Menu
How to answer an ask_deputy call concisely. Only meaningful in Phase 2+ (long-lived deputies).
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
How to orchestrate work across large documents and multi-agent delegations without burning token budgets. Reconnaissance-first, search-then-read, parallel-after-reconnaissance.
Decision criteria for delegating to the writer agent vs synthesising directly in your own response.
When and how to use the document library
How the Coder agent drives an OpenCode session via dispatch_coder — the engineering loop, when to dispatch, and how to verify the result is real (vs. claimed).
Maintain a personal index of interesting things, open questions, and follow-ups across sessions
When and how to render a promoted export to PDF via the export_pdf tool. Reads the writer's draft → export → PDF lifecycle and the operator-facing failure modes when the host is missing pandoc or tectonic.
| name | deputy-ask-response |
| description | How to answer an ask_deputy call concisely. Only meaningful in Phase 2+ (long-lived deputies). |
| agents | deputy |
This skill applies to Phase 2+ (ask-for-help mode). In Phase 1 (briefing-only),
deputies don't handle ask_deputy — they run once and die. Keep reading for
when Phase 2 lands.
In Phase 2, any agent in the delegation hierarchy can call ask_deputy(question, context?).
Your job is to answer concisely from memory — or admit you don't know.
Answers should be terse and cite their sources. Prefer:
From CBR-042: the last test-patching attempt failed because the date FFI
wasn't overridden. The fix was to rewrite the test rather than patch it.
Over:
Well, I found a case that might be relevant. It was CBR-042, about testing.
It seems like there was a fix involving FFI. You might want to look into that.
Cite the specific case_id or fact_key. Do not paraphrase past the point of recognizability.
Answer in 1–3 sentences. Cite your sources. Don't pad with caveats.
Structure:
Say so honestly. Emit a sensory event tagged unanswered with the question
and what you tried. Don't fabricate an answer.
I don't have a relevant case or fact for this. Escalating to cog as
`unanswered`.
If the question reveals that the agent is stuck in a way you can't help with,
emit a wtf sensory event alongside your unanswered answer. Cog may want to
investigate.
unanswered.