diff --git a/.agents/skills/clawhub-security-dataset/SKILL.md b/.agents/skills/clawhub-security-dataset/SKILL.md
new file mode 100644
index 00000000..02b1016f
--- /dev/null
+++ b/.agents/skills/clawhub-security-dataset/SKILL.md
@@ -0,0 +1,127 @@
+---
+name: clawhub-security-dataset
+description: Use when generating, publishing, validating, or changing the ClawHub security/eval dataset export, nightly snapshot workflow, Hugging Face dataset upload, Convex export ingestion, or sanitizer/redaction logic.
+---
+
+# ClawHub Security Dataset
+
+This skill covers the maintainer-facing ClawHub security/eval dataset export.
+The dataset is public; raw Convex exports, broad production credentials, storage
+ids, internal document ids, and obvious secrets are not.
+
+## First Pass
+
+- Read `AGENTS.md`, `convex/_generated/ai/guidelines.md`, and the current
+  Linear coordination issue when this is issue-backed work.
+- Check recent commits and security branches before editing:
+
+```sh
+git fetch --all --prune
+git log --oneline --decorate --date-order -18 origin/main
+```
+
+- Inspect the current command and data path:
+
+```sh
+node -e "const p=require('./package.json'); for (const [k,v] of Object.entries(p.scripts)) if(k.includes('dataset')) console.log(k+'='+v)"
+rg -n "securityDataset|dataset:snapshot|hf-dataset|convex-export|storageId|redaction" convex scripts .github/workflows package.json
+```
+
+## Command Map
+
+Dry-run a tiny live export against the configured Convex deployment:
+
+```sh
+bun run dataset:snapshot -- --dry-run --limit 10
+```
+
+Dry-run production only when the target deployment and auth boundary are clear:
+
+```sh
+bun run dataset:snapshot:prod:dry-run
+```
+
+Build sanitized files from a local Convex export ZIP:
+
+```sh
+bun run dataset:snapshot -- \
+  --convex-export-zip /path/to/export.zip \
+  --source-snapshot-id <snapshot-id> \
+  --out-dir .data/security-dataset/snapshots \
+  --hf-dataset
+```
+
+Useful flags:
+
+- `--prod` or `--deployment <name>` selects the Convex runtime. Do not use both.
+- `--push` pushes current Convex functions before `convex run`; use only when
+  that is intentional.
+- `--limit <n>` caps source artifacts for proof runs.
+- `--source-kind skill|package|all` narrows source rows.
+- `--created-after <date>` and `--created-before <date>` constrain by created
+  time.
+- `--hf-dataset`, `--hf-repo`, `--hf-revision`, and `--hf-commit` control the
+  Hugging Face-shaped split files and manifest metadata.
+
+## Security Rules
+
+- Prefer the live sanitized Convex export path for nightly automation unless a
+  genuinely narrow Convex backup-reader credential exists.
+- Do not automate backup downloads with broad deploy/admin credentials.
+- Do not upload raw Convex backups, raw `_storage` blobs, raw `storageId`, or
+  internal Convex document ids to Hugging Face.
+- The backup ZIP importer must resolve `_storage/<storageId>` only so sanitizer
+  logic can redact and cap public output content.
+- Reuse the shared sanitizer/normalizer for live exports and backup ZIP imports.
+  Do not add a second redaction policy without a manifest version bump.
+- Hugging Face writes should use the dataset repo `OpenClaw/clawhub-security-signals`
+  and OIDC/trusted publishing when available.
+
+## Output Shape
+
+The maintained output is:
+
+- normalized JSONL: `artifacts.jsonl`, `scan_results.jsonl`,
+  `static_findings.jsonl`, `clawscan_findings.jsonl`, `labels.jsonl`,
+  `splits.jsonl`
+- optional HF rows under `hf-dataset/data/{train,validation,test,eval_holdout}.jsonl`
+- `manifest.json` with source id, exporter git SHA, redaction policy version,
+  row counts, output sizes, selected splits/configs, and HF repo/revision/commit
+  metadata when applicable
+
+## Validation
+
+Run focused tests after changing this subsystem:
+
+```sh
+bun test scripts/security-dataset
+```
+
+For workflow or publish-path changes, also produce a dry-run proof with a small
+limit and inspect the generated manifest/output:
+
+```sh
+bun run dataset:snapshot -- --dry-run --limit 10 --hf-dataset
+```
+
+Before claiming output is safe, search generated sanitized files for forbidden
+markers:
+
+```sh
+grep -R -E 'storageId|skillVersions:|packageReleases:|_storage/' <snapshot-dir>
+grep -R -E 'gh[pousr]_[A-Za-z0-9_]{20,}|sk-[A-Za-z0-9_-]{20,}|AKIA[0-9A-Z]{16}|-----BEGIN [A-Z0-9 ]*(PRIVATE KEY|CERTIFICATE)-----' <snapshot-dir>
+```
+
+If credentials are available, validate a manual workflow run against a test HF
+revision before enabling or trusting the scheduled production upload.
+
+## Common Mistakes
+
+- Treating Convex Team Access Tokens as read-only backup credentials without
+  verifying Custom Roles or an equivalent permission boundary.
+- Blessing a backup-based workflow after the team decided a live sanitized
+  export is safer without narrow backup-read auth.
+- Documenting a successful dry run as a successful HF upload.
+- Forgetting that `--dry-run` does not write snapshot files.
+- Removing manual/admin export commands before the nightly path has equivalent
+  proof and a maintained entrypoint.
diff --git a/README.md b/README.md
index 8ac45802..5a1735a7 100644
--- a/README.md
+++ b/README.md
@@ -81,6 +81,21 @@ export CLAWHUB_DISABLE_TELEMETRY=1
 
 Details: [`docs/telemetry.md`](docs/telemetry.md).
 
+## Security dataset export
+
+Maintainers can generate a sanitized ClawHub security/eval dataset snapshot with:
+
+```bash
+bun run dataset:snapshot -- --dry-run --limit 10
+bun run dataset:snapshot -- --convex-export-zip /path/to/convex-export.zip --hf-dataset
+```
+
+The exporter writes normalized JSONL files, optional Hugging Face split files, and a manifest with
+row counts, source metadata, output sizes, and the redaction policy version. Do not use broad Convex
+admin credentials for automated backup reads; prefer a live sanitized export path or a genuinely
+narrow backup-reader credential. Future agents should read the repo-local `clawhub-security-dataset`
+skill before changing the exporter, workflow, or Hugging Face publishing path.
+
 ## Repo layout
 
 - `src/` — TanStack Start app (routes, components, styles).