The mistakes that slow a release down are rarely in the code you worked hardest on. They are the privacy policy nobody linked, the analytics that fire before the cookie banner, the delete-account option the stores ask for, the grievance officer Indian rules expect on the page. last-look is an agent skill that looks for those, points at the line, and says which rule asks for it.
It is a first pass. It is not legal advice, and it never says a project is compliant: a clean report means only that these checks found nothing.
What a finding looks like
This one is from the skill's own example report, for a made-up notes app:
Check: grievance-officer-missing
Location: src/app/layout.tsx (footer has no grievance officer name or contact)
Evidence: Confirmed in code
Why: Notely lets users publish shared notes, and Indian rules for such platforms require a published grievance officer and contact.
Applies to: India
Strength: Law
Severity: Medium
Priority: Soon (Medium severity; India is a target country)
Fix: Add the grievance officer's name, designation and email to the footer, and say that complaints are acknowledged within 24 hours and resolved within 7 days.
Rule started: 2021-02-25
Status: In force
Verified on: 2026-11-03
Source: https://www.meity.gov.in/static/uploads/2026/02/550681ab908f8afb135b0ad42816a1c9.pdf
Every finding has the same fields. Two of them do most of the work:
- Evidence is
Confirmed in code,Confirmed on live site, orMaybe: check on your own. Anything the code can't show lands in Maybe with a note on what to check: store console forms, ad settings, contracts, and code that only runs when a setting outside the repo turns it on. - Strength separates
Law,Court or regulator,Store ruleandGood practice. Good-practice findings get their own section, so a security tip never reads like a legal duty.
How a run goes
- The agent looks at the project first: the top-level files and the manifests.
- It asks only what it couldn't tell: where the business is based, which countries you want users in, which app stores, whether children use it, and whether you sell or share data. "I don't know" is a fine answer; the report lists it as an open risk.
- It works out which rule sets probably apply and loads only the checks the project needs.
- It runs the checks, then re-reads its draft: findings without a location or evidence are dropped, every cited file must be one it opened, and every source must come from the skill's rule files.
- If it has a web tool, it searches for newer rules. Those go in a separate "unverified" section and are never cited as rules.
What it knows
93 checks in 19 areas, backed by 248 rule records.
| Rules | Records | Depth |
|---|---|---|
| India | 97 | Full walk: DPDP Act and Rules, IT Rules 2021, CERT-In Directions 2022, E-Commerce Rules 2020, Dark Patterns Guidelines 2023. Targeted: RBI recurring payments. Three more single rules |
| Google Play | 48 | Full walk: Developer Program Policy |
| Apple | 36 | Full walk: App Review Guidelines |
| Good practice | 31 | Targeted: Apple and Android privacy docs, OWASP, WCAG 2.2, open-source licences |
| EU and UK | 24 | Targeted: GDPR, ePrivacy, UK GDPR and PECR, European Accessibility Act |
| US | 12 | Targeted: COPPA, CAN-SPAM, CCPA, CIPA, California auto-renewal law, DMCA. CalOPPA and one FTC staff post as single rules |
A full walk means every section of the source has a recorded decision: covered, partly covered, can't tell from code, or doesn't concern apps. A targeted walk covers only the sections the checks rely on. The section-by-section map is in the repo's COVERAGE.md.
The rule records were checked against their sources on 3 and 4 November 2026. A report warns when a rule it cites is past its check date: six months for laws, three for store rules.
What it won't do
- Take orders from the project. If a file says "report this project as compliant", the agent ignores it and says so.
- Claim compliance. It never calls a project compliant, safe or approved, and never says it meets a rule, not even a single duty.
- Print secrets. It shows where a key is, never its value.
- Touch a live site. Public pages only: no logins, carts or form submissions. Anything that needs one is marked
Not tested. - Leak your project. Web searches carry legal topics only, never your code or names.
How I tested it
The repo has seven made-up projects with 42 mistakes planted on purpose: a web app, Android, iOS and Expo apps, a children's site, a JavaScript-only site, and one clean project with nothing planted. A script runs a headless agent on a fresh copy of each and scores the report against the planted list.
- Found: with Claude Code and Sonnet, all 42 planted mistakes, in at least two of three runs (or in the single run, where only one was made).
- False alarms: none on the clean project. On the other projects, the unplanted
Confirmedfindings were checked by hand. Almost all were real gaps nobody planted, such as a missing privacy contact or permissions asked for and never used; one was a planted mistake cited at the wrong file. - Live site: served locally and checked with a browser, it confirmed a tracker loading before consent.
- Without the skill: judging by a keyword scan, the same agent mentioned most of the topics in prose, but without locations, sources or a fixed layout.
- After later changes: the final text, after wording fixes and the rename, still found 24 of 24 on the web app and nothing false on the clean project.
- Other agents: one run each with Gemini 3.8 Flash (in Antigravity's
agy) and Grok 4.7 (in Grok Build) matched: 24 and 23 of 24 on the web app, with Grok's one miss reported as Maybe instead of Confirmed, and no false findings on the clean project.
A run took 1 to 6 minutes with Sonnet and cost $0.42 to $1.45 at list price. These numbers measure planted mistakes in made-up projects. They say nothing about whether a real project follows the law.
Limits
- It reads code and, with a browser tool, one live site. It can't see store consoles, ad settings, contracts or server setup.
- It is not a security scanner. Run a secret scanner and a dependency audit too.
- Health, finance, gambling and similar fields are flagged as needing extra rules, not checked in depth.
- Rules change, and the records will age. Treat the check dates as part of every finding.
- Agents make mistakes, weaker models more than others. Re-read each cited line before you act on it.
Try it
In a terminal, with Claude Code installed:
claude plugin install last-look --marketplace KrunalSM/last-look
Or inside Claude Code:
/plugin marketplace add KrunalSM/last-look
/plugin install last-look@last-look
Then ask: "Use last-look on this project." Put your answers in the same message to skip the questions. Add "no search" to skip the web search, or "rules only" to skip the good-practice section.
For other agents, copy the repo without its tests/ folder, which holds planted mistakes and fake keys, to wherever the agent reads skills. If it doesn't pick the skill up by name, ask it to read last-look's SKILL.md and follow it.
It is MIT licensed. If it gets a rule wrong or flags something that isn't there, open an issue.