User Persona: The Legal & Compliance Professional
Profile: Private practice, in-house counsel, paralegals, and compliance officers — contracts, correspondence, discovery material, policies, and regulatory filings. Primary Goal: Move faster through long documents without sending privileged or client-identifying material somewhere it shouldn't go.
Rephlo is a tool, not legal advice, and not a compliance program. It provides technical controls — local processing, redaction, and routing — that help you work carefully with sensitive material. Whether a particular use is consistent with your professional obligations, your client engagement terms, your jurisdiction's rules on confidentiality, or your firm's policy is a judgment only you and your firm can make. Confirm before putting client material through any AI tool.
Recommended Setup
Providers
- Long-document work (review, summarizing, comparison): Claude Sonnet 4.6 or GPT-5.4 — you want context length and careful reading here.
- Hardest analysis: GPT-5.5 or Claude Opus 4.8 (Pro Plus or higher).
- Privileged or client-identifying material: an on-device model, so none of your content leaves the machine — or a cloud model with redaction and Block send if redaction is uncertain both on.
- Firm data residency: OpenAI-Compatible pointed at your firm's Azure OpenAI endpoint.
Turn these on first
Open Settings ▸ Privacy:
- Scrub PII Before Cloud — strips detected personal information before a request reaches a cloud provider.
- Also detect dates & addresses (on-device AI) — the one-time ~1.16 GB download that adds names, dates of birth, addresses, and other IDs (passport, driver's licence, national ID, tax ID, bank account numbers). Without it, only email, phone, SSN, and card details are caught. In legal work, names are usually the category that matters most, and names are in this layer.
- Block send if redaction is uncertain — Rephlo holds a request back rather than sending something it isn't sure it cleaned.
- Force Local for Sensitive — routes anything you've flagged sensitive to an on-device model instead of the cloud. Requires Pro or above and at least one on-device model downloaded.
Knowledge Spaces
Space: Precedent & Clause Bank(approved clauses, fallback positions, negotiated variants).Space: Firm Policies & Style(house style, formatting rules, engagement-letter templates).Space: Regulatory Source Material(statutes, guidance, regulator publications you rely on regularly).- Keep matter files out of shared Spaces. Attach client documents to a single conversation instead. A Space is a persistent, searchable index; an attachment is not.
Essential Commands
-
"Clause Extraction"
- Prompt: "Extract every clause in this agreement covering termination, liability, indemnity, governing law, assignment, and confidentiality. Quote each verbatim with its clause number. If a category is absent, say 'not present' — do not summarize a nearby clause in its place."
- Usage: Select a contract or attach it to a chat.
-
"Issue Spotting"
- Prompt: "Read this agreement and list provisions that are unusual, one-sided, internally inconsistent, or that reference a defined term that is never defined. For each, quote the text and state the concern in one sentence. Do not propose fixes yet."
- Usage: Select a draft received from the other side.
-
"Plain-English Translation"
- Prompt: "Rewrite this passage in plain English for a business audience. Preserve every obligation, condition, exception, and deadline exactly. Flag anything genuinely ambiguous rather than resolving it."
- Usage: Select dense drafting for a client explainer or a board note.
-
"Obligations Table"
- Prompt: "Produce a table of every obligation in this agreement: obligated party, what is required, trigger, deadline, and consequence of failure. Quote the clause number for each row."
- Usage: Select a signed agreement when building a compliance calendar.
-
"Redline Rationale"
- Prompt: "Compare these two versions of the clause. State exactly what changed, and what each change does to the risk allocation between the parties. Ignore purely stylistic edits."
- Usage: Select the old and new text together.
-
"Policy Gap Check"
- Prompt: "Compare this internal policy against the attached regulatory guidance. List requirements in the guidance that the policy does not address. Quote both sides. Distinguish what the guidance requires from what it merely recommends."
- Usage: Select the policy with the Regulatory Source Material Space attached.
Workflow Example: First Pass on an Inbound Contract
- Check privacy state: redaction on, on-device layer downloaded, block-if-uncertain on.
- Attach, don't index: attach the contract to a chat rather than adding it to a Space.
- Extract: run "Clause Extraction" to get the structural map.
- Spot: run "Issue Spotting" for the anomalies.
- Interrogate: switch to Multi-Turn Chat and ask follow-ups — the conversation keeps context, so you can push on a single clause without re-pasting the agreement.
- Verify: open the contract and read every clause the model quoted. The quotes are the checkable part; the characterization is not.
Workflow Example: Scanned and Photographed Material
Discovery, court filings, and executed originals often arrive as scans.
- Set the recognizer: Settings ▸ Privacy ▸ Scanned document language should match the document.
- Attach the scan: text is recognized on your own computer — nothing is uploaded to be read. Long documents are indexed to a 200-page limit and marked Partial OCR, so check that badge before you rely on a bundle being complete.
- Images are covered too: an attached image containing personal information is replaced by its redacted text rather than sent as a picture.
- Know the gap: OCR reads text. Signatures, handwritten annotations, and margin notes it cannot read are not caught — which is exactly where identifying information hides in litigation material. Keep block-if-uncertain on.
Choosing Where a Request Runs
| Material | Recommended setup | Why |
|---|---|---|
| Public statutes, published guidance, general research | Any cloud model | Nothing confidential involved |
| Firm precedent and templates, de-identified | Cloud model + redaction | Full model quality |
| Client matter files, correspondence | On-device model, or Force Local for Sensitive | None of your content leaves the machine |
| Anything you would not put in an email to an outside party | On-device model | Apply the same instinct you already have |
That last row is the practical test. If you would hesitate to forward it externally without thinking, run it locally.
Pro Tips for Legal & Compliance
Verify every citation. Every one.
Language models generate citations that look correct and are not — cases that don't exist, sections that say something else, quotations that were never written. This has ended badly for practitioners in real proceedings. Rephlo does not check citations for you. Use these Commands to find and organize text you already have, and treat any authority the model volunteers as unverified until you have read the source.
Quotes beat summaries
The Commands above ask for verbatim quotes with clause numbers on purpose. A quote is something you can check in ten seconds against the document. A summary is something you have to re-derive. Prefer prompts that make the model point at the text rather than paraphrase it.
Ask what's missing, not just what's there
"List provisions that are absent" is a much better use of a model than "summarize this contract". Absence is tedious for a human to check and mechanical for a model to enumerate.
Redaction changes what comes back
If names are redacted before sending, the response won't contain them either. That's the point — but it means output referring to "[REDACTED]" needs your own reinstatement pass. For a document where the parties' identities are integral to the analysis, run it on an on-device model instead.
Retention is a decision, not a default
Settings ▸ Privacy ▸ History retention offers Forever, 90 days, 30 days, or Immediate. History is a record of your inputs. For matter work, set this to match your firm's retention policy rather than leaving it as it came.
Next Persona: Return to For Financial Services or For Students.
Related pages
- Privacy & Data — every redaction and routing setting in detail
- On-Device vs Cloud Models — what a local model can and can't match
- How Rephlo Reads Your Files — attachments, Spaces, scans, and images