Custom skills and connected apps

Composable integrations the community describes as “it built a skill for Todoist” or “JIRA from Telegram”:

• Skills (or equivalent project conventions) that encode how to call an API, what to log, and how to recover errors.
• Thin wrappers around tools you already pay for: issue trackers, notes, CRMs, HR systems.
• Repeatable workflows with guardrails: allowed actions, required confirmations, and test prompts.

The point is not novelty—it is turning one-off hacks into maintained capabilities.

• Large skills ecosystem culture in the project: agents stay grounded when instructions live in-repo.
• Multi-channel access means the same skill can be invoked from the surface you prefer.
• Approvals prevent a skill from exfiltrating data or closing tickets you did not mean to close.
• Pair with features docs for what ships by default vs what you customize.

1. Pick one integration with a clear API or CLI.
2. Document inputs, outputs, and failure modes before coding.
3. Implement read paths first; add writes behind explicit approval.
4. Version skill text when behavior changes.
5. Add integration tests or manual test prompts you run after each change.

• API keys or OAuth as required by the vendor; rotate them on a schedule.
• Data classification: what must never leave your network.
• A single owner per skill so changes do not conflict.

1. Spike a smoke test: one read call, one logged response.
2. Add error taxonomy: rate limit vs auth vs bad payload.
3. Expand to multi-step flows with checkpoints.
4. Peer review before exposing to non-technical users.
5. Monitor usage and error rates in real traffic.

• “List minimal scopes needed for read-only vs write.”
• “What human confirmation is required before step 3?”
• “Draft rollback if this skill mis-files a ticket.”

• Dry-run succeeded against a non-production project or workspace.
• Audit trail shows who invoked what and when.
• Runbook exists for revoking credentials if leaked.

• Write-first skills that mutate data before you trust reads.
• Secrets in prompts—never paste production keys into chat logs.
• Unbounded polling that angers API vendors.
• Fork drift—three copies of the same skill with different behavior.