菜单
Checks whether a Swan workspace is structurally healthy. Use for configuration audits that surface missing knowledge, disconnected integrations, exhausted senders, broken triggers, sparse CRM data, and credit or subscription issues.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use this skill when the user wants to submit, contribute, or publish a GTM skill to the GTM Skills library (github.com/swan-gtm/gtm-skills, gtmskills.com), or update one of their existing skills or their creator profile there.
Use this skill when [the concrete trigger — a signal fires, a moment in the deal, a recurring cadence]. Produces [the concrete output]. [One or two more sentences dense with the phrases someone would actually say when they need this.]
Use this skill when a deal closes won and you want to turn the win into pipeline immediately. It extracts the win profile, finds lookalike companies not yet in your CRM, researches the right persona contact at each, drafts outreach from the deal owner using an anonymized win reference, and queues a single batch review task — nothing sends without human approval.
Use this skill when scoring a paying customer (closed-won account) for expansion. It evaluates product usage depth, persona quality, and expansion headroom to produce an expansion tier with short actionable reasoning. Designed as a sub-skill of your lead-scoring methodology — the parent runs the ACV assessment first and this returns a tier to it.
Use this skill when a profile-view webhook (e.g. LeadShark) fires for a new view on a team member's LinkedIn profile. It resolves the viewer's identity and company, applies dedup, internal, relationship, and ICP gates, and routes qualified viewers into the MQL pipeline — treating a profile view exactly like a website visit. The CRM is never polluted: only fully resolved, ICP-matched viewers reach a CRM write.
Use this skill when a tracked influencer publishes a LinkedIn post focused on your product (sponsored, review, or organic mention) and you want to capture the engagers as leads. It verifies the post is genuinely about you, scrapes the full engager list, filters to ICP-fit companies and buyer personas, runs lead scoring, and posts per-lead alerts plus a reconciled digest. Detection and scoring only — no outreach.
| name | audit-swan |
| title | Audit Swan |
| description | Checks whether a Swan workspace is structurally healthy. Use for configuration audits that surface missing knowledge, disconnected integrations, exhausted senders, broken triggers, sparse CRM data, and credit or subscription issues. |
| category | Ops |
Eight structural dimensions. Each is small — most produce 0-2 findings. Process them one at a time, write down the verdict (OK / WARN / FAIL) + a one-line summary, and move on. Do not try to gather all eight dimensions' raw data before reasoning. By the time you reach the report, you should have eight short verdicts in hand, nothing more.
swan-get-subscription. Check three things:
swan-search-sequences recent-activity sample (page 10) — don't pull the full log.Verdict:
One-line summary. Move on.
swan-get-memory for org memory.
Required:
Verdict:
One-line summary.
swan-get-org-senders (this returns a small list — usually < 10 senders). Per sender, check auth status and capacity headroom.
Verdict:
One-line summary.
For each integration referenced by an active trigger:
If integration is referenced by an active trigger but not connected: FAIL.
Don't enumerate integrations the org doesn't use. One-line summary.
Pull active triggers, page size 10. For each:
Most triggers are healthy. Stop after the first 10 unless ≥ 3 problems surface — then pull more.
Verdict per trigger group, summarized to one line.
Only if a CRM is connected. Two tiny calls:
hubspot-list-object-properties for companies. Check that the properties active plays depend on exist.hubspot-search-objects with page size 10 to sample populated-ness of hubspot_owner_id + lifecycle stage. 10 is enough to spot a hygiene problem.Verdict:
Pull execution log, last 7 days, page size 20.
If > 20 failures returned in the truncated preview: switch to swan-execute-code. Read files/tool-outputs/<filename>, group by errorClass in pandas, print classes with count > 3.
Verdict based on top class count: > 10 → FAIL, 3-10 → WARN, < 3 → OK.
Quick check: any orphan child skills (path points to a deleted parent)? Any duplicate display names across personal + org scope?
Usually clean. If WARN, one-line note. Otherwise skip.
After all eight dimensions, the agent has eight one-line verdicts. Compose:
SWAN DOCTOR REPORT
FIX NOW:
- <every FAIL, one line each with the exact tool to use>
PAY ATTENTION:
- <every WARN, one line each>
ALL CLEAR:
- <every OK dimension, one line each — "Subscription: OK ($45 credit, ~28 days runway")>
NEXT STEP:
<one sentence: the single highest-impact thing to fix>
If no FAILs: lead with "no critical issues" before the WARN section.