<kebab Case Skill Slug>

Skill:

Authoring rule (binding). This template is the framing-first (v2) scaffold. Write only procedural guidance, legal frameworks, analytical scaffolds, output-shape conventions, and role-specific guards. Do not preload subject-specific dollar amounts, percentages, share calculations, verbatim short quotations from the parties' internal documents, named subsidiaries / contracts / JVs / precedents, or any rubric-derived "must-name" checklist. The agent must derive those from documents_text/ itself. See this repository's README → Authoring principle: framing first, not rubric reverse-engineering for the empirical basis and the leakage A/B protocol.

Class-applicability rule (SkillsBench-aligned). Although each SKILL.md is routed by a single task_id, its content must read as guidance for the class of matters this task represents, not as a cheat sheet for this one instance. Concretely:

• Rubric-blind test. Ask: "If I were a senior practitioner writing a CLE module or internal practice note for this kind of matter, having never seen this scenario's source documents or the hidden rubric, would I write this sentence?" If the answer is no — if the sentence only makes sense because you know which jurisdiction / statute / entity type / counterparty appears in this scenario — cut it or generalize it to the category.
• 3-variant test. Ask: "Would this same SKILL.md still be useful if the parties, jurisdictions, dollar amounts, and counsel were swapped for three other plausible variants of the same matter type?" If swapping a US-Delaware borrower for a UK-Cayman borrower (or a $500M facility for a $50M facility) would break the skill, the skill is over-fit to this instance.
• Statute / jurisdiction handling. Naming the category of authority is fine ("apply the borrower's jurisdiction-of-incorporation corporate statute"); naming the specific statute, code section, provincial act, or regulator that happens to govern this scenario ("Ohio Rev. Code § 1701.58", "Alberta ULC", "Ontario EBCA registration") is leakage if you only know to mention it because you read the documents. Generalize unless the statute is genuinely load-bearing for every plausible variant of this task class.

Delete any section below that does not apply to this task. Keep section headings concrete and short.

1. Subject-matter triage (when applicable)

Use only if the task's documents/ plausibly contains material from more than one matter / transaction / counterparty / period, and the baseline failure mode includes the agent writing about the wrong one. Otherwise delete this section.

<Describe the procedure the agent should run to identify the single in-scope subject before drafting — e.g., "walk every file under documents_text/, group by counterparty / signing date / advisor of record, and identify the single cluster whose deal-summary / engagement letter / board minutes matches the task instructions." Mark every other cluster as out-of-scope background.

Do not name which specific cluster is the subject (the agent should infer that from instructions + documents). Do tell it how to decide.

If the pipeline includes a separate planner / checker step, the planner should write the triage outcome into its task contract / notes (subject cluster + explicit list of out-of-scope files), and the checker should grep the final deliverable for any out-of-scope entity names that would indicate the deliverable migrated into the wrong matter. Phrase the rule once here; do not split into per-role sub-sections.>

2. Failure modes the skill is correcting

2–4 bullets naming the structural / procedural patterns the baseline tends to drop — not the specific facts. Source: reading the baseline run's draft deliverable (you may inspect it). Do not read the hidden rubric, prior-run scores, or evaluation traces for this task.

• <e.g., "Baseline paraphrases pricing / market-power language from internal communications into its own words instead of preserving the source phrase.">
• <e.g., "Baseline stops at national-scale market analysis and does not walk state- or MSA-scale overlap.">
• <e.g., "Baseline collapses a comparison table into prose, losing the row-level structure the deliverable type conventionally requires.">

3. Legal frameworks / domain conventions that apply

What statutory / regulatory / industry framework the deliverable conventionally sits in. This is general practice knowledge an expert would write rubric-blind. Naming a regulation, an Act, a guideline, or a Form is fine; naming the specific dollar threshold the rubric happens to be checking for this filing year is not.

• <e.g., "HSR pre-merger notification mechanics: identify Ultimate Parent Entity per 16 C.F.R. § 801.1, name the Acquired Person, state the size-of-transaction threshold for the filing year named in the source documents (do not import a different year's number), and the reviewing agency under the FTC-DOJ interagency clearance protocol.">
• <e.g., "If the matter involves a JV between competitors, raise Section 8 of the Clayton Act interlocking-directorate analysis.">
• <e.g., "If the buyer is a private-equity-sponsored platform that has assembled the business through multiple add-ons, walk the 2023 DOJ/FTC Merger Guidelines Guideline 8 serial-acquisition theory and FTC heightened scrutiny of PE-backed consolidation in the relevant industry.">

4. Analytical scaffolds

What scales / granularities / dimensions the analysis must walk to be defensible. This is procedure, not answer-feeding.

• <e.g., "Walk geographic market at every scale: national for context, then each state with overlap individually (do not lump states), then the most-concentrated MSA. Identify the local competitors and their shares to contextualize the combined position. Treat the narrow MSA as a plausible agency-defined geographic market.">
• <e.g., "Walk product market at both broad (industry-wide) and narrow (service-line) granularity.">
• <e.g., "For every share figure, show the arithmetic (numerator / denominator = percentage). All numbers must come from the subject's source documents, not be invented.">
• <e.g., "Surface verbatim short quotations from internal documents with document name and date. Do not paraphrase pricing, market-power, or competitive-elimination language into your own words.">

5. Vertical / structural / temporal relationships (when applicable)

If the task has structural elements the analysis must address (org chart, exclusive customer contracts, subsidiary supply relationships, JVs, timeline tensions, lifecycle events), name the categories here, not the specific instances.

• <e.g., "Exclusive customer contracts the target holds — analyze each as a foreclosure / entry-barrier issue with customer, scope, and revenue from the source documents.">
• <e.g., "Subsidiaries that supply inputs or platform services to *competing* independent operators — raise the post-merger raise-price / refuse-to-supply foreclosure question for each.">
• <e.g., "Timing tensions between contractual deadlines (renewals, outside dates, closing windows) and regulatory review windows.">

6. Output structure conventions

The conventional shape an expert deliverable of this type takes. List standard section names an industry practitioner would use; do not copy the rubric's section list or its minimum counts verbatim.

• <e.g., "Organize as a formal client memorandum with named sections. Conventional sections for an HSR pre-notification briefing paper: Executive Summary, HSR Filing Requirements, Market Definition, Competitive Effects, Timeline / Process Risks, Pre-Filing Strategy, Risk Rating and Conclusion.">
• <e.g., "Use an explicit risk-rating scale (Low / Moderate / High / Very High) with supporting rationale, not a single qualitative adjective.">
• <e.g., "If the deliverable is an issue list, keep one row per issue with severity, source citation, recommended action, and deadline — do not collapse to prose.">

Forbidden in this body (anti-leakage checklist)

Before committing, grep your draft for these patterns; any hit is a sign you have crossed from framing into rubric reverse-engineering or from class-level practice guidance into instance-level cheat sheet:

• Specific subject-cluster dollar amounts, percentages, share calculations, ratios, or reconciliation arithmetic.
• Specific verbatim short quotes from the subject cluster's internal documents (the general instruction "surface verbatim quotes from internal documents" is OK; pasting the actual phrase the agent must surface is not).
• Specific named subsidiaries, contracts, JVs, counsel, advisors, precedents, or competitors from the subject cluster.
• Jurisdiction- / statute- / entity-type-specific citations you only know to mention because you read this scenario's documents (e.g., pre-loading "Ohio Rev. Code § 1701.58 director-proxy prohibition" or "Alberta ULC shareholder-resolution requirement" or "Ontario EBCA registration" when the task class — "review corporate authorization documents for a multi-jurisdiction guarantor group" — does not by itself imply those specific authorities). Write the category ("apply each entity's jurisdiction-of-incorporation corporate statute, including any director-proxy prohibition that jurisdiction imposes"), let the agent name the instance from the documents.
• Task-specific filenames, paths, document titles, or matter identifiers from the subject cluster (SkillsBench R1).
• The exact deliverable section list copied from the rubric's presence-check criterion (use the industry-conventional shape described in section 6 instead).
• Minimum counts the rubric is testing for (e.g., "≥ 12 distinct issues") — let the analysis dictate the count from the documents.
• Any sentence whose specificity is only justifiable by having read the hidden rubric, prior-run scores, evaluation traces, or the scenario's source documents. If the same sentence would not appear in a rubric-blind CLE module on this kind of matter, cut it.
• Any sentence that would stop being true / useful if the parties, jurisdictions, dollar amounts, and counsel in this scenario were swapped for three other plausible variants of the same matter type (class-applicability / 3-variant test).