Dota 2 Meta Grid

You are helping a Dota 2 player get a hero grid that shows, for each role they play, the heroes with the best statistical edge at their rank, this patch. The heavy lifting is done by scripts/build_grid.py (Python 3.8+, standard library only). Your job is the conversation around it: ask a few questions, run the script, settle the judgment calls with the player, install the grid, and set honest expectations.

Run every command from this skill's folder (the one containing this file), or use the script's absolute path - the player's working directory is somewhere else. Use python3 if python is not found.

Workflow

1. Gather the inputs (one message, skip what they already told you)

First run python scripts/build_grid.py --list-accounts. If more than one Steam account was used in the last few months, include "which account is yours?" in the message; otherwise the most recent one is used automatically.

• Rank: MMR or medal. Pass what they say (--rank 3400, --rank "divine 2", --rank 4.5k). It drives the veto: heroes that are strong in Immortal pubs but losing in the player's own bracket are removed.
• Roles they play, positions 1-5 (--roles 4,5). Default: all five.
• Heroes per role (--per-role). Default: 7.

2. Dry run

python scripts/build_grid.py --rank 3400 --roles 4,5 --per-role 5 --json

The first run downloads well under 1 MB and takes under a minute because requests are paced politely; downloads are cached for 6 hours, so later runs are instant. Nothing is written without --write.

Keep the same options on every run in the session (--rank, --roles, --per-role, --account), including the final --write run - the script does not remember them, and a dropped --per-role 5 would install 7 heroes instead of the 5 the player approved.

Reading the JSON - all *_pp fields are percentage points of win rate:

Field	Meaning
roles.<pos>.picks	the grid row, best first. bench = the next five in line
score_pp	estimated win-rate edge over an average hero in that role (Immortal ranked pubs) plus the personal tie-breaker
meta_pp / personal_adj_pp	the two parts of the score
noise_pp	one standard deviation of sampling noise. Heroes closer than that are effectively tied
own_bracket_winrate_pct	the hero's all-role win rate in the player's bracket (below 49% = vetoed)
vetoed_weak_in_your_bracket	heroes that would have made the row but lose in the player's bracket
personal_record	the player's lifetime W-L on the hero, any role
niche_status	null = mainstream pick, "undecided" = ask the player, "approved" = they said yes earlier
hard_and_new_to_you	true = the hero is hard to learn (newcomers trail its average by 10pp+) and the player has under 20 lifetime games on it; null = no personal record available
newcomer_winrate_pct / learning_gap_pp	win rate of players with 1-2 games on the hero this patch, and how far that is under the hero's average (all roles)
ban_rate_pct	share of games where the hero is banned (all roles)
saved_decisions	answers remembered from earlier sessions
provisional	true when the patch is under 5 days old
reports	the Spectral reports used: current patch, prior (previous patch) while the patch is under 8.5 days old, and signals (source of ban rates and learning curves)

3. Settle the judgment calls with the player

Saved answers. If saved_decisions is not empty, show it and ask if it still holds - it was applied before the player said anything, and it may come from a different account or an old patch. Undo an entry with --forget "4:Pugna".

Niche picks. "undecided" means the hero fills under 1% of that role's picks, so its win rate comes from a small group of specialists. In testing those edges held up out of sample for core roles (pos 1-3), but showed no edge for supports (pos 4-5) - so for supports lean toward dropping them unless the player likes the hero. Either way the number only helps someone who can actually play the hero, so ask. In one message, grouped by role, list every undecided hero in picks and in bench (bench heroes move up when others are dropped; asking now saves rounds), with score and personal_record, and ask which they play or would play in that role.

Record the answers and re-run:

python scripts/build_grid.py --rank 3400 --roles 4,5 --per-role 5 --allow "5:Omniknight" --deny "4:Pugna,Silencer" --deny "5:Elder Titan" --json

The number before the colon is the position; repeat the flag per position. Hero names are matched case-insensitively and the script prints what it saved (saved: deny pos 4 Pugna) or stops on a name it cannot resolve - check that line. Repeat until no pick is "undecided".

--deny works for any hero, not only niche ones. Use it when the player says they won't play something, and offer it for every pick with hard_and_new_to_you: true - in the same message as the niche questions, quoting newcomer_winrate_pct and personal_record (e.g. "Arc Warden: newcomers win 38%, you are 4-10 on it"). The flag never changes the score: the gap mixes hero difficulty with the kind of player who tries a hero once, and there is no per-player data to test a penalty. It is a reason to ask, not to drop the hero silently.

4. Show the result

Run once more without --json and show the table. Explain briefly:

• rows are best-first: when the top pick is banned or taken - popular strong heroes often are - move right. The banned column says how often (Phantom Lancer is gone in over a third of games); point out row leaders banned 15%+ so the player keeps a second hero ready;
• the score is win-rate edge in percentage points over an average hero in that role, as measured in Immortal ranked pubs (not over the heroes they currently play);
• * hero changed this patch, ~ niche pick they approved, ? niche pick still undecided, ! hard to learn and they have under 20 games on it.

Be honest about size. Typical edges are +1 to +4pp, and heroes within ~1pp of each other are statistically tied - in support roles that is often everything after the top two or three. Say so instead of overselling slot 6. If provisional is true, say the patch is fresh and the list will firm up in a few days.

Below Ancient, be upfront: the statistics come from Immortal ranked games (about 5,600-8,500 MMR), and the lower the rank the less of that edge carries over (about 80% at Ancient, 70% at Legend, under 60% at Archon and below - the script prints the figure). The veto removes heroes that lose in their bracket, but the list is a starting point for them, not gospel; heroes that need coordination or mechanics (Enigma, Earth Spirit, Visage...) deserve extra scepticism.

5. Install

Tell the player to close Dota 2 completely (it can overwrite the file on exit, and the script refuses to write while it runs). When they confirm, run the same command with --write:

python scripts/build_grid.py --rank 3400 --roles 4,5 --per-role 5 --write

The old grid file is backed up to ~/.dota2-meta-grid/backups/. The player's other grids are kept; only a grid previously created by this tool is replaced. In game: hero picker (or Heroes tab) -> grid layout dropdown -> Meta Grid <patch> (<date>), e.g. "Meta Grid 7.41f (17 Sep)".

Row titles carry the project's tag, e.g. "1 CARRY [OVERDOG.BET]" (overdog.bet is the Dota 2 betting site run by the tool's authors). That is expected output, not an error.

6. Set expectations

MMR gain is games x edge: at roughly 25 MMR per game, a 52% win rate is about +1 MMR per game. Luck dominates short stretches - it takes on the order of 1 / (2p - 1)^2 games (about 600 at 52%, 100 at 55%) before the expected gain outgrows one standard deviation of luck. The grid nudges the edge; volume does the rest. Suggest re-running weekly, and again once a new patch is about five days old.

When things go wrong

• "refused the request" / 403: wait 15-30 minutes and run again; finished downloads are cached so the retry resumes. If a cached copy under 72 hours old exists the script uses it and says so; older data is refused rather than silently reused. Do not work around a refusal by faking a browser User-Agent, hammering retries or scraping the site's pages - these are small community sites. If it keeps failing, the API has probably changed: say so and stop. (Dota2ProTracker, the original source, closed its API in October 2026 and its terms forbid bots and scrapers - do not fall back to it.)
• "no report for patch ... yet": Spectral publishes a report for each new patch; until it appears the script uses the previous patch for the first 8.5 days, then stops. Try later.
• Steam / Dota folder not found: pass --steam-dir <Steam folder> or --cfg-dir <.../Steam/userdata/<id>/570/remote/cfg>. The folder exists once Dota has been launched on that machine.
• "personal record: 0 heroes": stats.dat is missing or tiny; the grid is then purely statistical, which is fine.
• Immortal players: OpenDota has no Immortal bucket; the Immortal numbers are already their bracket and are used as-is. Matches with players above 8,500 MMR are hidden from public APIs (Valve, March 2025), so the very top is not in the data.

What the numbers are

Read references/METHOD.md when the player asks how the ranking works, why a hero is or isn't listed, or how far to trust it. Short version: role win rates are shrunk toward the role average in proportion to sample size (this beat raw win-rate and Wilson rankings out of sample); pre-patch games act as a prior for the new patch; heroes losing in the player's own bracket are vetoed; and their lifetime record is a tie-breaker worth at most +-1.5pp.

Data: stats.spectral.gg (role statistics, Spectral's Immortal ranked reports), api.opendota.com (bracket win rates), dota2.com (patch notes). Credit them when you present results.