Live Draft Assistant

---
name: live-draft-assistant
description: This skill should be used during a live draft, when the user says "I'm on the clock", "who should I pick", "best available", "my draft is live", "we're drafting now", "should I reach for him", "what's my max bid", "who should I nominate", or reports draft picks as they happen. Tracks a live snake or auction draft in real time - best-available logic, reach/value alerts against ADP, auction budget and nomination strategy, and roster-construction tracking. Not for pre-draft board building (draft-strategy draft-prep), keeper cost decisions (draft-strategy keeper-evaluation), or in-season waiver bids (waiver-wire faab-bidding).
---

# Live Draft Assistant

Run the user's draft in real time: track every pick, keep a best-available view, and warn about reaches, tier cliffs, and roster imbalance. Speed matters — on the clock, lead with the answer, then the reasoning.

## Setup (first message of the draft)

1. Read `leagues.md` from the project root first — the fields that matter here are teams, scoring, and starting slots; snake math depends on them. If the file is missing or those fields are blank, ask for them directly and suggest running the `fantasy-league-setup:league-config` skill to persist the answers. If `leagues.md` defines more than one league, use the one marked `(default)` unless the user names another.
2. Get the board: use the tiered board from a prior draft-prep session if one exists in the conversation or a file the user points to; otherwise ask them to paste rankings/ADP. No board means the skill can only do roster-construction tracking — say so. If the board is rank-only or missing ADP, say up front which outputs are unavailable (numeric VOR, reach alerts, exact value-fall alerts) instead of pretending precision you do not have.
3. Get the user's draft slot and confirm format (snake vs. auction; linear third-round-reversal if applicable).

**Live platform source routing.** Honor a browser the user explicitly names. If `leagues.md` records a Preferred browser, use that when it has a signed-in session for the platform. Otherwise use any authenticated browser the current assistant already has. For Yahoo league data, prefer an authenticated browser over a connector. If a Yahoo connector returns `403`, `unauthorized`, or an equivalent authorization failure, do not retry it during the same task. For non-Yahoo platforms, use a purpose-built connector when it is available and returns complete current data; otherwise use the browser. Read league rosters, the free-agent pool, standings, transaction history, and any rankings site directly instead of making the user paste them. Timestamp live data and name the source. The session rules from `roster-ops` apply unchanged: the user's session is the auth; never ask for, read, store, or type credentials; use the visible UI rather than platform endpoints; and stop and hand back on any login, 2FA, captcha, consent, or unusual-activity screen. If no usable live source exists, state the access gap and do not fabricate league-specific analysis.

**This skill is read-only.** Reading a page needs no confirmation, but never click anything that changes a roster, files a claim, or sends an offer from here — that is what the `roster-ops` skills and their confirmation gates exist for. If a live draft board is open in the browser, read picks from it rather than making the user retype them.

## Draft state tracking

Maintain throughout, and restate briefly every few picks:

- **Picks made** — the user reports picks ("Vellum went 1.03", "I took Renner"). Cross players off the board; log which team took them when the user says.
- **User's roster so far** — filled slots vs. required starters.
- **Checkpointed draft state** — keep an append-only, read-only state summary after each confirmed pick: current round/pick, drafted players removed, user's roster, and any unknown picks still pending ingestion. If the draft feed is partial, mark missing selections as unknown and keep working from the last confirmed checkpoint rather than inventing backfill. The checkpoint should be durable enough that the user can paste it into a fresh thread and resume cleanly.
- **Snake math** — distinguish `next-pick gap` (the difference in overall pick numbers, used for reach math) from `intervening picks` (how many other selections happen before the user is up again).
  - In a standard N-team snake at slot `s`, after an **odd**-round pick (order runs 1→N), the next-pick gap is `2 × (N − s) + 1`.
  - In a standard N-team snake at slot `s`, after an **even**-round pick (order runs N→1), the next-pick gap is `2 × s − 1`.
  - Pick labels flip too: slot `s` picks `s` in odd rounds and `N − s + 1` in even rounds.
  - Third-round reversal, if the league uses it: round 3 runs reversed too, so the round-2-to-round-3 gap is exactly `N` selections for every slot (`N − 1` intervening picks), and round 3 uses the same pick label as round 2. From round 4 onward, resume the normal alternation with the parity swapped relative to a standard snake.
  - Always show both counts plainly, e.g. "you pick again in 12 picks; 11 selections happen before your turn."

## On-the-clock recommendation logic

When the user asks "who should I pick," answer in this order:

1. **Best available by value over replacement when the board supports it** — top 3 remaining, with tier labels. If the board is rank-only, use best remaining tier/rank instead and say the recommendation is structural rather than projection-priced.
2. **Tier urgency override** — a tier is dying if it will not survive to the user's next turn. Compare players left in the tier against **expected picks at that position among the other managers' selections**, not against total picks: `expected_takes = intervening_picks × positional share of picks` (roughly 0.30–0.40 for WR or RB in the middle rounds). Take from the tier now if `players_left ≤ expected_takes + 1`. The next-pick gap belongs in reach math; using it here counts the user's own upcoming selection as though another manager could take the player first. Comparing against all intervening picks without positional share is worse still — 4 players left before 10 other selections is a coin flip, not a comfortable wait.
3. **Roster construction override** — never leave a required *skill* slot unfillable. Counting only QB/RB/WR/TE/FLEX, by round ⌈skill starters × 0.75⌉ the user should have no more than one empty skill-position group. Don't take a third TE while 0 WRs are rostered, no matter the value.
   - In standard 1-K/1-DST formats, exclude K and DST from this rule and leave them for the final two rounds. They have essentially zero value over replacement and free replacements exist all season, so counting them as required groups will push a kicker into round 7 and waste a bench asset. If the league's actual scoring or lineup rules make K, DST, return specialists, or IDP slots scarce, treat them like any other position instead of forcing the standard prior. A QB2 in a typical 1-QB league is nearly as wasteful.
4. **Bye stacking check** — warn (don't block) when a pick creates 3+ shared byes among likely starters.

## Reach and value alerts

- **Reach alert**: fire when **`player's ADP − current pick > picks until your next turn`**. That is literally the test the alert claims to run — "he'll likely be there at your next pick" — and it uses the gap already computed above, so it self-adjusts for league size, draft slot, turn picks, and third-round reversal. A fixed spot-count threshold is wrong at both turns and in every non-12-team league, and it ignores that ADP dispersion widens from ~2–4 picks in round 1 to ~20–30 by round 10: a 12-spot deviation is meaningful early and pure noise late. If the tier-urgency rule justifies the reach, say so — reaching into a dying tier is correct.
- **Value alert**: a player has fallen a full round past ADP → flag proactively, even unasked: "Quincy Marsh (ADP 2.03) is still on the board at 4.01 — value pick."
- Reaches the user makes anyway are their call; update the board and move on without relitigating.

## Worked example (fictional)

12-team snake, user at slot 7, pick 3.07 ("Gridiron Gazette" league). Roster so far: RB Tobias Renner, WR Cassius Bell.

> **Pick now: Errol Fontaine (RB, Tier 3, value +54).** He's the last Tier 3 RB, so the tier dies here by definition. Your next-pick gap is 11 (odd round, `2 × (12 − 7) + 1`, so you're back at 4.06), while 10 selections occur before your turn. Keep the two counts separate: the 11-pick gap is for the reach test, but only those 10 intervening selections matter for tier urgency. The WR call is closer than it looks: Deion Marsette (Tier 2, ADP 3.10) has four tier-mates left, but at roughly a 35% WR pick rate across those 10 intervening selections, about 3.5 — roughly 4 — WRs come off. So `players_left (4) ≤ expected_takes (3.5) + 1`, and that tier is genuinely at risk too. Fontaine still wins because his tier has one player left against four; just don't count on Marsette returning. Reach check: Fontaine's ADP 3.09 is one pick away, far inside your 11-pick gap — no reach. Bye note: Fontaine shares week 9 with Renner — acceptable, but avoid a third week-9 bye.

## Auction variant

If the league drafts by auction:

- **Track live inflation** — `remaining league-wide money / remaining market value of the player pool`. This is the most valuable single number in an auction. If $1,200 chases $1,000 of value, everything left costs about 20% over par and hoarding is a losing strategy; if the ratio is below 1, bargains are coming and patience pays.
- **Hard ceiling** = `budget − open roster slots + 1`. Never exceed it; bidding to it leaves a roster of $1 players.
- **Max meaningful bid** = `budget − Σ(realistic price of remaining slots you still need)`, reserving the going rate for remaining *starter* slots and $1 for bench spots. Quote this, not the hard ceiling — the hard ceiling prices a starting WR2 and a last bench spot identically.
- **Budget shape**: roughly 75–80% of budget on starters, $1–3 total for K and DST.
- **Enforcement pricing**: bid up a player when a rival with that positional hole is the only other bidder — but only to a price you would happily pay yourself, because you will sometimes own him.
- **Nomination**: nominate players the user doesn't want while rivals still have money.