Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Mandatory external verification workflow for ALL non-trivial factual claims before presenting them as true. This skill MUST be used whenever Claude is asked to research any topic, answer factual questions, provide current information, draft documents containing factual claims, give legal advice or cite legal authority, discuss current events or public figures, provide technical or scientific information, state statistics or data points, or answer ANY question where the answer could have changed
.claude/skills/lawve-ai-mandatory-verification/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-10 | ✗→✓ | ▲ Improved | 1743% | 0% |
| case-18 | ✓→✗ | ▼ Worse | 251% | 0% |
| case-19 | ✓→✗ | ▼ Worse | 107% | 0% |
| case-20 | ✓→✗ | ▼ Worse | 197% | 0% |
| case-01 | ✗→✗ | = Same ✗ | -24% | 0% |
This skill exists because Claude's internal knowledge, while broad, can be outdated, imprecise, or simply wrong on specifics. Users relying on Claude's outputs for professional, legal, academic, or decision-making purposes need factual accuracy they can trust. The only way to provide that is to verify claims externally before stating them.
This skill overrides Claude's default tendency to answer from memory. Internal knowledge is a starting point for identifying what needs checking — it is never sufficient to establish truth.
This skill applies to ANY task where factual accuracy matters:
If the response includes non-trivial factual claims → This skill is MANDATORY
The only exception is purely stable background knowledge that is definitional or conceptual in nature (e.g., "the Pythagorean theorem states..." or "a contract requires offer and acceptance"). Even for these, if the user's context suggests precision matters, verify.
Internal knowledge may be used only to:
Internal knowledge is never sufficient to establish the truth of a proposition. No material factual claim should be presented as accurate unless it has been externally verified. If external verification is not possible, the proposition must be clearly marked as unverified or uncertain.
Before drafting any answer, break the task into individual factual propositions and classify each:
For every proposition classified as time-sensitive, legal/technical/high-risk, or where precision matters:
Only after completing Phase 2:
The reason inline citations matter is that the user needs to be able to trace any individual claim back to its source instantly. If citations are grouped at the end, the reader cannot tell which source supports which claim — and that defeats the purpose of citing at all.
Good — citation immediately follows the claim:
> The current Bank of England base rate is 4.5% (Bank of England, Monetary Policy Summary, February 2025). The MPC voted 7–2 to hold the rate at this level, with two members preferring a cut to 4.25%.
> Under section 6 of the Data Protection Act 2018 (legislation.gov.uk), a "controller" is defined as the person who determines the purposes and means of processing.
Bad — citations grouped away from claims:
> The current base rate is 4.5%. The MPC voted 7–2 to hold. Two members preferred a cut. > > Sources: Bank of England, February 2025; Financial Times, 6 Feb 2025.
In the bad example, the reader cannot tell which source supports the rate figure vs. the vote split vs. the dissent.
Every material claim in the response should carry one of these labels (either explicitly stated or clearly implied by context):
This labelling matters because it lets the reader instantly assess how much weight to give each claim. A verified current rate and an analyst's prediction deserve very different levels of trust, and the response should make that distinction obvious.
Prefer sources in this order:
Do not rely on summaries where a primary source is available. Do not cite commentary for propositions that should be supported by primary authority.
When the task involves legal content, apply these additional requirements (this complements the legal-citation-verification skill — use both together for legal tasks):
Citations must be:
Include where applicable:
If a claim is inferred rather than directly stated in a source, label it: "Based on source], it appears that..." or "This suggests..." rather than stating it as established fact.
A consolidated sources list at the end of the response is acceptable as an additional convenience for the reader, but it does not replace inline citations. The inline citation is the primary mechanism; the end-of-response list is supplementary.
These exist because Claude has a well-documented tendency to generate plausible-sounding but fabricated citations, case names, and factual details. The following rules specifically counteract this:
When a proposition cannot be externally verified:
At the end of every response, include a Verification Notes section. This serves two purposes: it shows the user exactly what was checked (building trust), and it forces a final self-audit of whether each claim has actually been verified.
The verification notes should include:
Example: > Verification Notes > > Propositions verified: > - Current base rate (4.5%) — verified via Bank of England MPC decision, Feb 2025 > - Last rate change (Nov 2024, cut from 4.75%) — verified via BoE historical data > > Uncertainties flagged: > - Rate direction beyond Q2 2025 — inherently unpredictable; analyst forecasts cited but labelled as forward-looking > > Primary sources: Bank of England (bankofengland.co.uk) > > Current as of: date]. Next MPC decision: date] — position may change after this.
This section is not optional. It is the user's assurance that the verification workflow was actually followed.
Before finalising any response:
This skill works alongside and complements:
Accuracy takes precedence over speed.
It is better to:
Than to:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 24,636 | 6,320 | -74% | 1 | 1 | 0% | 4,447 | 3,366 | -24% | 0 | 0 | — |
case-02 | fail→fail | 11,211 | 29,774 | +166% | 1 | 1 | 0% | 2,204 | 3,451 | +57% | 0 | 0 | — |
case-03 | fail→fail | 21,419 | 9,620 | -55% | 1 | 1 | 0% | 3,602 | 3,747 | +4% | 0 | 0 | — |
case-04 | fail→fail | 8,453 | 6,079 | -28% | 1 | 1 | 0% | 1,553 | 3,190 | +105% | 0 | 0 | — |
case-05 | fail→fail | 15,429 | 7,781 | -50% | 1 | 1 | 0% | 2,654 | 3,482 | +31% | 0 | 0 | — |
case-06 | fail→fail | 9,483 | 5,345 | -44% | 1 | 1 | 0% | 1,550 | 3,236 | +109% | 0 | 0 | — |
case-07 | fail→fail | 11,164 | 7,238 | -35% | 1 | 1 | 0% | 1,643 | 3,430 | +109% | 0 | 0 | — |
case-08 | fail→fail | 8,254 | 6,703 | -19% | 1 | 1 | 0% | 1,440 | 3,291 | +129% | 0 | 0 | — |
case-09 | fail→fail | 14,693 | 32,115 | +119% | 1 | 1 | 0% | 2,178 | 7,708 | +254% | 0 | 0 | — |
case-10 | fail→pass | 1,688 | 6,434 | +281% | 1 | 1 | 0% | 189 | 3,483 | +1743% | 0 | 0 | — |
case-11 | fail→fail | 14,070 | 7,085 | -50% | 1 | 1 | 0% | 2,155 | 3,336 | +55% | 0 | 0 | — |
case-12 | fail→fail | 6,986 | 8,547 | +22% | 1 | 1 | 0% | 1,068 | 3,426 | +221% | 0 | 0 | — |
case-13 | fail→fail | 14,857 | 6,229 | -58% | 1 | 1 | 0% | 2,296 | 3,224 | +40% | 0 | 0 | — |
case-14 | fail→fail | 4,818 | 6,568 | +36% | 1 | 1 | 0% | 787 | 3,240 | +312% | 0 | 0 | — |
case-15 | fail→fail | 7,951 | 5,413 | -32% | 1 | 1 | 0% | 1,302 | 3,261 | +150% | 0 | 0 | — |
case-16 | fail→fail | 8,653 | 5,871 | -32% | 1 | 1 | 0% | 1,267 | 3,285 | +159% | 0 | 0 | — |
case-17 | fail→fail | 15,900 | 7,906 | -50% | 1 | 1 | 0% | 2,700 | 3,381 | +25% | 0 | 0 | — |
case-18 | pass→fail | 8,020 | 11,405 | +42% | 1 | 1 | 0% | 1,214 | 4,257 | +251% | 0 | 0 | — |
case-19 | pass→fail | 12,146 | 7,009 | -42% | 1 | 1 | 0% | 1,950 | 4,041 | +107% | 0 | 0 | — |
case-20 | pass→fail | 8,918 | 6,971 | -22% | 1 | 1 | 0% | 1,355 | 4,020 | +197% | 0 | 0 | — |
case-21 | fail→fail | 10,868 | 7,124 | -34% | 1 | 1 | 0% | 1,804 | 3,325 | +84% | 0 | 0 | — |
case-22 | fail→fail | 15,605 | 8,338 | -47% | 1 | 1 | 0% | 2,383 | 3,330 | +40% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 4 counted toward the lift figure. The other 18 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of -100 percentage points is the difference between those two pass rates over the 4 comparable cases. 3 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.