Bill Rausch

Ask first

• Where do recordings or transcripts land today? (default one intake folder)
• How should folders be split? (default Clients, Prospects, Partners, Internal)
• What happens when no company matches? (default leave it in intake and flag it)

Needs

• Transcripts as files in one folder, from any call recorder.
• Calendar access, to find the meeting by start time.
• A list of companies with a stage (client, prospect, partner). Fallback: a spreadsheet of domains.
• A small store that remembers which files are done.
• If your assistant can only chat: once a week, paste the file names and meeting list and ask where each belongs.

Build

1. Trigger on each new file in the intake folder.
2. Read the file's timestamp and find the calendar event within 15 minutes of it.
3. Collect attendee domains and drop your own and free-mail domains.
4. Decide: one matched company = file it under that company's stage folder. Several matches = take the one with the most attendees. None = hold it.
5. Move the file and rename it "date - company - meeting title".
6. Write the file ID to the ledger. Skip any file already there.
7. Post a one-line log message.

Check

• Drop three test files: one matching a client, one with only free-mail attendees, one duplicate. Expect one filed, one held, one skipped.

In my version

• AWS Lambda on an EventBridge schedule, a Google service account to read calendars, HubSpot for lifecycle stage, an S3 ledger.
• Feeds "Search all our call transcripts".

Ask first

• What time should the evening run land? (default 6pm)
• What time should the morning refresh land? (default 6:30am)
• Where should it arrive? (default a direct message in your chat tool)
• Which domains are internal, so those meetings are skipped? (default your email domain)
• How far back should news go? (default 14 days)

Needs

• Your calendar for the next working day.
• A list of the companies you work with and their status (client, prospect, partner). Fallback: the CRM, or a plain spreadsheet.
• Open tasks with due dates from your CRM or task list.
• Web search for company news.
• The CRM connection reads companies, contacts and tasks; it changes nothing.
• If your assistant can only chat: paste tomorrow's calendar and your task list each evening and ask for the brief.

Build

1. Read tomorrow's events (morning run: today's). Drop events with only internal attendees, and declined or all-day events.
2. For each external domain, match the company in your list and pull its status, owner and last activity.
3. Search news on each matched company from the last 14 days. Keep at most 2 items that bear on the meeting; drop press-release noise.
4. Pull open tasks due on or before tomorrow, marking overdue ones.
5. Decide one suggested angle per external meeting, quoting a fact from steps 2 or 3. If there is no fact, write "no angle".
6. Morning run only: compare against the saved evening brief and list moved, added or cancelled meetings and new attendees under "Changed overnight".
7. Deliver one message. Save the brief to a dated file, so the morning run has something to compare against.

Check

• Run it now with delivery off for today's calendar. Every external meeting must name a matched company or say "not in your list", and every angle must quote a fact.
• Move a test event, run the morning step, and confirm it appears under "Changed overnight".

In my version

• Claude scheduled tasks with HubSpot, Google Calendar, Slack and Google Drive connected.
• Runs at night and at dawn. Delivered to my Slack DM.
• The same run also does "Chase the replies I'm still waiting on".

Ask first

• What label or tag will you put on sent emails that need a reply? (default "Need Response")
• How many days before a nudge is suggested? (default 3 business days)
• Should Slack or chat replies from the same person count as an answer? (default yes)

Needs

• Your email, with labels on sent mail.
• Your team chat, searchable by person. Fallback: email only.
• The email connection reads labeled threads; it does not send anything.
• If your assistant can only chat: paste the list of emails you are waiting on each morning and ask which to nudge.

Build

1. Read every thread carrying the label.
2. For each one, look for a reply from any recipient after your last message, in email or in chat from the same person.
3. Decide: a reply exists = resolved. No reply and older than the nudge window = needs nudge. Otherwise = waiting.
4. For each "needs nudge" item, draft a two-line follow-up in your voice. Do not send it.
5. Deliver the list in the morning brief, grouped as Still waiting and Resolved.
6. Remember which threads were reported resolved, so they show once and then drop off.

Check

• Label two old threads, one answered and one not. Run with delivery off: one must show Resolved, the other Still waiting with a draft.

In my version

• Part of my day-prep skill. Gmail label "0 Need Response", with Slack checked too.
• Drafts only. I send nudges myself.

Ask first

• Which time-tracking tool and which planning or BI tool? (default whatever you use today)
• How far back should each run reload? (default the last 35 days, to catch late edits)
• At what grain? (default person x client x month)

Needs

• API access to your time-tracking tool.
• An import endpoint or connector in your planning tool. Fallback: a CSV it can pick up.
• A place to run on a schedule (CI, a cloud function, or a cron job).
• If your assistant can only chat: export a time report monthly and ask it to reshape it into your model's import format.

Build

1. Pull time entries for the reload window.
2. Map each person, client and project to the model's lists. Collect any unmapped item.
3. Sum hours to the chosen grain.
4. Decide: unmapped people or clients are added only if they have more than 1 hour, otherwise reported.
5. Replace the window's values in the model so reruns are safe.
6. Post a summary of rows read, written and skipped.

Check

• Run against a test copy of the model with one month of data and compare the totals to the time tool's own report. They must match to the hour.

In my version

• Ruddr into Pigment, a Python script run by GitHub Actions on a schedule.
• Earlier history from my old time tool stays in a separate metric.

Ask first

• Which system is the source of truth? (default the CRM)
• Which stages count? (default client, prospect, partner)
• Which fields should copy? (default name, domain, stage, owner)

Needs

• CRM API read access.
• A Notion or other database with matching columns.
• A scheduler.
• If your assistant can only chat: export the CRM list weekly and ask for the rows that changed.

Build

1. Read companies in the chosen stages.
2. Read the destination database.
3. Match on domain, then on name.
4. Decide create, update or unchanged. The source wins every conflict, and the old value is logged.
5. Write the changes.
6. Post counts.

Check

• Dry-run with writes off. The create and update counts should equal the companies added or changed in the CRM since the last run.

In my version

• HubSpot to Notion, Python on AWS, once a day.
• Feeds the filing and search routines.

Ask first

• Which meeting words trigger it? (default Demo, Walkthrough, Scoping)
• Prep and overrun lengths? (default 15 minutes before, 30 after)
• What if the slot is taken? (default skip and report)

Needs

• Calendar read and write access.
• If your assistant can only chat: paste your week and ask where the blocks fit, then add them yourself.

Build

1. Read the next 14 days.
2. Find events whose title contains a trigger word.
3. Decide: add a block only if that time is free and no block already exists.
4. Create the blocks as busy, titled Prep or Overage.
5. List what was added and what was skipped.

Check

• Run it as a preview with writes off on a week that has one demo next to another meeting. Expect one block proposed and one skipped.

In my version

• Google Calendar connected to Claude. 15 minutes of prep, 30 minutes of overage.

Ask first

• Which people, hashtags or subreddits? (default 20 people, 3 subreddits)
• How far back? (default 7 days)

Needs

• A scraper or search tool for LinkedIn and Reddit. Fallback: paste posts.
• If your assistant can only chat: paste 10 posts and ask for the themes.

Build

1. Pull posts from the window.
2. Group them into themes. A theme needs 3 or more posts.
3. Decide "worth a comment": newer than 3 days, active, and in your lane.
4. Write the brief.

Check

• Every theme must link back to at least 3 source posts.

In my version

• A Claude skill using Apify scrapers.

Ask first

• Who should see which calls? (default shared client and team calls, plus a private folder per person)
• Where are transcripts stored? (default the folders from "File every call recording")

Needs

• Transcripts as text files, sorted by company.
• A vector search or knowledge-base service.
• Sign-in that can tell users apart.
• If your assistant can only chat: upload a handful of transcripts into one project and ask questions there.

Build

1. Index each folder into its own knowledge base.
2. Put each knowledge base behind its own access rule: shared ones for the team, private ones per person.
3. Expose them to your assistant as tools.
4. Decide: answers must quote the transcript and give the call date. If nothing matches, say so.

Check

• Ask about a call you remember well. The quote must match the file.
• Sign in as a second user and confirm a private call is not returned.

In my version

• AWS Bedrock knowledge bases behind AgentCore Gateway, with Cognito per person and an S3 prefix per person.

Ask first

• Which sections do you want? (default business model, people, use cases, focus, risks, quotes, POC)
• How many quotes? (default 5 to 8)

Needs

• Call transcripts or notes for the account.
• If your assistant can only chat: paste the transcripts and ask for the record.

Build

1. Read every transcript and note for the account.
2. Pull facts for each section, citing the call they came from.
3. Decide quotes: keep the lines that show pain or a priority, and drop small talk.
4. Write the record and flag any section with no evidence.

Check

• Every quote must appear word for word in a transcript. Search for three of them.

In my version

• A Claude skill run inside each prospect's Claude project before I archive it.

Ask first

• Paste 10 posts or emails you are proud of. (default your last 10 LinkedIn posts)
• Which words do you never use? (default none)
• Which formats? (default post, comment, email, article)

Needs

• Samples of your own writing.
• If your assistant can only chat: keep the profile in a project and paste a draft each time.

Build

1. Build a voice profile from the samples: openings, sentence length, words you avoid, how you end.
2. Draft mode: write to the profile, then check it against the profile.
3. Critique mode: score 1 to 10 against each trait, list the misses, rewrite.

Check

• Give it one of your own samples to critique. It should score 8 or higher. If not, the profile is wrong.

In my version

• Three Claude skills: a voice profile, a drafter and a critic. My rules: lead with how the buyer feels, then the business case, and no em dashes.

Ask first

• Which use cases does your practice deliver? (default your top 5)
• Is the person new to the use case? (default no; if yes, add a glossary)

Needs

• A short write-up of how you've built each use case before.
• Transcripts, plus read access to the client's system if you have it.
• If your assistant can only chat: paste your notes and the last call.

Build

1. Load the use-case notes.
2. Search the client's calls for that topic.
3. Read the client's live model structure.
4. Decide what differs from usual practice, and write 3 to 5 questions.
5. Write the brief.

Check

• Every client claim must cite a call or a model element.

In my version

• Claude skills built on use-case knowledge from our past Pigment builds, my transcript search, and the Pigment connector.

Ask first

• What does your SOW template look like? (default paste your last SOW)
• Which complexity drivers change your price? (default entities, currencies, integrations, users)

Needs

• Call transcripts.
• A SOW template.
• If your assistant can only chat: paste the transcript, then the template.

Build

1. Extract facts into a fixed scope profile.
2. Score complexity on each driver.
3. Write a questionnaire for the gaps.
4. Draft the SOW in your template's language.
5. Stop for your approval, then build the document.

Check

• Every scope line must cite the transcript, and gaps must appear in the questionnaire, not the SOW.

In my version

• Four Claude skills chained together, plus industry revenue notes.

Ask first

• What should it look for? (default allocations, spreading, rollforwards, cohorts, capacity)

Needs

• The files.
• If your assistant can only chat: upload the workbook.

Build

1. List every tab and its formulas.
2. Match patterns to the list of structures.
3. Flag hardcodes and circular references.
4. Write the readout.

Check

• Give it a workbook you built yourself and see whether it names your structures.

In my version

• Two Claude skills: one for Excel models and one for exports from our old planning tool.

Ask first

• What target utilization? (default 75%)
• Which window? (default last 4 weeks)

Needs

• A connection to your time-tracking tool.
• If your assistant can only chat: paste an export.

Build

1. Pull entries and allocations.
2. Compute billable divided by capacity.
3. Flag anyone 10 points below target.

Check

• Compare one person's figure to the tool's own report.

In my version

• Ruddr connector with four small Claude skills.

Ask first

• Which course, date and time window? (default Saturday, 7 to 9am)
• Who is in the group? (default your last group)
• Check how often? (default every 15 minutes)
• Book automatically or ask first? (default book)

Needs

• A browser the assistant can drive, already signed in to the club's booking site.
• A scheduler.
• If your assistant can only chat: this one needs a browser.

Build

1. Open the tee sheet for the date.
2. Read the open slots.
3. Decide: the first slot in the window with room for the whole group.
4. Book it and confirm the booking took.
5. Message you and stop the schedule.

Check

• Run it with booking off for a date with known open slots. It should name one slot and stop.

In my version

• A Claude skill driving Chrome on my club's tee-sheet site. It remembers my last course, window and group.