dota-live-coach
Governs live coaching UX and recommendation presentation in Dota AI Coach — how much the engine's analysis actually reaches the player during active gameplay, message priority (P0/P1/P2), message brev
SKILL.md.claude/skills/dota-live-coach/SKILL.md
GitHubDota Live Coach: Presentation and Attention Budget
Core rule
The engine may analyze hundreds of variables. The player should normally see only 1-3 actions.
dota-domain, dota-item-recommender, and dota-draft all produce
richly reasoned, multi-factor analysis — that's correct and necessary
upstream. This skill governs the very different problem of what actually
reaches the player during a live match, when their attention is almost
entirely consumed by playing the game. A player who's shown everything
the engine knows learns to ignore all of it; a player shown 1-3 clear
actions at the right moment actually uses them. Rich analysis and
minimal live presentation are not in tension — they're sequential stages
(see "Where this fits" below).
This is a distinct concern from dota-item-recommender's alternatives
field or dota-draft's six-factor score — those are for auditability,
review, and (per those skills' explanation-stage guidance) a
richer post-hoc or on-demand view. Live presentation is a hard filter on
top of that full output, not a rendering of it.
Priority tiers: P0 / P1 / P2
Every live message has exactly one priority. Full detail and examples in
references/priority-tiers.md; summary:
- P0 (critical) — time-sensitive, high-stakes, act-now information.
Interrupts or takes visual/audio precedence. Reserve for things where
missing the window has an immediate, significant cost (e.g.
No buyback.before a risky fight,Three dead → objective.). - P1 (actionable) — a concrete recommended action, not urgent enough
to interrupt but worth surfacing promptly (e.g.
BKB in 400g.,Farm → BKB → fight.). - P2 (informational) — useful context that doesn't call for
immediate action (e.g.
Power spike ready.). Lowest visual priority; the first thing suppressed under attention-budget pressure (see below).
The 1-3-message budget applies across all tiers combined — P0 doesn't
get its own separate budget on top of P1/P2; a P0 firing should generally
displace, not add to, whatever P1/P2 messages are currently shown. See
references/priority-tiers.md for the suppression/displacement rules.
Message style
Live messages are extremely concise — short imperative fragments, not
sentences, and never an explanation during active gameplay. Full style
guidance and more examples in references/message-style.md; the
reference examples:
BKB in 400g.
Farm → BKB → fight.
No buyback.
Power spike ready.
Three dead → objective.
Notice the pattern: no subject pronoun, no justification clause, no
hedging — just the fact or the action. Do not give lengthy explanations
during active gameplay. The why behind a message (the reasonCodes,
the six-factor breakdown, the full analysis) belongs in a different
surface — a post-game review, a details panel the player can pull up
between fights on their own initiative, never pushed into the live
message stream. See references/message-style.md for how to compress a
rich recommendation into one of these fragments without losing the part
that actually matters for the next 10 seconds of play.
Preventing recommendation spam
A live coach that fires a new message every time an underlying value
changes is unusable — GSI/CV data updates frequently (per dota-gsi),
and most updates don't represent a meaningfully new situation. Three
mechanisms, used together, keep messages meaningful rather than noisy —
full detail in references/anti-spam-mechanisms.md:
- Hysteresis — a condition must clear a meaningfully different threshold to stop triggering than the one that made it start triggering, so the message state doesn't flicker on/off around a single boundary value.
- Cooldowns — a minimum time between messages (per priority tier, and often per message type), so even a genuinely fluctuating situation doesn't produce a rapid-fire sequence of restatements. P0 has the shortest allowed cooldown (it needs to be able to fire when it matters), P2 the longest.
- Change thresholds — a new message only fires if the underlying value changed by enough to matter for the player's next decision (e.g. "BKB in 400g" shouldn't re-fire for every 10 gold of farm — only when the number crosses a threshold worth re-announcing, or the recommended action itself changes).
None of these alone is sufficient — see references/anti-spam-mechanisms.md
for how they compose and for concrete parameter guidance (starting points
for cooldown durations per tier, how to pick change thresholds per
message type).
Where this fits in the pipeline
Per dota-architecture, this skill governs the Overlay/Voice layer's
consumption of the Recommendation Orchestrator's output — it does not
change what Recommendation Engines compute or how the Orchestrator
prioritizes between them; it governs the final filter/compression step
between "what the system knows" and "what the player sees right now."
Concretely:
Recommendation Orchestrator output (rich, multi-factor, per
dota-item-recommender / dota-draft)
↓
Live presentation filter (THIS SKILL: priority assignment,
hysteresis/cooldown/threshold gating, message compression to
1-3 concise fragments)
↓
Overlay / Voice (rendering)
A recommendation engine bug (wrong item, wrong pick) is a
dota-item-recommender/dota-draft problem. A player feeling spammed,
overwhelmed, or confused by how a correct recommendation was presented
is a bug in this layer, and should be fixed here — don't fix presentation
problems by dumbing down the upstream analysis, and don't fix analysis
problems by adding more live message volume.
Relationship to other skills
dota-item-recommender/dota-draft— produce the rich, multi-factor recommendations this skill compresses for live display; theirreasonCodes/explanation output is the raw material for P1/P2 message content, and their full detail belongs in a non-live surface.dota-architecture— this skill governs the Overlay/Voice layer specifically; it doesn't change the Recommendation Orchestrator's own responsibilities.dota-gsi— the update frequency and partial-payload nature of live state (perdota-gsi's state-diffing guidance) is exactly why anti-spam mechanisms are necessary here, not optional polish.