agentleFS
link7373link7373/agentic-bi-team/.claude/skills/privacy-review/SKILL.md

privacy-review

Skill4 starsChanged 7 days agoMITView on GitHub

Review a new data source, model, dashboard, or analysis that touches personal data before it is built — purpose, minimisation, classification, audience, retention, and regime-specific obligations — or handle a data-subject request (access, deletion, correction, opt-out). Use before connecting a source with personal data, before a model scores people, before widening an audience, and whenever someone asks "are we allowed to do this?".

4 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

1 rule. 0 are common, 0 are distinctive, and 1 is very specific to this project.

  • 0

    Common

    Most of the 3 edited versions we read keep these.

  • 0

    Distinctive

    This author's own choices. Worth a look.

  • 1

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read knowledge/governance-policy.md first.

    line 12

“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “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

Privacy Review

Owner: governance-officer, with the owning agent for the thing under review. Args name the subject, e.g. /privacy-review churn model using support ticket text, or /privacy-review request: delete customer 4417.

This is a short, structured assessment, not a legal opinion. It exists so the privacy question is asked while changing the design is still cheap — and so the answer is written down. Where a regime pack says a privacy professional must decide, this skill stops and says so.

Read knowledge/governance-policy.md first. If regimes: lists any, read standards/privacy/<regime>.md for each — they add questions to Mode A step 6.

Mode A — Review a new use (default)

  1. Describe the use in one paragraph. What is being built, from which tables, for whom, to drive what decision. If you cannot say what decision it drives, that is the first finding.
  2. Screen. Does it involve data about identifiable people — customers, users, employees, patients, leads? Check the tables against the Column Registry and sweep for unregistered personal columns (python scripts/check_governance.py). If no personal data is involved, record "screened out" in the decision log and stop. Most BI work ends here, and that is the right outcome.
  3. Purpose. Was this data collected for this use, or a compatible one? Billing data used for revenue reporting: yes. Support-ticket text used to score employees: a new purpose. A new purpose is the user's decision, not the team's — and under most regimes needs more than a decision.
  4. Minimise — redesign before you approve. For each personal column, ask what breaks without it:

    • Can the question be answered with a count, a rate, or a bucket?
    • Can the join use a surrogate key in place of an identifier?
    • Can dates become months, ages become bands, postcodes become regions?
    • Is free text needed at all? It contains everything, and nobody has classified it.

    The best outcome of a review is a design that no longer needs one.

  5. Audience and exposure. Who sees the output, at what grain? Apply min_group_size, then check the indirect routes: differencing between two views, filters that intersect down to one person, quasi-identifiers in combination. Does the output leave the approved export destinations?
  6. Regime questions. For each enabled regime, work through the "review questions" section of its pack. Anything it marks as needing a privacy professional — record it as open, owner: privacy contact and do not proceed past design on that item.
  7. If it is a model that scores or decides about people, add: what happens to someone the model is wrong about? Who can override it? Which groups could it treat differently, and how will that be measured (analytics.md Part 1, fairness impossibility)? Is the person told? Record the risk tier:
    • minimal — aggregate forecasting, no individual decisions
    • limited — ranks or segments people; a human decides
    • high — materially affects access to a service, price, credit, employment, or health. Needs explicit user sign-off and, under most regimes, a formal assessment by a privacy professional.
  8. Retention and deletion. Where do extracts, features, and scored outputs live, and when are they deleted? If a person is deleted from the source, does this artifact still hold them?
  9. Write it down. PRIVACY-REVIEW.md in the artifact's own folder (models/<name>/, dashboards/<name>/, pipelines/<name>/, or the analysis folder):
    • Verdict: proceed · proceed with conditions · do not proceed · needs a privacy professional
    • The use, the personal data involved, and its classification
    • What was removed or reduced in step 4
    • Conditions (suppression, audience limits, retention date, human override)
    • Open items and their owners
    • Review date — a review is of a design, and designs change
  10. Update the policy. New columns into the Column Registry; standing permissions into the Exceptions Log with an expiry; the ruling into knowledge/decision-log.md. Run python scripts/check_governance.py.

Mode B — Data-subject request

A person asks what is held about them, or asks for it to be deleted, corrected, or not sold or shared. These have legal deadlines under most regimes; say which applies from the regime pack, and treat the clock as started.

  1. Confirm the request came through the proper channel and that identity was verified by whoever owns that process. The team does not verify identity and does not act on a request relayed informally.
  2. Locate. Using the identifier supplied, list every table that holds the person: start from the Column Registry and the cross-source join keys in knowledge/data-sources.md, then check derived marts, model feature tables, scored outputs, and local extracts. Queries that read restricted columns need the override marker with the request reference as the reason (/* AGENTIC_BI_ALLOW_RESTRICTED: DSAR <ref> */); under strict enforcement a person runs them.
  3. Report, do not act. Produce the location report: system, table, what is held, whether it is a source or a derived copy, and what deletion or correction would cascade to. The team never deletes or alters personal data itself — the destructive-SQL hook blocks it, and a human with authority executes the change in the source system.
  4. Check what regenerates. A deletion that the next pipeline run restores from an upstream copy is not a deletion. Name every upstream source and backup the owner must also handle, and any aggregate small enough that the person remains identifiable in it.
  5. Keep the record minimal. Log in knowledge/decision-log.md: the request reference, date, type, and completion date. Never the person's identity or their data — the log must not become a new copy of what was deleted.

What this skill does not do

  • It does not give legal advice or decide a legal basis.
  • It does not replace warehouse access control. A review that concludes "the team should not read this column" is implemented with a grant, and backed up by the query guard.
  • It does not approve by default. No answer from the user on a "needs your decision" item means not approved.

Created by Colin Beck — https://www.linkedin.com/in/beckcolin/

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.