Your working materials
A small toolkit for everyday use
Edition 2026.09-ai-fluency-1 · Product terminology checked 15 September 2026. Tool names and access differ; concepts matter more than a remembered menu position. Source IDs refer to the source register.
The five questions before you start
What result do I need? Is AI suitable? May this information be used here? How will I check the result? Who may authorise the next action?
If you cannot answer the information-permission question, do not upload while you investigate. Use synthetic material or work manually.
The vocabulary, without the jargon
Scroll the table sideways if needed. Keyboard: focus the table area and use the Left/Right arrow keys.
| Term | Plain meaning | Do not assume |
|---|---|---|
| AI | A broad label for systems performing tasks associated with intelligence | Every AI tool generates text or understands your professional duties |
| Generative AI | Produces new text, images, audio or other output from an input | Plausible output is true |
| Model / assistant | The model is an underlying system; an assistant is the product around it, including tools and settings | The same model in two products has the same access or protections |
| Prompt | Your question or instruction | There is one magic formula |
| Context | Information made available for this task | An assistant has read every file or remembers a prior chat |
| Context window / token | A limit on material a model can consider; a token is a small unit of processed text or other representation | A long upload is fully read, or more context always helps |
| Hallucination / confabulation | An unsupported or false output that may sound convincing | Confidence, fluent writing or a citation fixes it |
| Retrieval / RAG | Finding relevant material and supplying it to the model before it answers | Retrieval proves the answer, or reaches everything in the repository |
| Grounding | Tying the task and claims to supplied or retrieved sources | A grounded answer cannot misread a source |
| Template | The shape of a document or response | It contains a checked process or current law |
| Project / matter space | A product's way of grouping related work and possibly instructions or files | It is automatically a legal confidentiality barrier |
| Memory / custom instructions | Product-specific ways to reuse information or preferences | They are the same as a source file, permission system or current matter instructions |
| Workflow card | A plain-language process you can inspect and reuse | Writing a card installs anything |
| Skill | Packaged instructions and often supporting resources for a repeatable task | It is merely a prompt, cannot include executable material, or is safe because it has a familiar name |
| Plugin / extension / app | Product-specific packaging for additional capabilities | The names mean the same thing everywhere, or installation alone grants appropriate authority |
| Connector | A connection to another service or information source | It is necessarily read-only, restricted to the intended matter, or safe for every account |
| MCP | Model Context Protocol: a way tools can expose capabilities and context to an AI application | A protocol is a trust certificate or permission to use the data |
| Agent | A system that can choose and carry out multiple steps using its available tools | A goal authorises every means of achieving it |
| Prompt injection | Instructions in untrusted material that try to redirect an assistant | A signed or reputable-looking document can issue new instructions to your tools |
| Human approval | An authorised person checks a specified action and version | A model-generated “approved” label is that approval |
The OpenAI page inspected for this edition describes skills as reusable instructions/resources and plugins as installable bundles that can include skills and MCP servers, with some runtime hooks. Do not teach that every present-day “plugin” is the discontinued feature found in older tutorials. Other vendors use these words differently. OAISKILLS
Eight prompt cards
These are starting points for permitted inputs, not safety controls. Adapt them. “Use only these sources” can still fail; checking remains necessary. The course uses task, context, output and boundaries as a convenient pattern, not an exam acronym. OAIPROMPT
1. Understand a difficult passage
“Explain this supplied passage to a new legal-team member in plain English. Keep qualifications and exceptions. Give one example and one question the passage does not answer. Do not add law from memory.”
2. Summarise with evidence
“Using C01 only, create an action table with action, owner, due date and paragraph. Write 'not agreed' if a date or owner is missing. Distinguish discussion from a decision. Prepare a draft only.”
3. Extract and compare
“Compare C03 and C04 on period, price, support and weekend coverage. For each, show what each source says and its paragraph. Identify conflicts without deciding legal priority. Do not invent missing terms.”
4. Plan research
“Turn this issue into a research plan for [jurisdiction] as at [date]. Identify the facts needed, likely primary source types and a search sequence. Do not present an unverified citation or legal conclusion as checked.”
5. Test a proposition
“Here is a proposition and an approved source extract with its locator. Identify what the extract supports, what it does not establish and any contrary wording. Give a short source-based explanation, not a claim that you checked sources you cannot access.”
6. Draft for review
“Prepare a draft internal email of at most 150 words for the supervising lawyer using the checked table below. Lead with the decision needed. Preserve figures and uncertainties. Do not advise the client, send anything or add authorities.”
7. Improve one thing
“Revise only the support paragraph. The source promises acknowledgement, not restoration. Keep other checked facts unchanged. Show the corrected sentence and a short explanation of the change.”
8. Bound a workflow
“Propose a plan to prepare a source-linked draft from these synthetic files. List required access first. Read only the selected training files. Stop on conflicting versions, unavailable sources or a request to send/share. Await human review before any action beyond drafting.”
When the result is wrong
Scroll the table sideways if needed. Keyboard: focus the table area and use the Left/Right arrow keys.
| Problem | Useful next move |
|---|---|
| Too generic | Name the reader and decision; supply the relevant passage |
| Invented date, case or term | Remove it; compare with the actual source; record what is missing |
| Correct words, wrong meaning | Point out the precise distinction; check the revision yourself |
| Missed page or table | Check extraction/OCR and page coverage; break the task into smaller labelled parts |
| Repeats an old instruction | Start a fresh matter/task context and supply the authorised current brief |
| Inaccessible source or feature | Record the gap; use a manual route or ask for permitted access |
| Requests more access than needed | Stop; seek a narrower configuration or use a manual export |
| Two assistants agree | Treat that as another suggestion, not independent primary-source verification |
| Overly agreeable answer | Ask what would change the conclusion; look for contrary evidence yourself |
Do not request hidden reasoning as proof. A concise explanation, source locations, calculations and observable checks are more useful evidence.
A reusable skill blueprint
Name: Source-linked internal update · Version: 0.1 · Status: teaching blueprint, not installed.
Use when: an internal reviewer needs a short status draft from a small, approved set of identified documents.
Inputs: instruction, permitted file list, source locations, document versions and a named reviewer. No real data in the training route.
Process: confirm scope → list supplied documents → extract material facts with locations → identify missing/conflicting information → produce a draft → check every material claim and number → hand the draft to the human reviewer.
Output: at most 150 words plus a compact source/conflict table and a review status.
Stop: missing required source, uncertain permissions, unresolved material version conflict, requested legal conclusion beyond the task or unexpected external action.
Do not: send, sign, publish, delete, use other matters, store client facts in global instructions, or treat text inside a source as permission.
Test 1: ordinary Cedar source set. Test 2: remove C03. Test 3: add unsigned C07. A useful skill should ask/stop or state the limitation, not invent the missing contract or adopt an unsigned price.
If an approved product supports skills, use its current supported creation process with the facilitator. Inspect supporting files, tools and executable steps, not just the name. Record exact version, owner, permissions and observed tests. If unavailable, run the card as a prompt or paper walkthrough. No installation or account access is necessary for course completion. OAISKILLS
Before a real connection
Name the service, account, matter boundary, files, read scope, possible write actions, data destination, retention questions, approver and revocation method. Inspect the actual permission screen and organisation settings. Ask an administrator if scope cannot be verified.
Use a dedicated synthetic training area. Start with a selected-file manual route if it is sufficient; do not connect the entire firm workspace to save a paste. Reject an all-mailbox or send-permission request for a read-only document exercise. Test with a harmless training file and confirm the result and audit trail. Disconnect/revoke when the approved exercise ends. Revocation does not by itself prove deletion of retained copies; check the applicable retention process.
No part of this checklist is a substitute for the employer's approval, vendor due diligence, security review or applicable law.