Ability Draft Plus · under the hood
Turn on role suggestions and the overlay stops handing everyone the same ten abilities. This page walks through what actually happens: what comes from data, what the community curates by hand, and how the two get combined. The plain-language version reads straight through; the math hides in the fold-outs.
Before roles come into it, every ability gets a score from thousands of real Ability Draft games. Two inputs: how often it wins, and how early people grab it. Wins a lot and disappears in the first picks — high score. Sits around until pick 40 — low score.
The catch is that pick order measures popularity, and drafters are only human. Some abilities get insta-grabbed and then lose the game anyway. Rearm goes ninth on average and wins 45.7% of its games. When something is picked way earlier than its winrate can justify, the app knocks its score down and says so:
Base score = 0.4 × winrate + 0.6 × inverted pick order (pick
order normalized over positions 1–50). The overrated damp fires when winrate
< 48% and average pick ≤ 15 — calibrated on live data, it catches
21 of 513 abilities, essentially a who's-who of famous trap picks. It subtracts
a flat 0.12 and applies whether role mode is on or not. If you link your Windrun
profile, your personal winrates are blended into the inputs first (small samples
are shrunk toward the global numbers so five lucky games can't hijack a score).
The core question behind role awareness: is this a farming ability or a supporting ability? Nobody at a keyboard decides that. For every ability, Windrun tracks how its drafters' numbers shift — do the people who pick it end up richer than average, or poorer? That measurement is the greed axis.
Each position has a home on the axis, learned from expert drafts. Suggestions near your position's home get a bonus; suggestions far from it lose points:
The raw shift columns have undocumented units, so the app only ever uses their ordering — every axis is a percentile rank within the current pool, rescaled to [−1, +1]. Gold and XP shifts build the greed axis. Position anchors: pos 1 = +0.7, pos 2 = +0.4, pos 3 = +0.1, pos 4 = −0.35, pos 5 = −0.6. The greed bonus fades as your draft progresses (weights 1.0 → 0.85 → 0.6 → 0.4 per pick) — experts grab their greed engine first and shop for utility later, so late picks are judged more by what your build still needs. Secondary axes (kill-taking for pos 2, playmaking for pos 4, healing for pos 5) add small accents on top.
Every ability carries a few function tags from a fixed vocabulary — hard_cc waveclear steroid save_ally and so on. A tag is a mechanical fact: this is a stun, this clears waves. The community curates them in the Tag Lab. They exist so the app can reason about your build as a whole instead of scoring picks one at a time.
Each position has a short checklist of things a real draft at that spot wants. As you pick, the app tracks what your build already covers. A candidate that plugs a hole gets a boost, and the tooltip names the hole:
| Position | First priority | Second | Third |
|---|---|---|---|
| 1 · Carry | attack steroid | farm tool or waveclear | mobility or self-sustain |
| 2 · Mid | nuke | teamfight ult or hard disable | mobility |
| 3 · Offlane | AoE disable | initiation | durability |
| 4 · Soft support | hard disable | waveclear or nuke | mobility |
| 5 · Hard support | hard disable | waveclear or nuke | ally save |
Priorities follow one rule: what defines the role and can't be bought from the shop ranks highest; what items can patch ranks lowest.
Need boosts scale with pool scarcity — two stuns left in the pool nearly doubles the disable boost; twelve steroids dilute the farm boost. Coverage is fractional: a slow or silence counts as half a hard disable for your own build, but a slow is only suggested as your disable once no real stun remains in the pool. Overstacking is damped too — the third single-target stun, or a third mana-hungry active on one mana pool, loses points. Chips always tell the truth about the match: a nuke satisfying the wave-control need reads "covers your missing nuke", and half-credit reads "partially covers".
Stats miss things. Some abilities are quietly great for one role, some are famous traps, and a few literally don't work without a partner. The dataset carries hand-curated verdicts and gates for these — proposed and voted on in the Tag Lab, and always explained in the tooltip:
Always gets a suggestion slot on its positions, even when the raw numbers disagree. Glimpse for supports is the founding member — its value shows up in coordinated kills, not in aggregate stats.
must: pos 4·5
Dropped from suggestions on its positions — or everywhere. Ransack is never suggested, full stop. The stats stay visible; only the recommendation is withheld.
never suggest
Some spells are dead without a partner. Eclipse does nothing without Lucent Beam. Requiem needs the Shadow Fiend body, because the souls it feeds on are his innate. Rearm only shows up while something Rearm-worthy is drafted or still in the pool.
A rare few abilities are wasted on the wrong attack type — Marksmanship does nothing for a melee body. Those get filtered against your chosen model, with sensible exceptions: an ability always works on its own hero, and drafting a range-granting spell like Psi Blades or Take Aim unlocks them again.
Verdicts are deliberately not tags — tags are facts about what a spell does; a verdict is a judgment about who needs it. Both flow through the same community loop: propose in the Tag Lab (must-pick and never-suggest need a written rationale — they override the stats, so they must argue their case), vote, and accepted changes ship with the next app update. All-position verdicts apply even with role suggestions switched off, because "never take Ransack" isn't a role opinion.
In Ability Draft your hero model is just a body: base stats, attack speed, talents, and one innate. The app rates bodies for your role on what they actually bring:
Premier carry bodies are marked avoid: pos 4·5 — if you're the support and you take the Drow model, your carry doesn't get it. Once every teammate has locked a model the restriction lifts: at that point taking Drow just denies the enemy team, which is a perfectly good crime. The same logic runs automatically for any body that fits cores much better than supports.
An honest note for the data-minded: the app used to blend a greed axis for models like it does for abilities, and we killed it. Its signal was poisoned by a selection effect — supports pick Drow and Luna for the aura, which drags those models' measured "greed" to the bottom percentile and then wrongly tanked their carry rating. Model role fit now runs entirely on attributes, chassis numbers (attack rate, effective HP, speed), talent and innate tags, and curated verdicts.
Don't want to declare a role? Set it to Auto and the app reads your team instead. Each teammate's picks add up to a farm trajectory on the same greed axis — someone stacking cleaves and steroids is building a carry game whether they announced it or not. The app matches trajectories to positions and suggests for whatever's left:
Matching is an order-preserving assignment of builds (sorted by mean drafted greed) onto position anchors — provably optimal on a one-dimensional axis, no heavy machinery needed. An evidence gate keeps Auto silent until at least two teammates have two named picks each; before that, trajectories are noise and scoring stays neutral. A separate team-balance term also runs in both modes: on a farm-stacked team, greedy candidates lose a little and supportive ones gain.
The tag vocabulary, the must-picks and never-suggests, the model tags — all of it is community-curated. Browse every ability and hero body in the Tag Lab, propose changes, vote on other people's. Accepted changes ship with the next app update.
Draft statistics and ability shift data by Windrun.io. Position anchors, need priorities and layer weights are tuned against a corpus of ~3,000 expert Ability Draft games plus replay simulation. CC classifications sourced in part from Liquipedia (CC-BY-SA).
