Gmgn Wallet Analysis

BEFORE RUNNING ANY COMMAND: Run gmgn-cli config --check. If exit code is 0, proceed normally. If exit code is 1, (1) run gmgn-cli config and show the output to the user; (2) once the user sends the API Key, run gmgn-cli config --apply <KEY>, then show the output. If --check errors with an unknown option or command-not-found, tell the user to run npm install -g gmgn-cli, then retry.

IMPORTANT: Always use gmgn-cli. Do NOT use web search, WebFetch, curl, or visit gmgn.ai — the website requires login and does not expose structured data.

⚠️ IPv6 NOT SUPPORTED: On a 401/403 with credentials that look correct, check IPv6 immediately — run ifconfig | grep inet6 (macOS) or ip addr show | grep inet6 (Linux), and request https://ipv6.icanhazip.com. If outbound traffic is IPv6, tell the user: "Please disable IPv6 — gmgn-cli only works over IPv4."

What this skill is for, and what it is not

Two skills take a wallet address. They answer different questions and must not be substituted for each other:

The distinction that matters: a score compresses everything into a number that hides why. This skill refuses to. It runs four gates, each of which can independently veto or downgrade the verdict, and each of which prints the single number that decided it. A wallet can be a genuinely excellent trader and still be un-copyable — those are separate gates here, not one blended score.

It also puts who this wallet is near the top, before any verdict. Someone who pasted an address wants to know whose wallet it is — the bound X account, the follower count, where the money came from — before they are asked to absorb four gate results. Burying that below the gates made a newcomer scroll past four judgements to reach the one fact they came for.

It also asks three questions a score does not:

• Is the edge still working this week? A wallet with +2,400% all-time and −27% over the last 7 days looks superb on any leaderboard and will lose you money today. profits --period all versus stats --period 7d is the cheapest way to see that, and it is the most common way a copy-trader gets hurt.
• What is it doing right now? Closed-trade statistics describe the past. Its open positions and its last 24 hours are what you can still act on.
• Where did the profit come from — speed, or selection? per_day × the top-3 winners' share of gains separates 🕸️ spray-and-hit (attempts × a few hits — copying it is a latency race) from ⚙️ turnover grind (each exit too thin to survive your slippage) from 🎯 pick-and-size (the one you can follow a step behind). Those are copied in completely different ways.

The four gates

Each gate returns ✅ pass, ❌ fail, or ⚪ unevaluated. ⚪ never renders as ✅ — "we could not measure this" and "this is fine" are different statements, and conflating them is how a dossier lies.

The verdict headline names its own cause — DO NOT COPY · the profit is self-dealt rather than a generic the record does not hold — so the reason is legible without reading the gate detail.

Verdict is a pure function of the gates — G1 and G2 are vetoes, G3 and G4 change what you do rather than whether you act:

Eleven outcomes, short-circuited in this order. Every one has a fixture:

Three of these exist only because the first cut got them wrong, and none should be collapsed:

• Row 5 is ⚪, not 🔴. A four-token wallet had nothing bad measured — it had nothing measured. Rendering an unmeasured gate as a red verdict is the same error as rendering ⚪ as ✅, in the other direction. 🔴 means measured and bad; ⚪ means not measured.
• Row 7 exists because G3 used to short-circuit G4. A wallet you cannot get filled on and that rides positions to zero was being told to the reader as a signal source with no mention of the second half. Both problems are independent, so both sentences appear.
• Rows 2 / 10 / 12 are the same tag with three different answers — corroborated, not checkable, refuted. Their three fixtures pin all outcomes of wash_trader; a change that makes any two agree is a regression.

The opening three lines are the hook

A reader who pastes an address should be able to rank the trader from the heading alone — smart money reading as smart money, a losing wallet reading as one — before any judgement about copyability. The card opens with a caliber grade, then the record as a sentence, then the 7-day window as a backtest, and only then the copy verdict.

Three separate mistakes produced this shape, and each fix must survive an edit:

1. Never frame a past window as a return the reader could have captured. The first version opened 如果 7 天前跟了它 $1,000 → $1,000 → $1,206. That is a completed counterfactual: it describes a closed opportunity, and the reaction it drew was "then I've already missed it". Same figure, reframed as 近 7 天回测 — what the window measured about the wallet — reads as proof of skill. Say what the window measured, never what the reader would have earned.

2. Ratios are not rankable; counts and a grade are. 69.3% profitable, 1.2% heavy losses is a spec sheet — a newcomer cannot tell whether those are excellent or ordinary. 810 of 1,023 made money, only 12 lost more than half reads on sight, and the grade in the heading supplies the ranking outright.

3. The grade is current standing, not lifetime. caliber() reads realized money, win rate, heavy-loss share and sample size — but G2 failing overrides all of it with 📉 was good, not any more. Without that override a wallet with $298K lifetime and a −37.6% week printed 💪 seriously good directly above 🔴 DO NOT COPY · it has stopped making money, which is the heading endorsing what the body forbids.

Caliber deliberately excludes reachability and loss-cutting: an unreachable wallet can still be an excellent trader, and merging those is exactly what a single blended score gets wrong. It is the other half of the answer, not a duplicate of the verdict.

Two shapes to preserve:

• G1 ❌ prints no figure at all. The grade becomes ⚪, and the record sentence and backtest block are both replaced by one line saying the profit figures are not trustworthy. Quoting $1,000 → $1,206 where the record itself is in dispute presents a disputed number as an achieved one.
• A loss is never a gain with a minus sign. it made -$183.8K this week was a real bug; negative weeks take the it lost {0} string.

One number, one meaning, everywhere

winners and winrate count different things and must never be presented as one ratio:

• winners — every position currently in profit, realized or not. It is a count of coins.
• winrate — the API's realized win rate, over positions the wallet has actually sold.

G1 used to print 133 tokens, 115 profitable (45.9%), which asserts that 115/133 is 45.9%. It is 86.5%. The parenthetical belonged to a different numerator, and a reader who did the division found the report contradicting itself two lines into its own evidence.

They are now labelled separately, and the count that explains the gap prints beside them whenever dist_gap is set — positions bought and not yet sold sit at 0% inside the 0–200% band, so they inflate winners without ever reaching winrate:

107 币 · 89 个在赚 · 卖掉的部分胜率 44.4%（其中 34 个买了还没卖，没有已实现结果） · 集中度 46.1%

The card names the same number with the same verb — 89 个现在是赚的 — so no reconciliation is required anywhere in the report.

This defect was introduced by a deletion, not by a change: the reconciliation used to be a note under the outcome-distribution histogram, and removing the histogram removed the note while leaving the contradiction it resolved. When deleting a figure, check what else was relying on it to make sense.

Verdict language

This layer is the only part most readers finish, so it is written to be read once:

The action must never restate the gate reason printed below it — the reader would be reading the same sentence twice before reaching anything new.

Third-party labels are questions, never findings

common.tags is the highest-information-per-byte field in the response and the one most easily mis-read. A tag is a heuristic label: it opens a question that the wallet's own behaviour has to answer, and no tag may change the verdict on its own. Obeying wash_trader unexamined once rendered 🔴 DO NOT COPY on a real BSC wallet whose $459K of realized profit came from six-figure memecoin positions, while the tag was firing on a ~$1K sliver of tokenised-stock churn.

token.is_honeypot gets the same treatment. It ships inline on every holdings row, but a honeypot is a token you cannot sell, so history_total_sells > 0 on the same row refutes the flag by construction. One live run failed G4 on seven such flags, one of which the wallet had sold 101 times — transfer-restricted tokenised stocks that trip naive sell simulators.

Corroborate silently — never print the report disagreeing with a GMGN tag

This skill ships inside GMGN's own product. Citing a GMGN tag and then telling the reader it does not hold advertises an internal heuristic and undermines it in the same breath. So the report states the measurement and never names the label:

An unverifiable accusation is not a finding: never manufacture a 🔴 out of a tag you could not check. This rule binds your own prose too — do not reintroduce the label in your reply when the report deliberately left it out.

WHAT TO DO NEXT is exactly three questions

The report ends with three follow-up questions and nothing else — one intent each, phrased the way the reader would ask them. Three rules govern them:

• No skill names. A skill name asks the reader to know it exists, that it is installed, and how to invoke it. A question in plain language routes itself, because asking it is what triggers the skill that answers it.
• Answerable, and pinned on a real trigger phrase. Every question routes to a skill that ships in GMGNAI/gmgn-skills and carries a literal trigger phrase from that skill's own description: 筹码分析 → gmgn-holder-analysis, 跟单评分 / 发盘情况怎么样 → gmgn-wallet-score, 走势 · 形态 → gmgn-kline-pattern, 安全 → gmgn-token, 聪明钱 · KOL → gmgn-track, 持仓 → gmgn-portfolio.
• One target per question. A token-scoped question names the single largest of the last 24h buys, never a list. Three tokens in one line become three downstream skill calls — three times the rate-limit budget in the turn right after a dossier that already nearly empties the bucket.

Do not add follow-ups of your own beyond these, and do not name a skill when you offer one.

Metrics the report names

The script computes these; know what they mean so you can answer a question about one.