do not give final conclusions, flag what you do not know” is the most valuable part of the role.

   Example line

   ROLE: You are a drafting assistant for Turkish law; you give no final conclusions and mark unknowns with [ ].

   Common mistake: Inflating the role to “the perfect lawyer”. The model writes with more confidence, not more accuracy.

2. 02/06

   Fact summary

   Give the facts without identities, in order and tied to documents: date, amount, document name. State what is unknown too, so the model does not fill the gap with its own invention.

   Example line

   FACTS: On 12.03.2026 the client wired TRY 150,000 (receipt available). The client reported a delivery defect for the rest.

   Common mistake: Pasting the whole file with identity details. It creates both a confidentiality risk and noise.

3. 03/06

   Legal framework

   Supply the provisions yourself and say “rely only on these”. Pasting the provision text from the official source into the prompt keeps citation errors to a minimum.

   Example line

   FRAMEWORK: HMK arts. 127, 128, 129. Cite no other provision or decision; leave [Article] if needed.

   Common mistake: Asking “add relevant case law too”. The model may satisfy that request by inventing decisions.

4. 04/06

   Format and structure

   Ask for the heading order of the Code of Civil Procedure (HMK): Article 119 for a statement of claim, Article 129 for a statement of defence: court and parties, numbered facts, evidence, legal grounds, a clear prayer for relief. Confirm the article text yourself; the list in the prompt does not replace it.

   Example line

   FORMAT: court, parties, facts (numbered), evidence, legal grounds, conclusion and prayer.

   Common mistake: Naming no structure. The result is fluent free text that does not follow procedure.

5. 05/06

   Tone

   Name the reader and the register: calm and formal for the judge, firm but measured for the opposing party, clear and warm for the client. A length limit is part of tone as well.

   Example line

   TONE: Calm, formal, short sentences; no exaggeration or stacked adjectives.

   Common mistake: Writing “be aggressive”. You usually get needless accusations.

6. 06/06

   Constraints

   Write the prohibitions explicitly: no date, amount, court name, file number or decision that is not in the input; leave gaps in square brackets. Constraints are the biggest source of the difference you will see in the laboratory.

   Example line

   CONSTRAINTS: Do not write dates, amounts, courts, file numbers or decisions absent from the input; mark gaps with [ ].

   Common mistake: Setting no constraints. Gaps are filled silently with plausible inventions.

You do not rewrite the six parts every time. Build a prompt template for the firm and change only the fact summary and the framework per file.

Each example shows the prompt first, then a fictional output and which of its sentences must be verified and why. Personal and file details in the prompts are masked with square brackets; do the same in your own prompts.

• Error or invention
• To fill in or verify

ROLE: You are a drafting assistant for Turkish law.TASK: An outline for a statement of defence in a [CLAIM TYPE] case. Do not write the text; under each heading put a note on what belongs there.FACTS: Claimant [CLAIMANT] alleges [ALLEGATION SUMMARY]. Our client [DEFENDANT]: [DEFENCE SUMMARY].FRAMEWORK: HMK arts. 127, 128, 129. Add no other provision; [Article] if needed.FORMAT: Numbered headings.CONSTRAINTS: Do not invent dates, amounts, court names or decisions.

ROLE: You are a drafting assistant for Turkish law; you mark unknowns with [ ].TASK: A payment notice draft for a [RECEIVABLE TYPE] receivable.FACTS: Between client [CLIENT] and [ADDRESSEE] there is a [CONTRACT] dated [DATE]. The amount [AMOUNT] was not paid on [DUE DATE].FRAMEWORK: Cite no provision or decision; leave the legal basis as [Article].FORMAT: Sender, addressee, facts, demand, time limit, conclusion.TONE: Firm, measured; no threats or insults.CONSTRAINTS: Do not invent interest rate, time limit or IBAN; leave [ ].

ROLE: You are a communication assistant for a law firm.TASK: A client update message after [STAGE].FACTS: At the [STAGE] on [DATE]: [SUMMARY]. Next stage: [STAGE 2], date: [DATE 2]. Requested from the client: [DOCUMENT].FORMAT: At most 120 words, three short paragraphs; explain legal terms in one sentence.TONE: Calm, warm, clear.CONSTRAINTS: Promise no outcome (nothing like “we will win”); invent no date or document; leave [ ] where unknown.

ROLE: You are a file-summarising assistant.TASK: Build a chronology from the masked notes below.INPUT: [NOTES: Exhibit 1 contract, Exhibit 2 receipt, Exhibit 3 notice; identities masked].FRAMEWORK: Rely only on the notes given; make no comments or legal assessment.FORMAT: Date | Event | Source | Uncertainty.TONE: Neutral.CONSTRAINTS: Do not correct conflicting dates; show them in the Uncertainty column; point out missing periods.

A lawyer may not disclose matters entrusted to them or learned through their duties (Attorneys' Act art. 36). Sending identities, facts and document text to an AI service touches that duty. On the personal data side, Article 4 of the Personal Data Protection Law (KVKK, Law No. 6698) requires processing to be connected with its purpose, limited and proportionate; Article 10 sets the controller's duty to inform, and Article 12 sets technical and organisational security measures. If the service processes data abroad, the transfer is subject to the conditions of Article 9.

The practical consequence: send less first. Give the model the skeleton of facts needed to produce the text, not the identities or distinguishing details. Read from the provider's own terms how it processes data, whether it retains it and where it processes it, and record that in your firm's KVKK inventory. For the wider framework see the KVKK compliance guide for law firms, and for documents the legal document security guide.

What to send, what to mask, what to keep out

Data minimisation in five steps

1. 01Set the purpose: be clear what you want the model to write and choose only the facts that task needs (KVKK art. 4).
2. 02Mask: replace identities and distinguishing details with bracketed placeholders and keep the mapping list in your own environment.
3. 03Summarise: write your own fact summary instead of pasting the document.
4. 04Fill back: replace the placeholders in the output with real data only in your own environment.
5. 05Record: note in the file which tool was used, for what purpose and on which file.

On the notice side, the tools and transfers you use when processing a client's personal data should appear in your privacy notice (KVKK art. 10). Where data is obtained by others through unlawful means, a duty to notify the data subject and the Personal Data Protection Board as soon as possible arises (art. 12/5).

A hallucination is the model producing content that is not real but looks convincing. In law the damage is greater because the error hides inside fluent language. Learn the six types below so you know where to look for each.

Example: the response period

Error or invention

The model's sentence: “The response period is 15 days from service.”

May send

The correct position: HMK art. 127/1 sets the response period at two weeks from service of the statement of claim; the extension is available once only and is at most one month. Counting rules are in arts. 92 and 93 (for periods set in days the day of service is not counted; if the last day falls on an official holiday, the period ends at the end of working hours on the first working day after the holiday), and the judicial recess is in art. 102 (20 July to 31 August). Statutory periods are strict (art. 94/1), and the cost of an error is heavy.

Verification list

1. 01Every provision was read from the current text and matches the sentence.
2. 02Every decision or case-law citation was found in the official source; sentences that could not be found were deleted.
3. 03Every date matches the date in the document; deadlines were calculated with a tool.
4. 04Every amount and total was recalculated against the receipt and the contract.
5. 05Every fact matches the input or a file document; sentences not in the input were deleted.
6. 06Identity, court and file number placeholders were filled with real data.
7. 07The prayer for relief is clear, consistent with the narrative and includes costs.
8. 08Language, tone and length suit the client and the court; the responsible lawyer did the final read.

Review should not be a last-minute skim. A simple flow of five gates decides which error is caught at which stage; the same gates apply to a trainee's draft and to your own.

1. Gate 01/05

   Input

   Responsible: Whoever prepares the prompt

   Check the masking table before sending. Give a summary instead of a document and remove identities and distinguishing detail.

2. Gate 02/05

   Citation

   Responsible: Whoever prepares the draft

   Read every provision and decision citation from the official source. A citation that cannot be verified is deleted and an “[Article]” placeholder is left.

3. Gate 03/05

   Fact

   Responsible: Whoever prepares the draft

   Match facts, dates and amounts to the file documents. Close the review marks one by one.

4. Gate 04/05

   Prayer

   Responsible: Responsible lawyer

   The lawyer writes or fully rereads the legal grounds, strategy and prayer for relief. These decisions are not delegated to the model.

5. Gate 05/05

   Signature

   Responsible: Responsible lawyer

   A final read is done, a record is made and the document is filed under the lawyer's responsibility (Attorneys' Act arts. 34 and 35).

A trainee or support staff may prepare a draft, but the authority to prepare the document rests with the lawyer. Write in the file note which tool was used, on what date and for what purpose, and who reviewed the draft. That record serves in an internal audit and in any later dispute. For drafts and their versions, the version history in document management is useful.

The HukukBis AI module produces a draft petition in a flow consistent with the principles in this guide. Every step is under your control.

AI · Pleading draft

File2026/118 E. · Client AFile selected

Prompt

ROLE: You are a drafting assistant for Turkish law; you give no final conclusions and mark unknowns with [ ].FACTS: On 12.03.2026 the client wired TRY 150,000 (receipt available). The client reported a delivery defect for the rest.

This text is not a legal opinion; it needs lawyer review.

Draft (editable)

[COURT NAME]

FILE NO: [File No] DEFENDANT: [Defendant name], represented by Av. [Name] CLAIMANT: [Claimant name]

1. The claimant seeks TRY 250,000 as a goods price.

2. On 12.03.2026 the client wired TRY 150,000 (Exhibit 1: receipt).

3. For the remainder, a delivery defect was notified on [Date] (Exhibit 2: notice).

Deadline note: The response period is two weeks (HMK art. 127/1) from service; service date [SERVICE DATE].

Legal grounds: [Article].

CONCLUSION AND PRAYER: [Dismissal / acceptance only for the part left after setting off the TRY 150,000 paid]; costs and attorney fees to be borne by the claimant.

Open review marks00

1. 01/06
   Select a caseYou choose which file the draft is for.
2. 02/06
   Generate a draftYou start the request; a system prompt focused on Turkish law runs.
3. 03/06
   Quota checkThe monthly request quota of the HukukBis AI package is checked.
4. 04/06
   DraftThe generated text appears on screen in editable form.
5. 05/06
   Review noticeA notice appears stating that the output is not a legal opinion and needs lawyer review.
6. 06/06
   Edit and saveYou edit the text, close the marked places and save it.

‹›

01/06

Select a case

Sample file and draft; the people, amounts and dates are fictional.

The AI assistant is used through HukukBis AI, a separate package outside the plans and priced by monthly request quota. For a quote, see the HukukBis AI package on the pricing page. You can use the six prompt parts in this guide (role, facts, framework, format, tone, constraints) in your own prompt as well.

Case outcome prediction is for information only; it is shown with a disclaimer and is neither a legal opinion nor a guarantee. In HukukBis the AI runs on a local model hosted on MTY Cloud, entirely inside Turkey, and no data is transferred abroad; the data minimisation principle above still applies: give the model only the facts a draft needs. You will find the security approach on the security page.

You can turn the headings below into a one-page policy for your firm. The template is a general outline; adapt it to your firm's structure, client agreements and professional rules.

Template: AI use policy

1. 01/10

   Purpose and scope

   States which staff, which tools and which tasks (drafts, summaries, translation) the policy covers.

2. 02/10

   Approved tools

   Only tools the firm has reviewed, with provider terms assessed, may be used.

3. 03/10

   Prohibited data

   National IDs, IBANs, special-category data and the full text of documents are never sent to any AI tool.

4. 04/10

   Masking

   Identities and distinguishing details are replaced with placeholders; the mapping list stays in the firm's environment.

5. 05/10

   Verification

   No text is used until citations, dates, amounts and deadlines are verified against the source.

6. 06/10

   Responsibility and signature

   The responsible lawyer reviews every draft; the document is filed under the lawyer's responsibility.

7. 07/10

   Record

   Use is recorded in the file note with the date, tool, purpose and reviewer.

8. 08/10

   Client information

   Tool use and data transfers are stated in the privacy notice and, where appropriate, in the engagement agreement.

9. 09/10

   Incidents and breaches

   Data sent by mistake or a security incident is reported immediately to the responsible person; notification duties are assessed.

10. 10/10

    Review

    The policy is reviewed whenever tools and legislation change and at least once a year.

To enforce the policy, you can use roles and permissions to limit who can access which documents.