agentleFS
mikezamayiasmikezamayias/agent-skills/skills/promise-audit/SKILL.md

promise-audit

Skill1 starsChanged 29 days agoBSD-3-ClauseView on GitHub

Traces every promise in a privacy policy, terms, store privacy label or marketing claim to the mechanism behind it. Use when writing, signing off or publishing such a text, or when a decision changes data handling or processors. Also use when asked to document the mechanism behind each promise, or to audit drift in a published policy.

The file

Show its rules and text

We split this skill into the rules it gives an agent

19 rules. About 0 are common, 14 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈14

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Test: list each label category, find the SDK or request that produces it, and cut the category or fix the sentence.

    line 47
  2. Distinctive:

    Test: disconnect the integration in a test account, ask the product for something only that data could answer, and check the store for the …

    line 49
  3. Distinctive:

    every processor (vendor) and the country its data sits in A crash or session-replay SDK is a processor that receives screens, taps, network …

    line 55
  4. Distinctive:

    If the tool offers an AI agent, ask which model provider receives the telemetry, because a vendor's policy may not name it.

    line 57
  5. Distinctive:

    every data category, including derived data such as scores, flags, and inferences

    line 59
  6. Distinctive:

    every automated decision that acts on someone before a human reviews it, for example account suspension, content filtering, or automated …

    line 60
  7. Distinctive:

    every retention period and every place data is copied to, including backups

    line 63
  8. Very specific:

    every entity property an app exposes to a system assistant, Spotlight or Shortcuts With App Intents, a property is readable by Siri unless …

    line 66
  9. Very specific:

    Check what a Transferable export puts in the text that Siri hands to another app.

    line 69
  10. Distinctive:

    Test: read each entity's properties in the source and list which ones a passerby could hear from a locked phone (Apple App Intents pages, …

    line 70
  11. Distinctive:

    add the mechanism (a spec change, a test, and a build task)

    line 78
  12. Very specific:

    Put the choice to the decision-maker on an answer sheet (see answer-sheet), with the exact before and after text.

    line 81
  13. Distinctive:

    Verify vendor facts such as retention, region, or data processing agreement coverage on the vendor's live pages.

    line 82
  14. Distinctive:

    Never ask the decision-maker to check them.

    line 83
  15. Distinctive:

    Keep the sign-off state inside the text file, as a comment block at the top or an HTML comment per section:

    line 87
  16. Distinctive:

    Never treat an approval as covering different text.

    line 96
  17. Very specific:

    Every ruling recorded afterwards runs through a check: does it make any signed sentence untrue, or add behaviour no sentence covers?

    line 101
  18. Very specific:

    If so, set the sentence to awaiting and add it to the next sign-off sheet, showing exactly what changed.

    line 102
  19. Distinctive:

    When a question needs a lawyer, for example whether a vendor without a data processing agreement is acceptable, say so in plain words and …

    line 116

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Promise audit

A policy is a list of promises. Under the GDPR (the EU's data protection law) and store review rules, a promise with nothing behind it is a liability, and so is a behaviour the policy does not mention. Both drift over time, because product decisions change behaviour after the text was signed.

This skill traces each sentence to its mechanism in both directions and keeps a sign-off state per sentence.

Inputs

  • The texts: privacy policy, terms, deletion page, store privacy labels (the App Store "nutrition label" and the Google Play Data safety form), consent screens, and marketing claims.
  • The spec: schema, jobs and retention rules, integrations and processors, screens and copy, analytics events, and moderation or automated-decision rules.
  • The decision log (see decision-sweep).

Step 1: Trace every sentence to a mechanism

Split each text into sentences, and quote each one verbatim. For each promise, fill one row:

SentenceMechanismWhere it is specifiedTest or check that proves itBuild taskStatus
  • Mechanism: what makes it true, for example a database cascade, a scheduled purge job, a setting, a vendor setting, or a screen.
  • Where it is specified: a link to the owner document.
  • Test or check: a test that fails if the promise breaks. Critical promises such as deletion, retention, and visibility need one.
  • Status: backed, gap, or contradicted.

Typical gaps:

  • deletion that removes the account but not analytics profiles, crash reports, backups, or files in storage
  • retention that is stated but not enforced by any job
  • vendor retention longer than the policy says, for example an analytics plan that keeps events for years
  • "we never share X" while an integration sends X
  • metadata the policy says is stripped but a code path keeps
  • a store privacy label that disagrees with the policy and the marketing line in either direction Example: a journaling app's listing says "Every word stays on your phone" and its policy says no typed text leaves the device, while its App Store label lists User Content under both "Data Used to Track You" and "Data Linked to You" (listing read 2026-10-30, network traffic not tested, so which side is wrong is unknown). The label is the developer's own answer and Apple does not verify it, so a mistaken label is a promise of its own. Test: list each label category, find the SDK or request that produces it, and cut the category or fix the sentence.
  • a disconnect or revoke that stops new access but leaves data already pulled from the integration stored and still used Test: disconnect the integration in a test account, ask the product for something only that data could answer, and check the store for the copied rows.

Step 2: Trace every behaviour back to a sentence

Walk the spec for anything that touches personal data or acts on a person, and check the texts cover it:

  • every processor (vendor) and the country its data sits in A crash or session-replay SDK is a processor that receives screens, taps, network calls and device data, so list it with the replay window and the masking level. If the tool offers an AI agent, ask which model provider receives the telemetry, because a vendor's policy may not name it. A keyword search of Measure's cloud policy found no model provider named, and its self-hosting guide routes the agent through OpenRouter with the operator's own key (read 2026-11-05, the policy only partly read).
  • every data category, including derived data such as scores, flags, and inferences
  • every automated decision that acts on someone before a human reviews it, for example account suspension, content filtering, or automated holds The GDPR requires these to be disclosed, with a way to ask for a human review. A section that claims to be complete must list them all.
  • every retention period and every place data is copied to, including backups
  • special-category data, such as health or biometric data, and the consent that covers each use of it A new use of it is a new purpose, so it needs new consent text.
  • every entity property an app exposes to a system assistant, Spotlight or Shortcuts With App Intents, a property is readable by Siri unless it declares visibility: .internal, and a property is read aloud without an unlock unless it declares an authenticationPolicy. Apple's default is that entities "don't require authentication before Shortcuts or Siri present entity information", and a property-level policy "only applies to responses Siri provides", so a tool's return value needs its own policy. Check what a Transferable export puts in the text that Siri hands to another app. Test: read each entity's properties in the source and list which ones a passerby could hear from a locked phone (Apple App Intents pages, read 2026-11-07, not tested on a device).

Anything missing is a gap in the other direction.

Step 3: Close the gaps

For each gap, choose one of two fixes:

  • add the mechanism (a spec change, a test, and a build task)
  • change the sentence to match reality, with a proposed rewording

Put the choice to the decision-maker on an answer sheet (see answer-sheet), with the exact before and after text. Verify vendor facts such as retention, region, or data processing agreement coverage on the vendor's live pages. Never ask the decision-maker to check them.

Step 4: Sign-off state per sentence

Keep the sign-off state inside the text file, as a comment block at the top or an HTML comment per section:

<!-- sign-off: section N, sentences X-Y: signed YYYY-MM-DD by the decision-maker
     section N, sentence X: awaiting (changed YYYY-MM-DD: <what changed>) -->
  • A sentence is signed only in the exact wording the decision-maker approved. A reworded sentence goes back to awaiting. Never treat an approval as covering different text.
  • Publication waits until every sentence is signed and every placeholder, such as the legal name, the contact address, or the domain, is filled.

Step 5: Re-open on change

Every ruling recorded afterwards runs through a check: does it make any signed sentence untrue, or add behaviour no sentence covers? If so, set the sentence to awaiting and add it to the next sign-off sheet, showing exactly what changed. The decision-sweep loop calls this step.

Output

  • the trace table, with counts of backed, gap, and contradicted sentences
  • the behaviours with no covering sentence
  • the proposed fixes, as an answer sheet
  • the sign-off state per sentence
  • what blocks publication

Legal limits

This is an engineering consistency check, not legal advice. When a question needs a lawyer, for example whether a vendor without a data processing agreement is acceptable, say so in plain words and name the exact question to ask.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.