Explore: Add confirmation gates to state-modifying skills #69

New issue

Closed

opened 2026-07-08 11:54:35 -04:00 by jasper · 0 comments

jasper commented

2026-07-08 11:54:35 -04:00

Contributor

Pocock's skills use confirmation gates — explicit instructions like "do not enact the plan until I confirm we've reached shared understanding." This prevents the agent from racing ahead on destructive actions.

What to explore

Which skills would benefit — anything that changes state vs read-only queries
What the gate should look like — a standard phrase that works across skills
Whether the gate should be in the SKILL.md or in a shared reference file loaded by multiple skills
Whether some operations are low-risk enough to skip the gate (e.g. creating a resource vs deleting one)

Relationship to existing work

There's a parallel discussion in the SOUL.md about "reaching for irreversible system-level actions when blocked by access controls" — this would encode that principle in the skills themselves rather than relying on conversational memory.

Several skills in this repo describe state-modifying operations — deleting items via arr-cli, restarting services via docker-management, changing configs via crowdsec, etc. These actions are currently described as procedures to execute, without guardrails. Pocock's skills use confirmation gates — explicit instructions like "do not enact the plan until I confirm we've reached shared understanding." This prevents the agent from racing ahead on destructive actions. ## What to explore 1. Which skills would benefit — anything that changes state vs read-only queries 2. What the gate should look like — a standard phrase that works across skills 3. Whether the gate should be in the SKILL.md or in a shared reference file loaded by multiple skills 4. Whether some operations are low-risk enough to skip the gate (e.g. creating a resource vs deleting one) ## Relationship to existing work There's a parallel discussion in the SOUL.md about "reaching for irreversible system-level actions when blocked by access controls" — this would encode that principle in the skills themselves rather than relying on conversational memory.

jasper referenced this issue from a pull request that will close it,

2026-07-11 17:02:55 -04:00

docs(skills): standardize confirmation gates #105