Give an assistant a reviewed route, not a bag of prompts.

Recommended flow

1. Start with the reviewed Practice Finder or practice.json for a concrete reflection situation.
2. Use practice-recommendations.json to preserve fit, exclusions, and no_match behavior.
3. Use reviewed learning pages for concepts and worksheets for structured blank forms.
4. For raw toolkit material, check toolkit-review.json before claiming CBT Cards publication.
5. Use toolkit-audit.json only for source-corpus triage. Candidate status is not publication approval.
6. When current review status matters, check content-review.json.
7. Check locale and translation status before presenting localized text as official CBT Cards content.

Trusted practice layer

• Situation/mechanism ontology
• Evidence and claim provenance
• Reviewed relations/progression
• Stable RAG chunks with safety in-chunk
• Semantic recommendation eval manifest
• Practice-system agent guide

The current layer has eleven CBT Cards-owned reviewed practices. It is intentionally smaller than the 115-record source corpus.

Publication, audit, and freshness are different

Publication overlay: raw source records not explicitly listed as reviewed and published remain unreviewed and source_only.

115-record audit: exposes duplicates, similarity groups, framework buckets, treatment-like wording, rework gates, and review priorities. It cannot publish a record.

Freshness registry: covers 26 trusted items and records their latest editorial/safety review date and next due date. A newer Git commit is not automatically a newer content review.

Agent Skill v1.8.0

The mutable SKILL.md currently points to v1.8.0. Reproducible clients should resolve manifest.json and pin the immutable version. The strict portable copy is available at v1.8.0/cbt-cards/SKILL.md.

Installation notes cover OpenClaw, Hermes Agent, generic Agent Skills clients, and the dependency-free HTTP/RAG reference path.

Runnable generic integration

Clients that do not implement Agent Skills can use the dependency-free cbt_cards_http_client.py reference consumer. It loads the canonical practice RAG manifest and NDJSON distribution, verifies the distribution and per-record hashes, and preserves canonical URLs, review status, locale, safety scope, and the in-chunk Avoid when context.

The reference client supports exact stable practice IDs and exact mechanism filters only. It deliberately does not perform free-text semantic routing, diagnosis, opaque ranking, or model-quality evaluation. Unknown exact IDs/mechanisms return no_match instead of producing invented CBT Cards guidance. The generic path is exercised offline in repository CI; OpenClaw and Hermes runtime paths remain documented compatibility paths rather than falsely claimed continuous executions. See interoperability-fixtures.json for the explicit status of each client class.

Reuse and licensing

Current terms: CBT Cards repository material and CBT Cards-owned machine-readable data remain under CC BY-NC-SA 4.0 unless a more specific source/license field says otherwise. That means commercial reuse is not currently granted by the public license without separate permission; appropriate attribution and ShareAlike requirements still apply.

Each record in knowledge.jsonl now exposes its current license_url and rights_basis. Six learning records are marked cbt_cards_original; six source-card-* records are marked adapted_from_toolkit and retain the source toolkit's CC BY-NC-SA 4.0 license reference. The pinned 115-record toolkit source is separately documented in toolkit-source.json.

When reusing a record, preserve useful attribution context: CBT Cards/MetalHatsCats as applicable, the canonical CBT Cards URL, the applicable license URL, stable IDs, source-toolkit attribution for adapted records, and an indication of changes. A canonical URL is strongly preferred because it preserves review and safety context; that preference is not an extra legal condition beyond the applicable license.

CBT Cards names, logos, and trademark rights are not granted by the Creative Commons license. Product artwork is not silently included in any future permissive data grant. The current Agent Skill also remains explicitly CC-BY-NC-SA-4.0.

A possible commercial-friendly data license is under publisher/legal review only. LICENSING_DECISION.md compares keeping NC-SA, scoped CC BY 4.0 for CBT Cards-owned machine data, CC0 for narrow factual metadata, and separate commercial permissions. Mentioning those options is not a new license grant.

Languages

English is the canonical source locale for the current curated knowledge set. Russian records are machine drafts unless an individual record is explicitly human-reviewed and published. German remains planned. A host assistant may translate canonical English for the user but must not label its own translation as an official CBT Cards localization.

Research and evaluation

The research layer separates deterministic routing baselines, held-out challenge cases, semantic recommendation cases, and provenance-clean real-model execution. Deterministic harness scores are not model-quality or clinical-quality claims.

Public interface boundary

These resources do not provide access to a user's journal, check-ins, backups, account, or other private app data. Website, data, evaluation, or skill releases are separate from mobile-app releases.

Safety

Do not infer diagnosis, remove genuine safety measures, turn source-only material into individualized treatment instructions, or route medical/legal/financial/safeguarding decisions to a reflection practice. CBT Cards is for general wellness and self-reflection, not diagnosis, treatment, or emergency support.