メニュー
The 5-minute morning GTM ritual — read the workbook's Budget/Log, Queue, and Companies tabs, inline a short digest (new triggers, queue state, invites vs cap, acceptance %, budget), append the wave-log line, then offer exactly three actions.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Harvest LinkedIn-reachable buyer and champion contacts for the head-of-queue companies in the workbook, then build a $0 keyless per-company buying map (org roles, reporting edges, approach order) before any paid call, then reveal verified emails ONLY for the head of queue via the Apollo to Hunter to None waterfall. Contact harvest runs through LinkedIn-dork subagents that trust link titles and ignore SERP prose and return rows as text; the main session writes the People tab. Reveal is budget-gated and ask-first, logs every credit to the Budget tab, and never guesses an email. Run after targets has populated and tiered Companies, once per daily slice, before drafting messages.
Draft two-message DM packs (connect note + post-accept DM, email variant only on a verified email) for People rows that lack messages, write them to the workbook's Messages tab with claims and dial columns, fact-check every claim at the caution dial, and set STATUS per the dial gate. Use when the user asks to draft outreach, write messages, fact-check drafts, or re-check rows they edited in the sheet.
Copy passed Messages rows into today's Queue tab (ranked, capped at daily-N, both DM messages + LinkedIn links), render the self-contained HTML cockpit from that tab, then run the STATUS paste-back loop. User-invoked and side-effecting — never auto-run.
The Galois GTM setup interview. Creates the campaign folder and workbook (~/GaloisGTM/<campaign>/gtm.xlsx), then builds and confirms galois.profile.yaml beside it (ICP plus confirmed adjacent profiles, per-segment discovery plan, personas, voice, caution dial, channels, budgets). Use when the user says "set up galois" / "set up galois-gtm", "new campaign", "interview me", "onboard me", drops a URL or doc to profile from, or when any galois-gtm skill finds no confirmed galois.profile.yaml next to the workbook. Also handles single-field tunes and drift check-ins.
Discover the day's target companies straight into the workbook's Companies tab and tier them into a trigger-sorted slice of 10-20. Run the profile's discovery plan — gate-matched lanes only: web-search and hiring-signal discovery for any vertical, registry pulls (openFDA / USAspending / SEC EDGAR) solely when the ICP trait matches, plus any CSV the human dropped in the folder — dedupe in the Companies tab itself, apply the 2+ lane multi-signal gate as a tier-A proposal, rank by trigger recency, then subagent-tier the single-signal residue (subagents RETURN rows, the main session writes). Use to build or refresh the target list before /galois-enrich.
| name | daily |
| description | The 5-minute morning GTM ritual — read the workbook's Budget/Log, Queue, and Companies tabs, inline a short digest (new triggers, queue state, invites vs cap, acceptance %, budget), append the wave-log line, then offer exactly three actions. |
| when_to_use | Start of a prospecting day, a morning check-in, or when the user asks "what's on deck / what should I do today" for outreach. Keep it to five minutes. |
| argument-hint |
If a structured question tool is available (
AskUserQuestionon Claude Code — one question per call;ask_user_questionon Codex — at most 3 per turn, 2–3 options plus free text, pause the auto-resolve countdown), use it. Otherwise restate each question verbatim, then lettered options with the recommended one marked, and accept a letter, "recommended", or the user's own answer. Never present options without restating the question (the spec-kit #1147 lesson). On Claude Desktop, at most 2 questions per message.
A five-minute retention loop, not a work session. Read the workbook, open with the digest, then offer exactly three actions and stop. Do not run the research pipeline from here (that is /galois-targets), do not re-interview, do not chase rabbit holes.
The workbook is the state: ~/GaloisGTM/<campaign>/gtm.xlsx. Locate it — if exactly one folder exists under ~/GaloisGTM/ use that campaign; else ask which campaign (structured question per the preamble). Read it in the main session with inline python3 + openpyxl (CSV fallback = <campaign>/sheets/*.csv, same headers); on Desktop, the native xlsx tooling on the uploaded file. No galois CLI, no subagents — one deterministic read.
Read three tabs and compute a few numbers — do not dump raw rows:
monthly_cap_usd, spent_usd, apollo_credits, hunter_finds) → budget remaining; and the wave log (row 4+) → yesterday's last wave line.date and STATUS: how many are unsent (blank), sent, accepted, replied; invites this week = li_note rows marked sent/accepted/replied in the last 7 days, against the 80 ceiling; acceptance % = accepted+replied over sent-or-further among li_note rows.signals landed since the last wave's timestamp.python3 - <<'EOF'
import openpyxl, os
p = os.path.expanduser("~/GaloisGTM/<campaign>/gtm.xlsx")
wb = openpyxl.load_workbook(p, data_only=True)
def rows(name):
ws = wb[name]; hdr = [c.value for c in ws[1]]
return [dict(zip(hdr, [c.value for c in r])) for r in ws.iter_rows(min_row=2)]
# summarize Budget spend block + wave log, Queue STATUS this week, Companies new signals
EOF
Summarize it back in two or three lines — new triggers, queue state + invites/cap + acceptance %, budget remaining. Then append one wave-log line to the Budget tab's wave log (the gstack-style manifest): ts | stage=daily | surface | rows_added=0 | rows_changed=<statuses touched> | credits_spent=0 | note="morning digest". Write it with the standard discipline (copy to gtm.prev.xlsx, save to gtm.tmp.xlsx, atomic os.replace), and say the close-the-file warning before that write.
Ask one question (structured tool or the verbatim-restatement fallback per the preamble). Exactly these three, no more:
/galois-queue (copy ready Messages rows into the Queue tab, render the HTML cockpit from it, open it).STATUS / STATUS_NOTE cells from it (match on date, person, company, channel); otherwise they mark sent/accepted/replied/skip/bounced straight in the Queue tab. This is what makes tomorrow's queue and the acceptance meter reflect reality — nudge it if the STATUS column looks stale./galois-targets --since 1d (only triggers new since yesterday, appended to the Companies tab; feeds tomorrow's queue).Do the one action the user picks, then stop. One action per morning is the whole ritual.
Acceptance under 40% → warn; under 25% → suggest pausing invites (low acceptance shrinks LinkedIn's allowed volume). Acceptance and reply rates come straight from the Queue STATUS marks, so the usual fix is B above (log statuses), not a profile edit. If the shape of what's landing has genuinely drifted from the ICP, offer /galois-setup --tune <field> (a check-in of at most three questions, drifted fields only — never a re-interview). It is an offer, not a detour: if the user declines, move on.
Wave outro (mandatory): "Wave done: digest read, 1 wave-log line appended — edit anything in gtm.xlsx (STATUS marks, TIER calls); the next wave reads whatever is there."
Optional: deterministic CLI. Power users may instead run the old galois digest engine from the repo's src/ appendix; no skill requires it, and the workbook flow above is complete without it.