An agent is an AI that carries out steps on its own: reading files, searching, writing, and asking you when it needs a decision. This page shows how one works, how a brief turns into actions, how to get one running, and gives briefs for legal tasks you can copy.
What happens when you give an agent a brief
From your brief to a finished task
Your briefWhat to produce, from what, with what permissions, and when to stop and ask.
The agent plansThe model proposes the next step: open a folder, read a file, write a row.
The harness actsThe software around the model runs the step, if the permissions allow it, and records what happened.
You decideIn a correctly configured system, actions outside granted permissions are blocked or require approval. Anything inside runs, and you check the result.
A model may propose tool calls. Whether an action can run depends on the harness and its enforced permissions, which is why the brief spends as many words on limits as on the task.
The loop, repeated until the task is done or the agent must stop
LookRead the task, the current state, and the last result.
DecideChoose one next action.
ActThe harness runs it, or refuses it.
CheckDone? Blocked? Need to ask? Otherwise loop again.
The anatomy of a good brief
GoalOne sentence. What exists at the end that did not exist at the start.
May / may notFolders it can read, files it can create. Everything else is forbidden, and the tool should enforce it.
Rules for the outputColumns, quotes, pages. A rule for gaps: "write not found".
Stop and ask whenConflicts, unreadable files, anything that would touch a real record.
Report at the endEvery file read, every action taken. The report is how you audit the run.
Install the tool, open it in a folder that is a copy of your documents, paste a brief, and approve each permission request. Good for files on your own machine. Start with the read-only briefs below.
Chat apps with projects and tasks (Claude, ChatGPT, Copilot)
Put the documents in a project, paste the brief as the project's instructions, and run it in a conversation or as a scheduled task. Good for recurring summaries and single-document tasks. The app's own permissions and retention settings apply.
Legal platforms (Harvey, Legora, CoCounsel)
Use the platform's workflow or table feature inside the matter workspace. The brief becomes column definitions or workflow steps. Good for many documents and the same questions.
Whatever the tool: work on copies, grant the narrowest permissions it offers, and read the report of what it did before you read the output.
Starter agents (read-only, safe to try first)
2 items
Document inventory
When: A folder of documents that needs a manifest before anyone reads them.
Agent brief
Goal: create an inventory of every file in the folder "Alder - productions".
You may: read files in that folder; create one new file, inventory.csv, in the same folder.
You may not: change, move, or delete any file; open anything outside the folder; send anything anywhere.
For each file record: file name, type, page count if available, date created, date modified, and the first line of readable text. Flag files that cannot be opened or contain no readable text.
Stop and ask me if a file looks like it contains a password-protected or encrypted document.
When you finish, list every file you read and confirm that inventory.csv is the only file you created.
How to run it
Copy the folder first; the agent works on the copy.
Paste the brief into an agent that can read files (Claude Code, Codex, or a desktop agent), with its working folder set to the copy.
Approve the file reads when asked. Decline anything else.
Open inventory.csv and spot-check five rows against the files.
Permissions it needs: Read files in one named folder. Write one new CSV file. Nothing else.
What to check: Compare the row count with the file count. Open the files flagged as unreadable; they are usually scans that need OCR before any other agent task.
When: A set of letters or emails where the timeline matters.
Agent brief
Goal: list every date and deadline in the documents in folder "Correspondence".
You may: read the files in that folder; create one file, dates.csv, in the same folder.
You may not: change any existing file or go outside the folder.
For each date: the exact words, the file name and page, what the date refers to, and whether it is firm, estimated, conditional, or disputed. Quote any condition.
Do not infer dates that are not written. If a document has no dates, record one row saying so.
Stop and ask me if two documents give conflicting dates for the same event, and show both.
When finished, report how many documents you read and how many rows you wrote.
How to run it
Run on a copy of the folder.
Approve reads; the agent should not ask for anything else.
Sort dates.csv by date and read it against the documents.
Permissions it needs: Read files in one folder. Write one new file.
What to check: Every "conditional" row should carry its condition. Pick three "firm" rows and confirm nothing in the surrounding text qualifies them.
When: You have the inventory and want a first source-linked chronology.
Agent brief
Goal: build a chronology from the documents in folder "Alder - productions" and save it as chronology.csv in that folder.
Columns: Date | Event (one factual sentence) | Source file | Page or paragraph | Exact supporting quote | Certainty (stated / conditional / disputed / inferred).
Rules: one row per event; a conflict between documents means two rows marked "conflict"; no facts from outside the documents; unknown dates written as "undated" with what is known.
You may read the folder and create chronology.csv. You may not change, move, or delete anything.
Stop and ask me if a document cannot be read or if you are unsure whether two descriptions are the same event.
When finished, report the number of rows and the documents you could not read.
How to run it
Run after the inventory, so you know which files are readable.
Expect questions about possible duplicate events; answer them rather than letting the agent guess.
Open chronology.csv beside the documents.
Permissions it needs: Read files in one folder. Write one new file.
What to check: Spot-check at least one row per source document. Then look for what is missing: events you know occurred that have no row.
When: Checking whether a response to undertakings or a request to admit answers every item.
Agent brief
Goal: map each numbered request in "requests.pdf" to the part of "response.pdf" that answers it.
Write mapping.csv with columns: Request # | Request (short) | Response quote | Answered fully / partly / not at all | Qualification quoted.
A request with no corresponding response gets "no response".
You may read the two files and create mapping.csv. Nothing else.
Stop and ask me if the numbering in the two documents does not line up.
When finished, report counts for fully, partly, and not at all.
How to run it
Give the agent only the two files.
Read the "partly" and "no response" rows first; that is the follow-up letter.
Permissions it needs: Read two files. Write one new file.
What to check: Check every "fully" row; agents are generous. Confirm each quoted qualification is in the response.
When: A discovery transcript that needs to become a list of claims with page and line references.
Agent brief
Goal: from "transcript.pdf", list every factual claim the witness made about [topic], with page and line references and the exact words.
Write claims.csv with columns: Claim (one sentence) | Page | Lines | Exact words | Subject.
Flag pairs of claims that appear inconsistent, citing both locations, in a second file, inconsistencies.csv.
Do not add claims the witness did not make. Do not comment on credibility.
You may read the transcript and create the two files. Nothing else.
How to run it
Confirm the transcript is text, not a scan. If it is a scan, OCR it first.
Run, then read each flagged inconsistency in the transcript before relying on it.
Permissions it needs: Read one file. Write one new file.
What to check: Page and line references are the whole value of this output; confirm a sample of ten.
When: A draft factum or memo with authorities that must be confirmed before filing.
Agent brief
Goal: for every authority cited in "draft-factum.docx", confirm that it exists and that the pinpoint supports the proposition.
For each citation write a row in authorities.csv: Citation as written | Found in source (yes / no / different citation) | Source checked | Pinpoint confirmed (yes / no) | Note.
Search only recognized sources: CanLII, the court's website, or the firm's subscription database if connected. Do not use general web search results as confirmation.
Do not change the draft.
Stop and ask me before marking any authority "not found", and show what you searched.
When finished, list every authority marked anything other than "yes / yes".
How to run it
This agent needs a connected legal database or CanLII access through the tool; without one it can only check formatting.
Treat its "yes" as a reason to open the authority yourself, not as the check.
Permissions it needs: Read one file. Search a recognized case-law source if the tool has it connected. Write one new file.
What to check: You will be certifying these authorities. The agent narrows the work; a person opens every case before filing.
When: An automated first pass over a table or chronology before a person reviews it.
Agent brief
Goal: check "chronology.csv" against the documents in folder "Alder - productions".
For each row, find the quoted passage at the cited file and page. Report "found", "not found", or "wording differs" (with the actual wording).
Write check-report.csv with columns: Row # | Result | Actual wording if different.
Do not edit chronology.csv. Do not add rows. Report only.
How to run it
Run with a different model than the one that built the table, if you can.
Read every "not found" and "wording differs" row in the source.
Permissions it needs: Read the output file and the source folder. Write one report file.
What to check: This catches citation errors, not reasoning errors. It does not replace reading the rows that matter.
When: A first draft that must be traceable to the record.
Agent brief
Goal: draft [document] for [audience] from the documents in folder "Record", and save it as draft.docx in folder "Drafts".
Use only facts in the record. After each factual sentence add a bracketed reference: [file, page].
Where a fact is needed but not in the record, write "[FACT NEEDED: describe]" instead of inventing it.
Tone: [tone]. Length: about [n] words.
You may read "Record" and create draft.docx. Nothing else.
When finished, list every "[FACT NEEDED]" marker.
How to run it
Keep the record folder small and relevant; the agent reads all of it.
Work through the FACT NEEDED list before anything else.
Permissions it needs: Read a folder. Write one new document.
What to check: Check each bracketed reference before removing it. Then compare conditions and hedges in the draft with the record.
When: Long threads that need a short, accurate record on the file.
Agent brief
Goal: write a file note from "thread.eml" (or the exported PDF) and save it as note.md.
Include: participants, date range, each decision or commitment (sender, date, exact quote), and open items, in chronological order.
End with attachments mentioned and whether each was included.
Do not characterize motives. Do not send anything.
How to run it
Export the thread first; do not give the agent your mailbox.
Read the note against the thread once.
Permissions it needs: Read one exported thread. Write one note.
What to check: Confirm each quoted commitment and who sent it.
A review-table workflow (Harvey, Legora, or similar)
When: Many documents, the same questions for each, inside a platform that offers tables.
Workflow pattern
Documents: the uploaded set for [matter].
Columns (one question per column):
1. Document date (quote the date line).
2. Parties named (quote).
3. Mentions [topic]? Quote the passage, or "no mention".
4. Any deadline or notice requirement? Quote it with the clause number.
For every cell: quote and page. Leave a cell blank rather than guess.
Export the table when complete.
How to run it
Create the table in the platform, add the documents, define the columns from the list above.
Run, then sort by the "no mention" column and sample those documents by searching their text yourself.
Permissions it needs: Whatever the platform grants; confirm the matter workspace is the approved one.
What to check: Sample cells from every column. Platforms differ in how they handle scans and tables inside documents; those are where cells go wrong.
A scheduled summary (Claude or ChatGPT scheduled tasks)
When: A recurring brief you want without asking each time, from material that is already in the tool.
Workflow pattern
Every weekday at 7:30 am, using only the documents in the "[matter] - weekly" project, produce a one-page note: what changed since the last note, with the document and date for each item, and any deadline in the next 14 days.
If nothing changed, say so in one line. Do not search the web. Do not send the note anywhere; leave it in the project.
How to run it
Put only the material you want summarized into the project.
Review the first three notes carefully before trusting the routine.
Permissions it needs: Read access to the connected source only.
What to check: Dates and deadlines in the note against the documents, every time.
