AI Agents for Lawyers

Playbooks

Build a Prompt Library for Your Practice: Structure, Governance and 12 Starter Prompts

Build a law firm prompt library lawyers will use: folder structure, a header for every prompt, a review cadence for model changes and 12 tested starter prompts.

By Dr. Niklas SchmidtUpdated 13 November 202611 min read
On this page
  1. Why forty private prompt files are not a library
  2. The header every prompt carries
  3. Folder structure, naming and where the library lives
  4. Governance: the quality bar before a prompt goes in
  5. The twelve starter prompts
  6. Review cadence and retiring prompts as models change
  7. What to do this week

A prompt library for a law practice is a governed set of tested, versioned prompts, each carrying a named owner, a data classification and a review date, stored where lawyers already work and mirrored into Claude Projects, Skills or custom GPTs. This playbook gives you the folder structure, the header every prompt needs, a review cadence and twelve starter prompts to copy today.

Why forty private prompt files are not a library

Most firms already have a prompt library; it is spread across forty personal note files and nobody can find it. Thomson Reuters' Future of Professionals report (June 2026; 1,816 respondents in 62 countries) found that 34% of professionals use unsanctioned "shadow AI", and that 66% of those in organisations with a named AI strategy say the technology meets expectations, against 22% without one. Clio's 2025 Legal Trends Report put the governance gap more bluntly: 53% of legal professionals say their firm has no AI policy or do not know whether it has one. A shared library is the smallest artefact that turns individual habit into firm practice. It says which prompts are approved, on which tools, with which data, and who answers when one goes wrong.

The five-part formula that 2026 guidance converges on, role and context, task, sources, output format and constraints, produces more reliable output than conversational prompting, and positive examples outperform lists of prohibitions (LegalFly, 17 Sep 2026). None of that helps if the well-built prompt lives on one associate's laptop. The prompt engineering guide covers how to write one prompt well; this page covers how to keep sixty of them alive.

The header every prompt carries

Treat each prompt like a precedent in the know-how system: it has a reference, an owner, a version and a review date, and it says what it may be used for. The header below sits at the top of every library entry and is copied into whichever tool the prompt is deployed to.

FieldWhat goes in itWhy it matters
IDSection code and running number, e.g. REV-03"Ran REV-03 v3.1" in a matter note gives an audit trail
OwnerA named lawyer or knowledge professional, not a committeeSomeone must answer when the prompt misfires
Version1.0 on release; +0.1 for wording; +1.0 when task, model tier or data class changesA partner can tell whether the version used was current
Model tier"Any frontier model", "reasoning model", or a named model with the test datePrompts tuned for one model degrade on another
Data classGREEN (no client data), AMBER (client data, business or enterprise tier only), RED (never in a general-purpose model)Mirrors the traffic-light structure the North Carolina Bar recommended in January 2026
JurisdictionOne, fixed or set per runEvery system Vals tested in October 2025 lost roughly 11–14 points on multi-jurisdiction questions
Last testedDate, number of test documents, tester's initials"Tested" means by someone other than the author
Review byA date no more than one quarter outUntested prompts expire

The data class is the field lawyers skip and the one that matters most. The North Carolina Bar's January 2026 note recommends a traffic-light structure, red for client data in consumer tools, yellow for research and drafting with verification, green for administrative and marketing work, and observes that blanket bans push lawyers towards shadow AI rather than away from it (NC Bar, 13 Jan 2026). Putting the class in the header means the question "may this prompt see client data on this tool" is answered before anyone pastes. As of September 2026 consumer tiers train on conversations unless the user opts out, business and API tiers do not, and OpenAI retains conversations for 30 days for abuse monitoring even after an opt-out (Sonomos, 1 Sep 2026). The confidentiality checklist maps each vendor's tiers onto the three classes.

Folder structure, naming and where the library lives

Keep the canonical copy as plain Markdown or text files in the document management system or SharePoint, permission-bound like any other know-how. Projects, Skills and GPTs are deployments of that copy, not a second source of truth.

prompt-library/
  00-governance/
    README.md              how to submit, test and retire a prompt
    CHANGELOG.md           one line per version change, newest first
    data-classes.md        GREEN / AMBER / RED definitions and the tool matrix
  10-research/
    RES-01-issue-framing-v1.2.md
    RES-02-verified-authorities-memo-v2.0.md
  20-drafting/
    DFT-01-clause-from-precedent-v1.4.md
    DFT-02-plain-english-clause-v1.1.md
    DFT-03-tax-irac-skeleton-v1.0.md
  30-review/
    REV-01-contract-triage-v3.1.md
    REV-02-red-team-argument-v1.3.md
    REV-03-processor-terms-v2.2.md
    REV-04-lease-abstract-v1.0.md
    REV-05-termination-letter-check-v1.2.md
  40-client-communication/
    CLI-01-note-to-client-update-v1.5.md
  50-operations/
    OPS-01-time-narrative-cleaner-v1.0.md
  90-retired/              every superseded version, kept for audit
  tests/
    REV-01/                the ten test documents and the expected outputs

Naming is section code, running number, a short slug and the version. The version in the filename is deliberate: a lawyer who opens REV-01-contract-triage-v3.1.md knows what they are running, and the retired folder keeps every earlier version. The tests/ folder is what separates a library from a scrapbook. Without the test documents nobody can tell whether v3.2 is better than v3.1, and the argument will be won by seniority.

Where the prompts run is a separate decision; the containers differ more than their marketing suggests.

ContainerWhat it holdsSharing and dataFits
Markdown in the DMS or SharePointCanonical text, header, changelog, testsFirm-wide, permission-bound; reachable by Claude, ChatGPT and Copilot through the iManage and NetDocuments MCP servers (May 2026) under existing ethical wallsSource of truth
Claude ProjectInstructions plus static reference files in one workspaceShared on Team and Enterprise, never used for trainingPractice-group or matter-level bundles
Claude SkillA SKILL.md folder (YAML frontmatter plus instructions) with optional scripts, references and assets; fires only on a matching trigger; runs in Claude.ai, Claude Code and the APIPackaged and installed; single-user unless distributedPrompts that should trigger themselves, such as a citation-format check
Custom GPTInstructions, knowledge files, optional actionsWorkspace sharing on Business and Enterprise, which do not train by defaultClient explainers and internal checklists
Copilot Studio agentPrompts over Microsoft 365 data, with DLP over connectors and Purview audit (documentation updated 4 Aug 2026)Tenant-wide under the Microsoft DPAFirms standardised on Microsoft 365

The Skills description comes from a guide updated on 7 September 2026, which draws the line that matters for library design: Skills fire on a trigger, Projects are static context in one workspace (Layer3 Labs, 7 Sep 2026). Two cautions. Zack Shapiro's post on hand-written Claude Skills that edit Word XML for tracked changes drew seven million views and a week of argument over whether a general model plus Skills beats purpose-built tools (Artificial Lawyer, 2 Mar 2026); for a library, Skills are a deployment format, not a substitute for governed text. GC AI's critique of Anthropic's legal plugin lists the gaps a library must fill: context resets between sessions, no citation verification without a connector, and single-user markdown files where a team playbook is needed (GC AI, 2026). Anthropic's open-source Claude for Legal repository, at 150+ skills and 13 scheduled agents as of 18 September 2026, starts every plugin with a cold-start interview that writes a practice profile every other skill reads (github.com/anthropics/claude-for-legal); the firm's library is that profile, written once and shared. The playbooks on setting up a Claude Project for a matter and creating a custom GPT cover deployment, and the glossary entry on skills and plugins explains the packaging.

Governance: the quality bar before a prompt goes in

A prompt enters the library when it passes four checks, none of them performed by its author.

  1. Tested on five to ten documents or fact patterns by a second lawyer, with the outputs kept in tests/. A prompt that has only ever worked on the author's favourite example has not been tested.
  2. Data class assigned and defensible. If the answer to "which tier may this run on" is "it depends", the class is RED until someone decides.
  3. No citation requests to a general-purpose model. Research prompts follow the safe pattern: the model frames issues and writes search strings, the lawyer runs the search in Westlaw, Lexis or the national database, and verified authorities go back in with the instruction to work only with what has been provided (The Legal Prompts, 18 Feb 2026). Courts now describe verification as a "nondelegable duty" that AI "in no way abrogates" (Kemp, Q1 2026 sanctions review), and Damien Charlotin's database counted 2,041 decisions involving hallucinated material by 14 September 2026, 811 of them involving lawyers. The research verification protocol is the companion to RES-01 and RES-02 below.
  4. Output format fixed. Tables, word limits and a mandatory closing line ("Draft for lawyer review") make outputs comparable across versions and across lawyers.

Ownership sits with a folder, not with the library: the employment knowledge lawyer owns 30-review/REV-05, and the changelog line names them. One library owner, usually the head of knowledge or a partner who cares, runs the cadence described after the prompts. Anyone may submit, including trainees; the firm AI policy should say so, because a library that only partners may add to fills up with prompts nobody under forty uses.

The twelve starter prompts

Each prompt follows the five-part structure and carries its header. Replace the braces, keep the constraints, and run the test set before changing the wording. Owners and test dates are illustrative. The red-team prompt has a longer treatment in the red-teaming playbook.

Research

ID RES-01 · Issue framing and search strings · Owner: litigation KM · v1.2
Data: AMBER (anonymised facts, business tier) · Tested 2026-09-12 on 10 fact patterns · Review by 2026-12-31

You are a senior {jurisdiction} litigator with 15 years of experience, helping a colleague frame a research task before they run the search themselves in {Westlaw / Lexis / RIS / juris}.

Fact pattern (anonymised): {facts}
Question to be answered: {question}

Produce, in this order:
1. The three to five legal issues the facts raise, each in one sentence.
2. For each issue, the elements or tests a court would apply as you understand them, with a confidence label (high / medium / low).
3. Eight to twelve search strings suitable for {database}, using connectors and field restrictions where that database supports them.
4. The assumptions you made about the facts.

Constraints: Do NOT provide case names, citations, statute section numbers or quotations. Where you would cite, write "[authority to be located]" instead. Stay within {jurisdiction}. Plain English, no headings other than the four numbered parts.
ID RES-02 · Memo from verified authorities only · Owner: litigation KM · v2.0
Data: AMBER (business tier) · Tested 2026-09-12 on 6 questions · Review by 2026-12-31

You are a {jurisdiction} litigation associate drafting an internal research note for a partner.

Work only with the authorities listed below. Each has been read and verified by a lawyer, and the quoted passages are the only text you may rely on. Do not add any case, statute, regulation or commentary that is not in the list.

Question: {question}
Authorities (number, name, citation, verified passage, date verified, verified by):
{authority 1}
{authority 2}
{authority 3}

Output: a note of no more than 600 words under the headings Question, Short answer, Analysis, Open points. In the Analysis, pin every proposition to a numbered authority in square brackets, e.g. [2]. Where the authorities do not answer a point, say so under Open points rather than filling the gap. British English. End with the line: "Draft for lawyer review. Authorities verified on {date} by {initials}."

Drafting

ID DFT-01 · Clause in the style of a precedent · Owner: commercial KM · v1.4
Data: GREEN if the precedent is the firm's template, AMBER if it is a client's · Tested 2026-09-05 on 8 clauses · Review by 2026-12-31

You are a commercial contracts lawyer drafting for the {supplier / customer} under {governing law}.

Here is a clause from our precedent. Match its defined terms, numbering, sentence length and tone exactly:
{paste precedent clause}

Draft a {type of clause, e.g. limitation of liability} for the same agreement that:
- {position 1, e.g. caps liability at 100% of fees paid in the preceding 12 months}
- {position 2, e.g. carves out death, personal injury, fraud and anything that cannot be limited by law}
- {position 3}

Then, under the heading "Drafting notes", list in bullet points: any defined term you used that does not appear in the precedent; any point where the positions above conflict with each other; and any provision that in your understanding is unusual under {governing law}, marked "check". Do not cite cases or statutes. Do not alter the precedent clause itself.
ID DFT-02 · Plain-English clause explanation for a client · Owner: private client KM · v1.1
Data: AMBER (remove client names before any consumer-tier use) · Tested 2026-08-29 on 10 clauses · Review by 2026-12-31

You are a {practice area} lawyer explaining a clause to a client who is intelligent, busy and not a lawyer.

Clause: {paste clause}
Context: {one sentence, e.g. the survivorship clause in a will for a married couple with two adult children}
Governing law: {e.g. English law}

Write an explanation in plain English of no more than 120 words. Say what the clause does, when it bites and what happens if the client does nothing. No legal jargon unless you define it in the same sentence. No advice on whether to sign, amend or reject. One paragraph, no bullet points, no headings. Then, on a separate line, list any question the client would need to answer before a lawyer could advise further.
ID DFT-03 · Tax analysis skeleton with assumptions log · Owner: tax KM · v1.0
Data: AMBER (anonymised structure only, business tier) · Tested 2026-09-10 on 6 structures · Review by 2026-12-31

You are a {jurisdiction} tax adviser preparing the skeleton of an analysis for partner review. You are not producing the final advice.

Facts: {anonymised structure and steps}
Rule as we understand it (pasted from the firm's own note or the statute text, never from memory): {rule text}
Question: {question}

Work through the question in IRAC form: Issue, Rule, Application, Conclusion. Under Rule, use only the rule text supplied; do not add sections, thresholds, rates or dates from your own knowledge. Under Application, walk through each step of the structure in order and say whether the rule applies, might apply or does not apply, with the reason. Under Conclusion, give a provisional answer and a confidence label. Finish with an "Assumptions and open points" list containing every fact you had to assume, every point on which the rule text is silent and every item that must be checked against current guidance. Maximum 700 words.

Review

ID REV-01 · Contract triage against the playbook · Owner: commercial KM · v3.1
Data: AMBER (counterparty paper, business tier) · Tested 2026-09-15 on 10 NDAs · Review by 2026-12-31

You are a commercial lawyer reviewing an inbound {NDA / DPA / SaaS agreement} for {our company / our client} against the attached playbook. The playbook is the only source of our positions.

Document: {paste or attach}
Playbook: {attach playbook file}

For each playbook row, report in a table with the columns: Playbook point | Document position (quote the words) | Clause | Rating | Suggested change (one sentence). Ratings: GREEN = matches our position or fallback; AMBER = within the fallback but needs a lawyer's judgment; RED = outside the fallback or an escalation trigger is met.

Below the table, give one overall rating equal to the worst row, then list the three changes that matter most. If the document is silent on a playbook point, rate it AMBER and write "silent". Do not invent positions that are not in the playbook. Do not propose wording for RED items; a lawyer will draft those.
ID REV-02 · Red-team a draft argument · Owner: disputes KM · v1.3
Data: AMBER (business tier; strip party names on any consumer tier) · Tested 2026-09-08 on 5 skeletons · Review by 2026-12-31

Adopt the role of opposing counsel before a {court / arbitral tribunal} in {jurisdiction}. You have just received the argument below and your job is to defeat it.

Argument: {paste the draft section or skeleton}

Give the five strongest arguments against it, ranked from most to least dangerous. For each: (a) the point in one sentence; (b) the weakest link in our reasoning it attacks; (c) what evidence or authority opposing counsel would need to make it stick, described generically without naming cases; (d) the question a judge or arbitrator is most likely to put to us about it. Then, under the heading "What we would have to concede", list any point where the draft overstates its position. Do not rewrite our argument. Do not cite cases, awards or statutes.
ID REV-03 · Processor terms against Art. 28(3) GDPR · Owner: privacy KM · v2.2
Data: AMBER (counterparty paper, business tier) · Tested 2026-09-01 on 8 DPAs · Review by 2026-12-31

You are an EU data protection lawyer reviewing a data processing agreement offered by a vendor to {our client, a controller}.

DPA text: {paste or attach}
Reference: the attached extract of Art. 28(3) GDPR and our checklist of required content. Use these, not your memory.

Produce a table with one row per checklist item: Required content | Present? (Yes / Partial / No) | Clause | Quoted wording | Gap in one sentence. Then list separately: sub-processor rules (notice period, objection right, flow-down); the international transfer mechanism named, if any; audit rights and who pays; deletion or return at end of term and the deadline stated. Mark "check" any point where the DPA refers to a document that is not attached. Do not assess adequacy decisions, standard contractual clauses or supervisory-authority guidance from your own knowledge; mark those "lawyer to confirm".
ID REV-04 · Lease abstract · Owner: real estate KM · v1.0
Data: AMBER (business tier) · Tested 2026-08-22 on 12 leases · Review by 2026-12-31

You are a real estate lawyer abstracting a commercial lease for a due diligence report under {jurisdiction} law.

Lease: {attach}

Extract the following into a two-column table, quoting the exact words and giving the clause number for every entry: parties; premises and any plan reference; term start and expiry; break options (who, when, conditions); rent, payment dates and review mechanism; service charge and any cap; permitted use; assignment and subletting restrictions; repair obligation; insurance; guarantor; any unusual or onerous provision you notice. Where a field is not in the lease, write "not found" rather than guessing. After the table, list in bullets any provision that refers to a side letter, licence or deed not provided. Do not summarise the law; abstract the document.
ID REV-05 · Termination letter check against the reference table · Owner: employment KM · v1.2
Data: RED on any consumer tier; AMBER on a business tier with the employee's name removed · Tested 2026-09-03 on 7 letters · Review by 2026-12-31

You are an employment lawyer in {jurisdiction} checking a draft termination letter before it goes to the client.

Draft letter: {paste, with the employee's name replaced by "the Employee"}
Reference file: {attach the firm's notice-period and formality table for {jurisdiction}}
Facts: start date {date}; role {role}; contractual notice clause {quote}; termination type {ordinary / summary / mutual}; delivery method {method}.

Check, using only the reference file and the facts given: the notice period stated against the minimum in the reference file; the termination date arithmetic; the required form and delivery method; the mandatory content listed in the reference file; any statement in the letter that goes beyond what the facts support. Report as a checklist with Pass / Fail / Cannot tell and a one-line reason for each. Do not redraft the letter. Do not add legal requirements that are not in the reference file; if you believe one is missing, put it under "Suggested additions to the reference file" for the owner to consider.

Client communication and operations

ID CLI-01 · Internal note to client update · Owner: BD and KM · v1.5
Data: GREEN (the internal note is already client-safe) · Tested 2026-09-11 on 6 notes · Review by 2026-12-31

You are a {practice area} lawyer turning an internal know-how note into a short client update.

Internal note: {paste}
Audience: {e.g. general counsel of mid-sized companies in Austria and Germany}
Purpose: inform, not advise.

Write a client update of 180 to 220 words: a one-sentence headline stating what changed and from when; two short paragraphs on what it means in practice; a closing sentence saying what the reader should consider doing now and whom to contact. Use only facts, dates and figures that appear in the internal note; where a date or figure is missing, write "[date]" or "[figure]" for the author to complete. British English. No case citations, no marketing adjectives, no exclamation marks. End with a line for the author's name.
ID OPS-01 · Time-recording narrative cleaner · Owner: finance and KM · v1.0
Data: AMBER (narratives name clients and matters; business tier) · Tested 2026-09-16 on 40 narratives · Review by 2026-12-31

You are a billing reviewer at a law firm rewriting time-recording narratives so that a client can understand what was done and why it was worth the time.

Rules for every narrative: start with a verb; name the document, meeting or issue; state the purpose; no internal jargon or initials; no more than 25 words; keep the recorded time unchanged; never add work that is not described in the original; keep any privileged description general (e.g. "advice on strategy", not the content of the advice).

Narratives (one per line, format: date | timekeeper | hours | narrative):
{paste}

Output the same lines in the same order with only the narrative text changed, followed by a list of lines you could not rewrite because the original was too vague, so the timekeeper can supply detail.

OPS-01 is the prompt-sized version of a tool that already exists at scale: a Linklaters associate's vibe-coded "Narrative Checker" cleans roughly 1,000 time-recording narratives a minute. REV-04 does by prompt what Ashurst reports doing with a Harvey Agent Builder workflow that "saves hours on producing lease summaries" (March 2026). When a library prompt starts to look like that, it is a candidate for the knowledge-management use case or for automation, and its entry should say where it went.

Review cadence and retiring prompts as models change

Prompts age faster than precedents because the thing that reads them changes every few months. Anthropic alone shipped Opus 4.6, 4.7 and 4.8, Opus 5 and Fable 5.1 between February and September 2026; OpenAI retires GPT-5.5 from ChatGPT on 14 October 2026 and closes its Agent Builder on 30 November 2026. A prompt tested in April against a model that no longer exists is an untested prompt.

TriggerActionOwnerDeadline
New prompt submittedSecond-lawyer test on five to ten documents; assign data class; complete header and changelogFolder owner10 working days
QuarterlyRe-run every active prompt's test set on the firm's current default model; log pass or failLibrary ownerQuarter end
Frontier model release, retirement or change of defaultRe-test prompts marked model-sensitive (research, extraction, tax)Library owner30 days
Sanction decision, regulator notice or bar opinionReview the constraints in every research and drafting promptRisk partner5 working days
Task absorbed by a tool (Word Legal Agent, a Claude for Legal skill, a vendor agent)Move to 90-retired/ with a note naming the replacementFolder ownerNext quarterly review
Prompt fails its test set twiceDeprecate; owner rewrites or retiresFolder ownerImmediately

Status moves from draft to active to deprecated to retired, and retired prompts are never deleted. Two reasons. First, the audit trail: if a client or a court asks in 2028 what instructions produced a 2026 draft, the answer should be a file, not a memory. Second, the training record. Article 4 of the EU AI Act, reworded by the Digital Omnibus into a duty to "take measures to support the development of AI literacy", was not deferred and applies to any organisation whose staff use tools like ChatGPT. The Commission's FAQ (updated 27 July 2026) requires no certificates or tests but expects an internal training record, national market-surveillance authorities have supervised the duty since 2 August 2026, and regulators expect auditable records of who was trained, when, on what and how it was assessed (MuchSkills, 2026). A library with a changelog, test logs and a list of who was walked through which prompts is that record; the Article 4 explainer sets out what the authorities expect to see.

Retire prompts when the task moves inside a tool. Microsoft's Legal Agent for Word (Frontier preview, 30 April 2026) reviews a contract clause by clause against a playbook and writes tracked changes with explanatory comments; once a firm enables it, REV-01 becomes a playbook file rather than a prompt, and the prompt goes to 90-retired/ with a line saying so.

What to do this week

  1. Create the folder tree, the README and the data-classes file. Half a day, most of it spent agreeing what AMBER means on each tool the firm has licensed.
  2. Collect the prompts lawyers already use. An email to each practice group asking for their three most-used prompts yields more than any workshop; read the replies for data-class problems first.
  3. Adopt three of the twelve above, one research, one review, one drafting, and run each on five documents with a second lawyer scoring the output.
  4. Deploy those three into one container only, a Team-plan Claude Project or a workspace GPT, and record the deployment in the header.
  5. Book the first quarterly review now, and put the next frontier model release on the same calendar.

Module 1 of the AI Lab for Lawyers course covers the prompting fundamentals these twelve prompts rely on, and the later modules show where a library prompt turns into a Project, a Skill or an automated workflow.

Frequently asked questions

What should a law firm prompt library contain?

Tested prompts stored as plain files, each with a header giving an ID, a named owner, a version, the model tier it was tested on, a data class saying which tools may see which data, a jurisdiction, the last test date and a review date. Beside the prompts sit a changelog, a data-class matrix and a tests folder holding the documents and expected outputs used to approve each prompt.

Where should a law firm store its prompt library?

Keep the canonical copy in the document management system or SharePoint, permission-bound like any other know-how, and deploy copies into Claude Projects, Claude Skills, custom GPTs or Copilot Studio agents. Since May 2026 the iManage and NetDocuments MCP servers let Claude, ChatGPT and Copilot read a DMS folder under the lawyer's own ethical walls, so the file store can also be the runtime source.

Can lawyers share prompts that contain client information?

Share the prompt, not the client data. A library prompt holds placeholders and a data class; the facts are pasted at run time on a tier whose terms permit it. As of September 2026 consumer tiers of ChatGPT, Claude and Gemini train on conversations unless the user opts out, while business, enterprise and API tiers do not, and ABA Formal Opinion 512 requires informed consent before a self-learning tool may see confidential data.

How often should prompts in a library be reviewed or retired?

Re-run every active prompt against its test set quarterly and within 30 days of any frontier model release, retirement or change of the firm's default model. Retire a prompt when it fails its tests twice, when the task moves inside a tool such as the Word Legal Agent, or when its reference files are superseded. Move retired prompts to an archive folder rather than deleting them; they are part of the audit and training record.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion