# Legal Compliance Obligation Extraction > Turns a supplied regulation, rule, contract or policy instrument into an atomic obligation register: one row per obligation, each carrying its provision reference, the obligated party, the trigger that makes it bite, the deadline or frequency, the modality (must, must not, may), any carve-out, and the evidence that would demonstrate compliance. Extracts only from the supplied text and marks anything the text does not settle as an open question rather than resolving it. Use when a rule or contract must be turned into something trackable, when a compliance register or obligations matrix is being built, when a counterparty agreement needs its duties listed, or when an impact assessment needs a register it does not have. Trigger on 'extract the obligations', 'obligations register', 'what do we have to do under this', 'build the compliance matrix', 'list the duties in this contract', 'clause register'. Not for assessing what a regulatory change means for the business and what to remediate (use legal-compliance-reg - Skill: `alihusains/legal-compliance-obligation-extraction` (Agent Skill) - Install (CLI): `npx skillmds@latest add alihusains/legal-compliance-obligation-extraction` - Raw SKILL.md: https://api.skillmd.com/api/skills/alihusains/legal-compliance-obligation-extraction/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: alihusains (https://skillmd.com/u/alihusains) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/alihusains/legal-compliance-obligation-extraction --- # Obligation extraction ## Purpose Obligations get lost between the instrument and the register. Someone reads a forty-page rule or a signed agreement, writes eight bullet points, and the organisation tracks the bullet points forever — losing the conditions, the carve-outs, the party that is actually bound, and the paragraph number a regulator or a counterparty will cite. This skill produces the atomic register: one testable duty per row, each traceable back to the provision it came from, so downstream control mapping and evidence collection have something to attach to. ## Data classification **Confidential, potentially privileged.** An obligation register is a list of what the organisation is bound to do, and by implication where it may fall short. Assume a regulator, counterparty or claimant may one day see it, and take a view with counsel on whether it should sit under privilege. **STOP CONDITIONS:** | Trigger | Action | | --- | --- | | The instrument or the work relates to a live investigation, enforcement action, skilled-person review or litigation | Stop. Privilege and disclosure decisions come first. Route to counsel. | | Counsel's advice, or without-prejudice correspondence, is pasted | Stop. Do not summarise or excerpt privileged material. | | Customer, client or employee personal data appears in the source or as an illustration | Stop. Ask for it to be removed; extraction works on provisions, not records. | | Client positions, balances, transactions or account identifiers appear | Stop. Name the category, not the value. | | The user asks whether the organisation currently complies, or is in breach | Stop short of that conclusion. The register states the duty and the evidence that would show compliance; whether it is met is control mapping, and breach determination belongs to counsel. | | A confidential third-party agreement is supplied without confirmation it may be processed | Ask before extracting. Many agreements restrict disclosure of their own terms. | **NON-FABRICATION RULE.** Every row must be derived from supplied text and cite the provision it came from. Never supply a deadline, threshold, frequency, retention period, notice period or monetary limit from memory, from a summary, from a law-firm briefing, from an earlier version of the instrument, or from the equivalent rule in another jurisdiction. Never convert a discretion into a duty because it is prudent to treat it as one — record the modality the text uses and flag the judgement separately. Never invent a defined term's meaning; if the instrument defines it elsewhere and that part was not supplied, the term is open. If the provision text was not supplied, the output is `SOURCE TEXT NOT SUPPLIED — no obligations can be extracted`, and nothing else. Partial source is a partial register and must be labelled as one: state which sections, schedules, annexes or incorporated documents were **not** supplied. A register that looks complete and is not is the failure mode this skill exists to prevent. ## Prerequisites - **Inputs (mandatory):** the operative text — provisions, clauses, schedules, annexes. Not a summary, not a marketing overview, not a press release. - **Inputs (mandatory):** which instrument version and as at what date, and whether it is in force, transitional, or in consultation. - **Inputs (mandatory for a regulatory instrument):** which of your applicable regulator(s) and jurisdiction(s) the instrument belongs to, and which entity or entities in the group it is being extracted for. **Ask.** The same activity is regulated differently by different authorities and a register attached to the wrong entity is worse than none. - **Inputs (mandatory for a contract):** the parties and their defined roles, the effective and expiry dates, and any documents incorporated by reference. - **Inputs (needed for the deadline column):** the dates the instrument keys off — in-force, first reporting date, transitional end, notice periods, term dates. - **Inputs (if extracting from a PDF or scan):** confirmation the text layer is reliable. If pages are images or the extraction is lossy, say which pages and do not extract from them. - **Access:** none required. If a mandatory input is missing, name it and stop. ## Procedure ### 1. Inventory the source before extracting List every section, schedule, annex and incorporated document, and mark each supplied or not supplied. Record the defined-terms section separately — obligations turn on definitions more often than on the operative words, and a scope term left undefined makes every row that uses it provisional. ### 2. Extract atomically — one testable duty per row Split on the duty, not on the sentence. A single clause commonly contains several: notify, notify **within a period**, notify **in a specified form**, and retain evidence of having notified. Each is separately testable, separately capable of failing, and separately evidenced. If a row cannot be answered "done / not done" on its own, it is not yet atomic. The register columns: | Column | Content | Failure if omitted | | --- | --- | --- | | ID | Stable reference for the row | Nothing downstream can cite it | | Provision reference | Section, clause, paragraph, page — precise enough to find in seconds | The register stops being auditable | | Verbatim text | The operative words, quoted | Paraphrase drift; the whole point of the register | | Obligated party | Who is bound: the entity, a specific role, a counterparty, or both | Duties get tracked against the wrong side | | Beneficiary | Who it is owed to — regulator, client, counterparty, employee | Notification routes are guessed later | | Modality | Must / must not / should / may — as the text words it | Discretions become duties and duties become options | | Trigger | The event or state that makes it bite; standing obligations marked "continuous" | Everything looks always-on and nothing is prioritised | | Deadline or frequency | The period, and **the date it runs from**, quoted | A deadline with no start date is untrackable | | Conditions and carve-outs | Thresholds, exemptions, proportionality, transitional relief | Applies rules to entities they do not bind | | Evidence of compliance | What a reviewer would need to see: a record, an approval, a report, a log, a signed document | The register cannot be assured against | | Consequence | The stated sanction, remedy or termination right, if the text states one | Prioritisation becomes arbitrary | | Dependencies | Other rows this one presumes | Sequencing errors in remediation | | Confidence | Direct / interpretive / open | Silently mixes reading with reasoning | **Confidence values.** *Direct* — the text states it. *Interpretive* — the duty follows from the text but needs a stated reading, which must be written out in the row. *Open* — the text does not settle it. Open rows are questions for counsel, not gaps to be closed by inference. Report the count of each; a register that is 100% direct on a complex instrument has probably lost the hard parts. ### 3. Test applicability at row level, not document level An instrument applying to the organisation does not mean every provision does. For each row, record whether it applies to the entity in scope and why — the permission, activity, threshold, client category or contractual role that brings it in. Where applicability is uncertain, mark it open. Do not delete non-applicable rows: record them as not applicable with the reason, because that reasoning is exactly what is challenged later when the business changes. ### 4. Normalise deadlines into something trackable Convert each period into: the trigger event · the duration and its unit · whether days are calendar or business as the text specifies · the resulting due date or recurrence. Where the text does not specify calendar or business days, that is an open row, not an assumption. Recurring obligations get their frequency and their first occurrence. ### 5. Cross-check for the classic omissions Sweep specifically for duties that hide outside the operative clauses: record retention and its period; the duty to notify a change of circumstance; training and competence; governance and approval requirements; reporting and its form; audit, inspection and access rights; sub-contracting and onward flow-down; exit, termination and transition assistance; and duties that survive termination. Flow-down and survival are the two most commonly missed in contract extraction. ### 6. Quality-check the register before issue Does every row cite a provision? Can every row be answered done / not done? Does every deadline have a start date? Is every defined term used in a row actually defined in the supplied text? Is the not-supplied list stated at the top of the output rather than in a footnote? Is the count of open rows stated? If the answer to any of these is no, the register is not ready to hand on. ### 7. Hand the register on — it is an input, not a deliverable The register states duties. It does not state whether they are met. Control mapping, gap assessment and remediation planning are the next step and belong to `legal-compliance-regulatory-change-impact`, which consumes this register directly and cannot be run properly without one. ## Boundaries - **Not for** assessing impact, mapping obligations to existing controls, or building a remediation plan — that is `legal-compliance-regulatory-change-impact`. That skill's applicability test and gap assessment start from the register this one produces; run this first when it has no register, and route to it immediately afterwards. - **Not for** negotiating, redlining or advising on the acceptability of contract terms — that is `legal-compliance-contract-review`. - **Not for** writing the internal policy or standard that implements an obligation — that is `legal-compliance-policy-drafting`. - **Not for** answering a counterparty's questions about the controls that satisfy these obligations — that is `legal-compliance-questionnaire-response`. - **Not for** giving legal advice, opining on interpretation where the text is ambiguous, or concluding on compliance or breach. Open rows go to counsel. ## Hand-offs - **Receives from:** the compliance monitoring function, contract owners, and `legal-compliance-regulatory-change-impact` when that skill is invoked without a register to work from. - **Routes to:** `legal-compliance-regulatory-change-impact` — the register is its mandatory input for applicability testing, control mapping and remediation planning. - **Routes to:** `legal-compliance-policy-drafting` where an obligation needs an internal policy to implement it. - **Routes to:** `legal-compliance-contract-review` where extraction surfaces a term that needs renegotiating rather than tracking. - **Routes to:** `operations-sop-authoring` where an obligation needs a repeatable operational procedure and evidence trail. - **Escalates to:** counsel for every open row and every interpretive row whose stated reading carries material consequence.