Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when reviewing or negotiating an NDA / confidentiality agreement from one party's side (the receiving party or the disclosing party) and you need an issue log with a preferred redline and a fallback for each clause. Applies the asymmetric positions each side takes on scope of Confidential Information, term and survival, residuals, non-solicitation, injunctive relief, and return-or-destroy — which direction to push and where to settle. Not for drafting an NDA from a blank page or for extracting clauses into a data schema.
.claude/skills/nda-review-issue-log/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | 583% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 599% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 591% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 587% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 606% | 0% |
Review a non-disclosure agreement from one side. The same clause is favorable or dangerous depending on which party you represent, so the first move is always to fix the side, then walk the clauses. For each issue produce two positions, never one:
your side's core interest.
Output an issue log: one entry per clause, each naming the issue, your Preferred redline, and your Fallback. A single-position answer ("make the definition clearer") is a miss — the value is the directional ask plus a pre-negotiated retreat.
The two sides pull in opposite directions on every issue:
escapable obligations: a tight definition, a short survival period, room to use general know-how, no covenants smuggled in, no waiver of the other side's burden of proof.
airtight protection: everything covered, obligations that outlast the deal, no residuals loophole, strong remedies, prompt return of everything.
Mutual NDAs blend both; when both sides will genuinely share, a Recipient of a one-way NDA should ask to make it mutual so the same protections cut both ways.
as confidential (and, for oral disclosures, confirmed in writing within a set window). Keep the standard carve-outs: already public, already known to the recipient, independently developed without use of the disclosed information, or rightfully obtained from a third party.
only if the standard carve-outs survive intact and marking is still required where practical.
marked or designated; no marking obligation, since designation is easy to forget.
anything a reasonable person would treat as confidential given the context.
Distinguish the term of the agreement (how long you can still disclose under it) from the survival of confidentiality obligations (how long the duty to protect lasts after disclosure).
clock — e.g. 2–3 years after each disclosure. Avoid perpetual survival.
period (3–5 years) for all other confidential information.
secrets with a long fixed tail for everything else.
Recipient red flag: perpetual survival applied to all confidential information (not just trade secrets) is an open-ended liability — push it onto a clock.
A residuals clause lets people use general knowledge, ideas, and know-how retained in unaided memory, without breaching the NDA. It is a large loophole — good for the Recipient, dangerous for the Discloser.
unaided memory, with no duty to track or segregate what was learned.
deliberately memorized to get around the NDA.
memorization, and state expressly that it grants no license to patents or copyrights.
An NDA is the wrong home for a non-solicit or no-hire covenant. When one appears, treat it as smuggled scope, not boilerplate.
negotiated agreement, not a confidentiality document.
employees met in connection with the deal, cap it at 6–12 months, and carve out general job postings, recruiter-sourced candidates, and employees who apply on their own.
general-advertising carve-out.
Recipient red flag: any non-solicitation or no-hire language inside an NDA — flag it as out of place before negotiating its terms.
Discloser clauses often stipulate that any breach causes irreparable harm and that injunctive relief issues automatically, without proof of damages and without posting a bond.
stipulate irreparable harm, do not waive the bond, and preserve the other side's burden to prove its entitlement.
bond requirement (or leave it to the court's discretion).
without proof of damages, no bond, in addition to all other remedies.
court's discretion.
and a carve-out to retain one archival copy for legal/compliance plus automatic system backups, each kept subject to continuing confidentiality.
keeping the backup/archival carve-out.
written certification and no copies retained.
ongoing confidentiality duty.
but the facts show who mostly receives vs. discloses, infer it and say which side you took.
non-solicit / no-hire covenant living inside the NDA.
mutual is the cleaner fix.
A secondary catalog of less common clauses (assignment, no-warranty, no-license, marking mechanics, governing law/venue, term of permitted use) lives in references/issue-catalog.md — consult it only when a review reaches beyond the six core issues.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 12,069 | 2,345 | -81% | 1 | 1 | 0% | 300 | 2,032 | +577% | 0 | 0 | — |
case-02 | fail→pass | 4,667 | 3,321 | -29% | 1 | 1 | 0% | 297 | 2,029 | +583% | 0 | 0 | — |
case-03 | fail→pass | 3,868 | 13,355 | +245% | 1 | 1 | 0% | 289 | 2,021 | +599% | 0 | 0 | — |
case-04 | fail→pass | 3,780 | 2,251 | -40% | 1 | 1 | 0% | 293 | 2,025 | +591% | 0 | 0 | — |
case-05 | fail→fail | 13,017 | 14,230 | +9% | 1 | 1 | 0% | 306 | 1,786 | +484% | 0 | 0 | — |
case-06 | fail→fail | 4,197 | 2,150 | -49% | 1 | 1 | 0% | 284 | 2,016 | +610% | 0 | 0 | — |
case-07 | fail→pass | 4,247 | 2,357 | -45% | 1 | 1 | 0% | 295 | 2,027 | +587% | 0 | 0 | — |
case-08 | fail→fail | 2,379 | 10,852 | +356% | 1 | 1 | 0% | 286 | 2,018 | +606% | 0 | 0 | — |
case-09 | fail→fail | 20,185 | 4,839 | -76% | 1 | 1 | 0% | 288 | 2,020 | +601% | 0 | 0 | — |
case-10 | fail→fail | 3,176 | 4,950 | +56% | 1 | 1 | 0% | 304 | 2,036 | +570% | 0 | 0 | — |
case-11 | fail→pass | 11,572 | 2,039 | -82% | 1 | 1 | 0% | 286 | 2,018 | +606% | 0 | 0 | — |
case-12 | fail→pass | 2,998 | 6,223 | +108% | 1 | 1 | 0% | 294 | 2,026 | +589% | 0 | 0 | — |
case-13 | fail→fail | 2,478 | 3,459 | +40% | 1 | 1 | 0% | 286 | 2,018 | +606% | 0 | 0 | — |
case-14 | fail→fail | 28,316 | 2,245 | -92% | 1 | 1 | 0% | 307 | 2,039 | +564% | 0 | 0 | — |
case-15 | fail→pass | 4,449 | 23,476 | +428% | 1 | 1 | 0% | 300 | 2,032 | +577% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.5-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 15 cases were attempted, and 14 counted toward the lift figure. The other 1 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 +47 percentage points is the difference between those two pass rates over the 14 comparable cases.
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.