A Claude Skill that turns a Bates-indexed record into a working deposition outline: a dated chronology, topic blocks whose every evidentiary question carries a record cite, a list of admissions to lock in before the examination turns adversarial, and impeachment pairs built from the deponent’s own prior inconsistent statements. It also does the two things ad hoc outlines skip — it budgets the outline against the seven-hour clock, and it reports what the outline does not reach.
The skill is built around one constraint: it can only cite documents that exist in a record index you supply. It never composes a Bates number. This page covers when to run it, when not to, what a run costs, and the failure modes to account for before an outline drafted this way goes into a deposition.
When to use
Reach for the skill after the record is assembled and indexed, and before the examining attorney starts drafting questions.
- Witnesses whose prior statements are scattered. When the deponent’s own words sit across two prior transcripts, a Slack export, a declaration, and forty produced emails, assembling the impeachment material by hand is the expensive part of prep. The skill builds the statement inventory in one pass and the outline in a second.
- Depositions on a compressed schedule. Federal practice caps each side at 10 depositions without leave or stipulation under Rule 30(a)(2)(A)(i), so a case with a wide witness list gets scheduled tightly. A draft outline in minutes moves the attorney’s work from drafting to editing.
- Rule 30(b)(6) examinations. Supply the noticed matters and every block is tagged to the matter numbers it serves; untagged blocks are moved to a separate appendix rather than folded into the main outline, which is what keeps a designee’s counsel from shutting a line down mid-deposition.
- Second-chair and litigation-support staff drafting for review. The block library encodes the sequencing decisions — authenticity early, admissions before impeachment — that junior drafters get wrong.
The artifact bundle lives at apps/web/public/artifacts/deposition-outline-claude-skill/ and ships:
SKILL.md— the Skill definition with inputs, the two-pass method, a literal output example, and the watch-outsreferences/1-topic-block-library.md— question modules by block type and deponent role, each with the reason its questions are phrased as they arereferences/2-record-index-template.md— the fillable record index, and the only source of cites the skill will usereferences/3-impeachment-pair-worksheet.md— the lock/confront/introduce pair format plus the pre-deposition verification checklist
When NOT to use
- Deciding whom to depose, or in what order. Witness sequencing follows a case theory. The skill drafts for a deponent who has already been noticed; it has no view on whether noticing them was right.
- Drafting case theory or a strategy memo. Which admissions matter to a defense theory, and what a witness’s testimony does to your summary-judgment position, is work product. The skill produces an examination outline built from cited record material, not a privileged analysis.
- Any record that has not been reduced to stable identifiers. If your exhibits are filenames in a shared drive rather than Bates numbers, transcript page-and-line references, or production document IDs, stop and index first. Running against an unindexed record is how you get questions that cite documents nobody can produce at the table.
- Records already hosted in your review platform. If the production sits in Everlaw or another review platform with a deposition module, use the platform’s tool — see the alternatives section below.
- Thin records. The skill refuses to draft impeachment when fewer than three index rows resolve to the deponent’s own statements. That refusal is deliberate; three rows is the floor below which a “prior inconsistent statement” is usually a paraphrase, and paraphrase impeachment falls apart on the record.
Setup
- Download the bundle from
apps/web/public/artifacts/deposition-outline-claude-skill/and place it in~/.claude/skills/(Claude Code) or upload it to a Claude project. - Fill in
references/2-record-index-template.mdfor the deponent. Setproduction_throughto the last Bates number produced andproduction_through_dateto the date of that production — the skill prints both in the outline header and warns when the production date is more than 14 days before the run date. An index of 80 to 400 rows is the working range for a single fact witness; under 50 rows you are indexing a subset and the coverage report will be full of gaps you cannot explain. - Mark the
privilegedcolumnyeson every row that was clawed back, withheld, or logged. Those rows are excluded from question generation, and rows adjacent to them draw a[PRIVILEGE-CHECK]annotation in the output. - Edit
references/1-topic-block-library.mdfor your practice. The role modules ship generic phrasings; replace the system names, record types, and process language with your matter’s. Jurisdiction matters here too — several states have not adopted the federal impeachment sequence described below. - Run the skill once against a closed matter whose outline you already have, and diff the two. Check that every cite resolves and that the block ordering matches how the examination actually went.
- For each new deponent: supply the record index, the deponent’s role, and 3 to 8 examination goals in priority order. Review before anyone walks into a deposition with it.
What the skill actually does
Two passes, and the separation between them is the point.
Pass 1 — build the statement inventory. The skill reads the record index and produces a dated chronology plus an inventory of the deponent’s own statements, each row carrying its identifier verbatim. Nothing is drafted here. If a required column is missing, the skill emits a structured error and stops.
Pass 2 — assemble the blocks. Drafting sees the inventory, not the raw record. A question therefore cannot cite a document the extraction pass did not find, because the drafting pass has no access to anything outside the inventory. This is the structural guard against invented exhibit numbers, and it is why a single-pass “read these documents and write me questions” prompt is the wrong shape for this job — in one pass, generating a question and generating its citation are the same act.
Blocks come out in a fixed order: background and foundation, document authenticity, substantive topics in chronological order, admissions to lock in, then impeachment pairs. Admissions precede impeachment because a witness who has been confronted with a prior inconsistent statement becomes guarded and stops conceding. Authenticity comes early because an exhibit the witness will not authenticate changes which substantive questions are worth asking, and discovering that in hour five wastes everything built on it.
The time budget. Rule 30(d)(1) limits a deposition to one day of seven hours absent stipulation or court order. The skill estimates minutes per block against a 330-minute default for question time, holding roughly 90 minutes back for objections, exhibit handling, colloquy, and breaks. When the outline runs long it does not trim silently — it emits the full outline plus a ranked cut list ordered by the priority of the goals each block serves.
Impeachment pairs are sequenced. Each pair comes out as lock the testimony, confront with the prior statement, then introduce. Rule 613(b), amended effective 1 December 2024, provides that extrinsic evidence of a prior inconsistent statement is not admissible until the witness has had an opportunity to explain or deny it and an adverse party has had an opportunity to examine on it. The pre-amendment rule imposed no sequence at all, so an outline that just says “impeach with Exhibit 14” is one an examiner can work in an order that forfeits the impeachment.
The final section is a coverage report: index rows no question cites, examination goals no block serves, blocks carrying no exhibit, and the count of [UNCITED] questions. That is the section to read first.
Cost reality
Token cost is dominated by how much record you hand over. Running against the index rather than the underlying documents keeps a run small: a 250-row index plus the block library lands around 8,000 to 20,000 input tokens, with 3,000 to 6,000 tokens of output.
At Claude Sonnet 5 list pricing ($3 per million input tokens, $15 per million output), that is roughly $0.07 to $0.15 per outline. On Claude Opus 5 ($5 and $25 per million), roughly $0.12 to $0.25.
Feeding raw transcripts instead of an index changes the arithmetic by an order of magnitude. A 300-page transcript runs about 105,000 tokens (estimate, at roughly 350 tokens per transcript page), so two prior transcripts plus exhibit text puts a single run near 250,000 input tokens — about $0.75 on Sonnet 5, $1.25 on Opus 5. Still small in absolute terms, but you are paying it for a worse result, because the model is then extracting and drafting in one step and the citation guard no longer holds.
The cost that matters is attorney time. Drafting an outline from a mid-sized record is a 6-to-10-hour job (estimate, based on a 200-to-400-document record and one prior transcript); reviewing and editing a generated draft against the index runs 60 to 90 minutes. That review is not optional and does not shrink — the skill replaces drafting, not judgment.
Success metric
Two numbers worth tracking from the first deposition.
- Cite resolution rate. Before the deposition, pull a sample of 20 cited questions and check each cite against the production. Anything below 100% means the record index is stale or wrong, not that the model needs better prompting — re-index and re-run. Track the
[UNCITED]count in the coverage report as the leading indicator: a rising count means the block library is asking about things your index does not cover. - Admission conversion rate. After the deposition, count how many of the drafted admissions actually got a yes or a qualified yes on the record. Below about half, the admissions are being drafted too broadly — they are asking the witness to agree with a characterization instead of a fact, and the fix is in the admissions block of
references/1-topic-block-library.md, not in the model.
vs alternatives
vs an associate drafting from scratch. Six to ten hours of drafting for a mid-sized record, and the predictable misses are structural rather than careless: impeachment material buried in a second production never makes the outline, admissions get asked after the first impeachment lands, and nobody budgets the seven hours until the deposition is already running long. The skill’s advantage is not writing speed. It is that the chronology, the statement inventory, and the time budget get produced every time instead of when there is time for them.
vs Thomson Reuters CoCounsel “Prepare for a Deposition.” CoCounsel takes the deponent’s prior statements and case documents and returns a topic-organized outline with suggested questions; the capability sits in the CoCounsel Essentials feature set, and Thomson Reuters publishes no per-seat price on its plans page (checked 19 August 2026). If your firm already holds CoCounsel seats, run that first — it is bundled, it is integrated with Westlaw, and the setup cost is zero. Use this skill when you do not hold seats, when the record spans sources CoCounsel is not holding, or when you want the outline to carry a coverage report and a time budget, which CoCounsel’s output does not.
vs Everlaw Deposition Analyzer and Storybuilder. Everlaw’s deposition workspace drafts questions, generates deponent backgrounds, summarizes testimony, and flags inconsistencies with exhibits linked inside the review platform. When the production already lives in Everlaw, that is the better tool by a clear margin: cites resolve to platform document IDs natively, so the fabrication problem this skill solves with a two-pass index does not arise. Pricing is quote-gated. The rule: if the record is hosted in your review platform, use the platform’s module; use this skill when the record is scattered across produced PDFs, prior transcripts, and third-party files, or when the matter is not hosted at all.
Watch-outs
- Fabricated record cites. A model asked to write questions about documents will produce document numbers that look right. Guard: cites are copied verbatim from index rows and never composed; any question the skill cannot tie to a row is emitted with an
[UNCITED]marker and counted in the coverage report rather than given a plausible number. Verify a sample anyway — as of 19 August 2026 the AI Hallucination Cases database maintained by Damien Charlotin lists 1,933 decisions worldwide, 1,324 of them in the United States, where a court found a party relied on hallucinated material. - Impeachment sequenced wrongly. Introducing extrinsic evidence before confronting the witness forfeits the impeachment under amended Rule 613(b). Guard: pairs are emitted only in the lock/confront/introduce format, and the introduce step carries an explicit instruction not to offer the exhibit until the confrontation question has been answered. The worksheet also flags that the federal sequence is not universal — check your jurisdiction before relying on it.
- Overrunning the clock. Guard: per-block minute estimates, the 330-minute default budget, and a cut list ranked by goal priority rather than block order, so the topics that get dropped are the ones you chose to drop.
- Privileged material pulled into the examination. A record index built from a production can contain clawed-back documents. Guard: rows flagged
privilegedare excluded from question generation, and blocks touching adjacent rows carry a[PRIVILEGE-CHECK]annotation naming the row. The flag is only as good as the index — a clawback that never made it into the spreadsheet is invisible to the skill. - A stale index against a rolling production. Guard:
production_throughand its date are required header fields, printed in the outline header, with a warning when the production date is more than 14 days before the run. Re-index rather than patching cites by hand. - Treating an instruction not to answer as an obstacle to route around. Rule 30(c)(2) permits an instruction not to answer only to preserve a privilege, to enforce a court-ordered limitation, or to present a Rule 30(d)(3) motion. Guard: the skill does not draft workaround questions for anticipated instructions; it flags the block and leaves the response to counsel on the record.
Stack
- Claude — statement inventory extraction, block assembly, time budgeting, coverage reporting
- Everlaw — review platform for hosted records, and the better path when the production already lives there
- Thomson Reuters CoCounsel — the bundled alternative for firms already holding seats
Related: privilege review batching for the review phase this outline draws its record from, and the e-discovery custodian questionnaire for the intake phase that determines whether the deponent’s data sources are in the production at all.