| name | deputy-briefing-format |
| description | How a deputy should structure its briefing XML — what to include, what to leave out, how to weight relevance. |
| agents | deputy |
The briefing format
Your output is a <deputy_briefing> XML block. Its structure:
<deputy_briefing>
<signal>high_novelty</signal>
<relevant_cases>
<case>
<case_id>CBR-042</case_id>
<similarity>0.87</similarity>
<summary>Short description of how the prior case applied and what the lesson was.</summary>
</case>
</relevant_cases>
<relevant_facts>
<fact>
<key>cycle_log_test_pattern</key>
<value>One-line answer to what this fact says.</value>
</fact>
</relevant_facts>
<known_pitfalls>Plain-prose note about a recent pattern of failure that the agent should be aware of.</known_pitfalls>
</deputy_briefing>
Rules
Quality over quantity
- 1-5 relevant cases. More than that is noise. Zero is often correct.
- 0-3 relevant facts. Only include facts the specialist might not otherwise
see (things not already in its skill docs or system prompt).
known_pitfalls is optional. Only populate it if recent narrative
entries show a clear pattern worth warning about.
An empty briefing (signal=silent, no cases, no facts) is a correct and
valuable output when nothing in memory applies. Don't pad.
Similarity scoring
Your similarity scores are your honest estimate of how applicable a prior case
is to the current delegation:
- 0.8+ — clearly applicable. The specialist should act on it.
- 0.5–0.8 — related. Worth a look but may not apply directly.
- Below 0.5 — generally not worth citing at all. Drop it.
If you cite a case below 0.5, explain in the summary why it's still useful.
Signal tags
routine — known pattern, similar delegations have succeeded; the agent
doesn't need special guidance
high_novelty — this looks unfamiliar; cog should be aware
anomaly — there's something off (repeated failures in this domain,
conflicting evidence)
silent — nothing in memory applies; use this freely, it's not a failure
Summary writing
Each case summary should be one sentence, specific to the lesson:
Good:
- "Prior fix to cycle_log was blocked by missing FFI date override — agent had
to rewrite the test, not patch."
- "Similar research: the DuckDuckGo path returned empty for API-adjacent
queries; switch to Brave."
Bad:
- "Similar coding task." (generic — no lesson)
- "CBR-042 was about tests." (duplicates what the id already tells you)
XML hygiene
- Escape entities:
& → &, < → <, etc.
- No markdown formatting inside the XML — just plain text.
- No additional sections. The schema is strict.
When nothing applies
Return:
<deputy_briefing>
<signal>silent</signal>
</deputy_briefing>
The specialist will proceed without briefing context. This is not a failure;
it is the honest answer when memory doesn't cover the situation.