recommendation-drafter·reporting·active
Recommendation Drafter
Produces specific, owner-implied recommendations with effort estimates and projected impact. Ties each one to its supporting evidence.
ARIAtrigger · agent
Recommendation Drafter
Produces specific, owner-implied recommendations with effort estimates and projected impact. Ties each one to its supporting evidence.
When this skill fires
·ARIA finishes synthesis + trend detection and needs to produce actions
·An engineer asks:"what should we do about this?"
·A leadership review asks for prioritized work for the next sprint
What "specific" means
→A bad recommendation:"Improve test coverage."
A good recommendation:"Add 4 scenarios to the checkout flow's test suite to cover the rate-limit edge cases that LEO's threshold-detector flagged at 240 concurrent users. EVAN can author them in ~3 hours; expected to recover the 5pp pass-rate decline observed over sprints 23–25."
The good version has:
·A specific code area (checkout flow's test suite)
·A specific gap (rate-limit edge cases)
·A specific source (LEO's threshold-detector)
·A specific owner (EVAN, implying the QA team)
·A specific effort (~3 hours)
·A specific impact (recover 5pp pass rate)
Owner implication
Recommendations don't name individuals — they name roles or agents. The mapping from agent → team is configurable per organization:
·EVAN-driven recommendations → QA / SDET team
·LEO-driven → SRE / performance team
·SARA-driven → security team
·TARA-driven → product / engineering managers (because they sign off on requirement quality)
→This protects the recommendation from going stale when individuals change roles. It also avoids the politics of naming someone in a leadership document.
Effort estimates
Effort is estimated in engineer-hours, not points or days. Hours are anchored to:
·Past similar work in the codebase (when the recommendation has analogues)
·A baseline from prior ARIA recommendations and their actual completion time
·A confidence interval (a 4-hour estimate with high confidence; a 1–3 day estimate with low confidence)
Impact projections
Each recommendation states what improvement is expected if executed:
·"Expected to recover the 5pp pass-rate decline"
·"Expected to push the bend point from 240 to 320 concurrent users"
·"Expected to close 4 of 7 open High security findings"
Projections are conservative. ARIA prefers under-promising and over-delivering — credibility compounds.
Tying to evidence
Every recommendation links back to:
·The specific finding(s) it addresses
·The agent run that identified them
·The historical trend that justifies prioritization
→A recommendation without evidence is opinion. ARIA refuses to draft those.
Output format
For each recommendation:
·Title (imperative, < 60 chars)
·Body (1–3 sentences explaining the action)
·Owner (team/agent)
·Effort (hours, with confidence)
·Impact (projected improvement, with confidence)
·Evidence links (to findings, trends, runs)
A typical sprint report has 3–7 recommendations. More than that is a sign of weak prioritization, and ARIA will surface that as a meta-recommendation:"prioritization is broken; reduce in-flight scope before adding more".