How we test
Skills run code next to client numbers. We act like it.
Every release passes the same gates before it reaches a library — and this page shows what those gates are, what each skill can touch, and what we will never see.
The release pipeline
- 01 · AUTHORSkill, scripts, examples and tests written together
- 02 · REVIEWAn independent reviewer reads every instruction and script
- 03 · SANDBOXFull run on sample agency data, isolated from real accounts
- 04 · DISCLOSEPERMISSIONS.md checked against what the run actually did
- 05 · ROLLBACKDowngrade to the previous version tested and documented
- 06 · RELEASEVersioned, logged, and added to your library
A release that fails any gate goes back to step 01. Nothing skips a gate for speed.
Release checklist
What every release must pass.
The results for each version are published on its skill page under “Test record”.
- Instructions in SKILL.md reviewed for scope and guardrails
- Every bundled script read in full; no hidden downloads or calls
- Outputs match tests/expected on the tested configuration
- Missing data is flagged, never filled with a guess
- Nothing is sent, published or shared on the user’s behalf
- PERMISSIONS.md matches the observed run
- Rollback to the previous version tested end to end
Your data
Two separate places. Client data only ever lives in one.
Your AI account
you control itexports.csvskills run heredraft for sign-off
- Client exports, notes and drafts
- Your AI usage and its billing
- The skills you install, once downloaded
This store
we hold- Your email, agency name and billing country
- Licenses, seats, invoices and download history
NEVER RECEIVED
- Client files, notes, drafts or reports
- Your AI account credentials
- Card details — payment is handled by the store’s payment service
Permissions
Matches PERMISSIONS.md in each packageWhat each skill can touch.
| Skill | Reads | Writes (in your session) | Runs | Network |
|---|---|---|---|---|
| ASK-00Reporting Diagnostic | Your answers only | reporting_scorecard.md | scripts/scorecard.py | None |
| ASK-01KPI Normalizer | CSV exports you upload | kpi_normalized.csv, qa_log.json | scripts/normalize.py | None |
| ASK-02Evidence-Linked Drafter | kpi_normalized.csv, approved client notes | draft_report.md, evidence_index.csv | scripts/link_evidence.py | None |
| ASK-03Report QA & Brand Check | Draft, evidence index, your brand sheet | qa_checklist.json, annotated draft | scripts/check_report.py | None |
Known limits, stated plainly
- Provider and model compatibility review is pending. Local Python script tests are included in each package.
- Drafts need a person to sign off. The skills are built for that, not around it.
- Time-saving results are still being measured. We publish them once independently tested.
- AI models change. We re-test and post status on the releases page.
FOUND A PROBLEM?
Report unexpected behaviour or a security concern — we answer every one.
Security reports: the support form below. Please describe what happened, never attach client data.