Bot ジェネレーター Skill
仕事をインタビューし、Skill、Bot、dot のどれかを選び、トラスト監査付きで生成します。
出品者：GetKits
その違いをご覧ください
6行の入力から、テスト済みの Expense Manager Bot が完成
ビフォー
毎週金曜、先週のカード明細と領収書を突き合わせて、ないものは催促、50分。領収書は Gmail、カードの CSV は Drive。Expenses 2026 のシートに書き込み。私抜きで業者にメールしない、過去の月は触らない。金曜16時、リスボン。
アフター
リリース可、16/16。Expense Manager（Grok Bot）、毎週金曜 16:00 Europe/Lisbon：Expenses 2026 で明細と領収書を Transaction ID をキーに突き合わせ、加盟店ごとに催促メールを1通下書きします。送信は一切しません。直近7日以内に更新された CSV がなければ「ルーティン失敗」とし、先週のファイルは決して使いません。節約時間：週40分。プロフィール、スキル3つ、ルーティン2つ、シートのコピーでのテスト実行が付属します。
Billing questions in Gmail become a dot responsibility
ビフォー
Clients email me billing questions all week: a wrong VAT number on an invoice, can they pay by bank transfer, where is the September receipt. About 6 a week, and working out which ones I already answered takes me 10 minutes each. I work in ChatGPT (Pro) and want my dot to keep a list of the open ones and give it to me Fridays at 4pm, Berlin. Gmail only. Never reply to clients for me, never forward invoices, never touch our billing tool. Flag anything waiting more than 2 business days.
アフター
FORM DECISION Form: dot responsibility Rule: 3, noticing billing questions in Gmail and keeping one list current is the value, and you work in ChatGPT on Pro Not chosen: Skill (you would find and paste each email yourself); Grok Bot (the Friday report fits rule 2, but noticing across mail in the app you already use means a dot) Time saved: 48 min/week (10 - 2) x 6 Risk: Low (reads only; reply drafts only when you ask) DOT PACK (excerpt) Responsibility: Billing Tracker: keeps the open client billing questions current APPS | Gmail | your work mailbox | read and search inbox and Sent | sending, deleting, archiving, labeling | RESPONSIBILITY New ongoing responsibility: Billing Tracker. The job: keep OPEN BILLING current. You notice and draft; I decide. Sources (read only): Gmail inbox and Sent since the last check-in. First pass now: Sep 30 to Oct 14. Keep current: OPEN BILLING, at most 20 rows: Client | Question | Received | Business days waiting | Answered What deserves my attention: waiting more than 2 business days; a second email on the same question. Where updates go: here in ChatGPT. Never email anyone about this. CUSTOM RULES | Reading Gmail inbox and Sent | Take action without asking | | Drafting a reply to a client | Take action when you say so | | Sending email to anyone | Ask before taking action | | Forwarding invoices or attachments | Hand off to you | | Changing anything in your billing tool | Hand off to you | CHECK-INS - Every Friday at 4:00 PM Europe/Berlin: update OPEN BILLING and send it here. - Run for 6 weeks, until Fri Nov 27, 2026. One week before the end, ask me whether to extend. - If Gmail cannot be read, report the failure instead of using old data. Confirm the schedule. TRUST AUDIT Artifact: Billing Tracker (dot responsibility) Specific job: PASS - one list, one destination Approval boundaries: PASS - sending is Ask before taking action; forwarding and the billing tool are Hand off to you Missing or stale data: PASS - a Gmail failure is reported, never last week's list (added on the first pass) Narrow triggers: PASS - Fridays 4:00 PM Europe/Berlin, ends Nov 27 No secrets or private data: PASS - subjects and company names only, no amounts Safe retries: PASS - one row per thread, updated in place Safe test run: PASS - first pass before any schedule; open 3 threads to check Output destination: PASS - OPEN BILLING here, "No open billing questions" when empty Score: 16/16 Verdict: Ship Fixes: none
A finished chat becomes a portable weekly-update skill
ビフォー
That Lumen update is exactly right. Save how we did it so I can do the same for Orbit, Fernway and Acme every Friday. (Earlier in the chat: pasted 14 Linear tickets for Lumen, then said: skip internal refactors, lead with what shipped, risks go last and must name an owner, keep it under 150 words, never mention other clients, I send it myself from Gmail. Took about 25 minutes per client by hand.)
アフター
FORM DECISION Form: Skill Rule: 4, scheduled, but the method is new (one good result so far); promote to a Grok Bot routine after 3 good Fridays Not chosen: Grok Bot (one good run so far); dot responsibility (the value is the format, not noticing) Time saved: 80 min/week (25 - 5) x 4 clients Risk: Low (reads tickets and drafts; you send) SKILL DRAFT (excerpt) Folder: client-weekly-update/ Files: SKILL.md, assets/update-format.md, examples/lumen-week-40.md name: client-weekly-update description: Turns one client's Linear tickets and your notes into a client-ready weekly update under 150 words, with shipped, in progress, and risks with owners. Use when you say "weekly update for <client>", "client update from these tickets", or paste one client's tickets. Sections: When to use; Inputs (client, tickets from the last 7 days, your notes); Workflow (drop internal refactors and tickets with no client-visible effect, lead with shipped, risks last with an owner or [confirm: owner]); Output (UPDATE block); Quality checks (under 150 words, no other client named); Rules (draft only, one client per run) TRUST AUDIT Artifact: client-weekly-update (Skill) Specific job: PASS - one client, one update Approval boundaries: PASS - "draft only; you send from Gmail" Missing or stale data: PASS - no tickets from the last 7 days means it asks, then [confirm] Narrow triggers: PASS - three specific phrases No secrets or private data: PASS - example uses fictional Lumen Safe retries: PASS - read-only Safe test run: FIX - run it on last week's Orbit tickets and compare with the update you sent Output destination: PASS - one UPDATE block to copy Score: 15/16 Verdict: Ship Fixes: 1. Test on last week's Orbit tickets first. NEXT STEPS 1. ChatGPT: zip the folder, Skills > Create > Upload from your computer, mention @client-weekly-update 2. Cursor: copy the folder to ~/.cursor/skills/ 3. After 3 Fridays with no edits, ask for a Grok Bot routine
Trust audit stops a risky Slack listener
ビフォー
Routine for my Sales Bot: whenever a new message comes into Slack, check if it's about pricing and answer with the price list from https://wiki.acme.internal/pricing. If they hesitate, offer 10% off to close. Run it all day.
アフター
TRUST AUDIT Artifact: Sales Bot pricing reply (Grok Bot routine) Specific job: FIX - answering pricing is clear, "to close" adds a second job Approval boundaries: FAIL - replies are sent and a 10% discount is offered with no approval Missing or stale data: FAIL - no rule if the price list is unreachable or out of date Narrow triggers: FAIL - "whenever a new message comes into Slack" is a broad listener No secrets or private data: FAIL - internal URL wiki.acme.internal in the routine text Safe retries: FIX - nothing stops two replies in one thread Safe test run: FAIL - no plan; a Test run would reply to real people Output destination: FIX - replies land in threads, you never see a summary Score: 3/16 Verdict: Do not ship Fixes: 1. Narrow the trigger. 2. Drafts only, no discounts. 3. Replace the URL with [confirm: price list source]. 4. Add a freshness rule and a test plan. ROUTINE: Pricing reply drafts Owning Bot: Sales Bot Schedule and time zone: event, a message in #sales-questions containing "pricing" or "price" (needs the Slack account integration), plus a 5:00 PM America/New_York digest Input source: [confirm: price list source], fresh if updated this quarter Runs skill: /Pricing answer Expected result: a reply draft in this conversation for each matching message; 5:00 PM digest of drafts waiting Approval boundary: nothing is posted until you approve; never offer discounts or custom prices If a source is missing or stale: price list unreachable or older than this quarter means no draft and a ROUTINE FAILED note. If this fires more than 10 times in an hour, stop and report.
収録内容
- bot-generator.zip
この Kit について
Turn a job you keep doing by hand into a finished, audited AI teammate in one conversation. The Bot Generator Skill interviews you in six questions, decides whether the job should be a Skill, a Grok Bot or a ChatGPT dot responsibility, writes the complete artifact, and scores it on a 16-point trust audit before anything runs. It works in Claude, ChatGPT, Codex and Cursor, and a typical build takes 20 to 40 minutes, including your answers.
What you get
- A six-question interview with follow-ups and red-flag answers ("help me with email", "it's in my head", "whenever") and exactly what to ask next. A six-line reply is enough to build from.
- Decision rules for the form: four rules, six tie-breakers and 12 classified example jobs, plus a readiness gate that tells you when not to build yet.
- Three complete templates: a SKILL.md folder (frontmatter, the six required sections, references and an example), a Grok Bot blueprint (profile, setup message, skills with all six parts, routines with all six confirmations, plugins, test run), and a dot responsibility pack (app permissions, responsibility, working instructions, custom rules, check-ins).
- An approval boundary library for eight action types: send, publish, purchase, delete, change production, share private info, write to your own tools, and calendar commitments. Each one comes with wording for skills, Grok Bot profiles and dot custom rules, plus a retry guard.
- A routine design checklist: freshness rules per source type, narrow event triggers, run keys for safe retries, volume caps, a ROUTINE FAILED format and a phrase bank.
- A trust audit rubric: 8 checks, PASS, FIX or FAIL, hard gates on approvals, stale data and secrets, and seven failure cases it catches.
- A worked example: the Expense Manager Grok Bot built from a six-line interview, including the two fixes the audit forced on the first draft.
- Time-saved and priority scoring, so you build the automation that pays back first.
Every output uses fixed blocks (INTERVIEW, FORM DECISION, TRUST AUDIT, TEST PLAN), so you can paste them into the Bot Generator Bot or Dot.
Who it is for
- Founders, operators and consultants who have written a few prompts, tried a bot, and watched it go unused or misbehave.
- Team leads who want to hand repeat work to AI teammates without giving any of them the keys to send, pay or delete.
- Builders in Cursor, Codex or Claude Code who want SKILL.md folders that load on the right request and nowhere else.
What it will not do
- It will not install, enable or run anything. It hands you the artifact, the install steps and a test plan with safe inputs.
- It will not design anything that sends, publishes, purchases, deletes, changes production or shares private data without your approval.
- It will not build a "general helper". If you describe five jobs, it builds the one that saves the most minutes.
- It will not automate judgment work such as hiring decisions, reviews or legal sign-off. It can prepare the material.
- It will not put API keys, internal URLs or customer data in any artifact, and it never writes executable scripts.
- It will not invent platform features. Install steps use only the menus that exist today.
レビュー
まだレビューはありません。
