Platform skills: kb-learn and kb-recall¶
Every workspace comes with two platform skills that turn the workspace's default notebook — the Knowledge Base — into your team's long-term memory. They appear on the Skills page with a Platform badge the first time anyone opens Skills (or syncs), and sync to developers like any other skill.
Which knowledge base?
The Knowledge Base here is the workspace's default notebook — living pages agents read and write. It's separate from document ingestion, the read-only reference documents that ground AI generation.
/<prefix>:kb-learn — capture durable knowledge¶
Records one durable, reusable team decision in the Knowledge Base. The agent invokes it when someone states a lasting rule ("from now on…", "we always…", "don't do X again"), when PR-review feedback generalizes beyond the current task, or explicitly: /acme:kb-learn use REST, not gRPC, for new APIs.
The skill keeps the Knowledge Base trustworthy rather than noisy:
- Distills a timeless rule, not session narrative — "Use REST, not gRPC, for new service APIs", with the why and who decided.
- Dedupes before writing — if a relevant entry exists it's updated, never near-duplicated. Replaced rules stay visible as "Superseded" so agents meeting legacy code know the change was deliberate.
- Refuses junk — no secrets, no task-scoped context (that belongs in the feature's workbook), no speculation, nothing already enforced by linters or CI.
- Never blocks work — if the Knowledge Base is unavailable or full, the agent says so in one line and moves on.
/<prefix>:kb-recall — consult before building¶
Pulls relevant team decisions from the Knowledge Base before substantive work — starting a new endpoint, a refactor, a design decision — or on demand: /acme:kb-recall error response format.
The agent runs a couple of targeted searches, reads the few most relevant pages, and applies what it finds silently, citing an entry only when it actually shaped the work. When things conflict, it surfaces the conflict instead of guessing:
- Task vs. Knowledge Base — "KB says REST-only (May 2026); this task asks for a gRPC endpoint — proceed anyway?" If you overrule the entry, it offers once to update it.
- Entry vs. entry — contradictory entries are presented with their freshness signals; you pick which applies.
- Knowledge Base vs. code — an entry that no longer matches reality is flagged, and a confirmed-wrong entry becomes a
kb-learnmoment.
Together they close the loop: corrections you give an agent today become rules every agent follows tomorrow.
Platform skill lifecycle¶
Platform skills stay updatable by Dezycro until your workspace takes them over:
| You do | What happens |
|---|---|
| Nothing | Dezycro ships improvements automatically; developers get them on next sync |
| Edit the skill | Your version wins — the skill stops receiving platform updates. An "update available" hint appears when the bundled version moves ahead |
| Delete the skill | Opt-out for the workspace. The card stays on the Skills page so you can re-add it later |
| Reset to platform version | Discards workspace edits (or revives an opt-out) and restores the current bundled content |
Platform skills can't be renamed — the name is how Dezycro identifies them across updates.