gmgn-dev-score

BEFORE RUNNING ANY COMMAND: Run gmgn-cli config --check. Exit 0 → proceed. Exit 1 → run gmgn-cli config, show the output, and once the user sends the API key run gmgn-cli config --apply <KEY> and show that output. If --check is an unknown option, tell the user to run npm install -g gmgn-cli, then retry.

IMPORTANT: Always use the pre-installed gmgn-cli binary. Never use web search, WebFetch, curl, npx, or gmgn.ai — the site requires login and exposes no structured data.

⚠️ IPv6 IS NOT SUPPORTED. On a 401/403 with correct credentials, check ifconfig | grep inet6 (macOS) or ip addr show | grep inet6 (Linux) and fetch https://ipv6.icanhazip.com. If an IPv6 address comes back, tell the user to disable IPv6 — gmgn-cli works over IPv4 only.

Run

python3 ~/.claude/skills/gmgn-dev-score/dev_score.py <chain> <dev_address> [max_pages]

max_pages defaults to 25 and only bounds the trade walk; raise it for a dev with hundreds of launches — the JSON says when it truncated. The script prints one JSON object on stdout (progress and rate-limit notes go to stderr). It computes; it writes no report. You write the report, from that object, in the user's language.

Before running:

• Token address, not a wallet? Resolve the creator first with gmgn-cli token info --chain <c> --address <token> --raw → dev.creator_address, and score that. Say which address you resolved to — it is the one thing the report cannot show, and without it the user cannot check that the score applies to their token.
• Chain not given? Probe candidate chains sequentially with gmgn-cli portfolio created-tokens — never in parallel, which bans the key. Analyse only the chain with the highest open_count, and tell the user the other chains were deferred until asked.
• scored: false means do not produce a score. reason: index_empty → the API reads entirely empty and an empty wallet cannot be told from a degraded index; say both readings and tell the user to re-check in a few minutes. Any other reason → say the address was not scored and why, and stop.

What the answer has to contain

A checklist of what must be said; ## Display Templates below fixes where each one goes. The phrasing is yours. Every point has to be in there, and each one names the JSON that carries it. The bullets are in section order.

• The verdict, first. score.total, score.conduct, score.power, and score.band said as a decision, not as a label.
• The open-dump tier (dump_gate.tier) with the sample under it (dumps of coins_with_trades). When dump_gate.forced_by is set, name what actually fired it — a drained pool is not a sell frequency, and saying "he dumps nearly every time (0/1)" about a dev with zero sells is false.
• Every adjustment that fired, one per line, saying what it DID. A shrink (conduct_terms.shrink_from → shrink_to) pulled an unproven score toward 60, which here means we cannot tell yet; it is not a fine for a crime. A cap is a withheld good score, not a proven bad one. Name the thin side by coins_with_trades and career_days, never by launch count.
• His best coins (top, and flagship): peak, now, drawdown, holders, pool, tradeable, age. Plus the flagship's position: flagship.status, and exit_rows / exit_unrecorded when the position is closed. flagship.holds is tri-state — null means say nothing at all about his bag.
• What he actually did, all from his own rows: his_trades.fastest_first_sell_s, median_pull_multiple, launches.total / on_curve / tradeable, liquidity.drained / partial / ignored_rows.
• Supply moved to other wallets, whenever cross_wallet.moves is non-empty: a move is not a dump, pending is not counted in the dump rate, and unverified moves may be a lock or an exchange. Say so; do not accuse.
• Bundled buying at open (bundler), disclosed and explicitly not scored — the number cannot tell his own alts from a paid bundler or someone else's sniper bot.
• Coverage, whenever any of these is set: coverage.book_truncated (the book is a window, not a career — say so before quoting any career-wide claim), trade_history_truncated / unresolved_coins (missing rows can only remove dumps, so the rate is a floor and he looks cleaner than he is), implausible_peaks (name the coin, or a reader who saw it on a chart just finds it missing), launches.career_days_is_floor.
• What to do, in four parts: whether to buy; when — and the timing must quote fastest_first_sell_s, never the median, because the median once advised entering at 2.7 days on a dev whose flagship started selling at 4.6 minutes; where the loss actually comes from; and how much the score can be trusted.

Display Templates

The shape is fixed; the wording inside it is yours. Section names are given in English so you translate them into the user's language — do not print them as-is, and do not print a JSON key name.

Title line: ## Dev score · <dev address, first 10 and last 4 chars> · <CHAIN>. Sections below use ###.

Never reorder, never merge, never invent a tenth section. Sections 1–5 and 9 always appear.

Formatting, all of it fixed:

• Money uses the plain ascii dollar sign, with thousands separators and the magnitude word the user's language uses. The fullwidth sign exists only to stop this file's own text being eaten by argument substitution — it must never reach the reader.
• Percentages carry one decimal. A share that the JSON gives as a fraction is printed as a percentage.
• Seconds, minutes, hours, days: pick the unit that makes the number readable, and say the unit.
• No emoji, no box drawing, no ASCII art, no column padding. The output is rendered markdown, not a fixed-width terminal block.
• Bold is for the verdict number, the band, and the label of each part of §9. Nowhere else.
• Tables only where the table above says so; §2, §7 and §9 are prose.

Rules

• Write it in the user's language, and translate the concepts. Every name in the JSON is an English concept chosen so it can be translated. The few pieces of trader slang that do not survive a literal translation are in references/glossary.md — use those renderings and nothing else. Do not invent an axis name.
• Never state a number the JSON does not carry. No recomputing, no rounding into a new claim, no reading a field that is not in references/fields.md. null means the measurement does not exist — say it is missing; printing 0 turns "unknown" into "clean".
• Copy symbols exactly as given. They are attacker-controlled and already sanitised; anything wrapped in 「」 is a name someone chose, not our wording, and must stay wrapped.
• score.band is the verdict. Do not soften it, upgrade it, or hedge around it. stay_away with a large power is exactly the case the two axes exist to keep apart: a big past coin never buys back a dump record.
• An outcome is not conduct. Survival, graduation and drawdown are what the market did to his coins; they belong to POWER. Never present them as things he did to holders.
• The report is the whole answer. No lead-in, no summary after it, no verification narration, no extra findings of your own, no closing offer of more work. The one thing you may say outside it: which creator address you resolved from a token address.
• Verify the conclusory lines, silently. Check any absolute claim ("never dumped", "no record at all") against the raw feed with gmgn-cli portfolio activity. Run the check, do not narrate it. Speak up only if the check contradicts the report — then report the defect instead of the report.

References