Shantanu ShekharWebsite

Ask first

• Which inboxes should I watch? (default: every connected inbox)
• What times should I sweep? (default 9am and 6pm)
• What counts as urgent? (default: a deadline within 3 days, money or a signature asked for, a named person waiting on you)
• Send a note when nothing is new? (default yes, one line)

Needs

• Read access to your email.
• A place to keep a running list of open items between sweeps (a note or a file).
• If your assistant can only chat: paste the day's senders and subject lines each morning and evening, plus the last open list.

Build

1. Read mail received since the last sweep in every inbox named.
2. Read the open-items list from the last sweep.
3. For each open item, look for your reply or a later message that settles it. If found, move it to Resolved.
4. Mark new mail Needs you if it has a deadline within 3 days, asks for money or a signature, or a named person is waiting on you. Skip newsletters, receipts and notifications.
5. Keep an item in Still open with no new mail if its deadline is within 7 days.
6. Write the note under three headings: Needs you, Still open, Resolved. One line per item, 10 lines at most.
7. Deliver the note in chat. Read-only: never send, file, pay or submit anything.
8. Save the updated open list.

Check

• Run it once with delivery off over the last 48 hours. Every Needs you line must point to a real message, and an item you already answered must show as Resolved.

In my version

• Instinct, a scheduled assistant, reading three inboxes: my day job, my firm and personal.
• Sweeps at 9am and 6pm, read-only, with a short note on quiet days.

Ask first

• How far back should I look? (default: posts from the last 14 days)
• Include DMs? (default yes, skipping obvious sales pitches)
• Where is your voice guide? (default: I write one from 10 of your past posts)
• Reply length? (default 1 to 3 sentences)

Needs

• A way to read your notifications, comments and messages (a browser agent signed in as you).
• A short voice guide: phrases you use and phrases you never use.
• If your assistant can only chat: paste the new comments and DMs each morning and ask for drafts.

Build

1. Open your recent posts and messages and collect anything new since the last run.
2. Sort into Worth replying (questions, disagreements, people you know, real opportunities), Like only (congratulations, emojis) and Skipped (pitches, spam).
3. For each Worth replying item, draft a reply that answers the actual point in under 3 sentences.
4. Check every draft against the voice guide's banned phrases and rewrite any match.
5. Deliver the digest in chat.
6. Stop there. Never post, comment, react or send a DM. You copy and send the ones you want.

Check

• Run it today with posting disabled. Every draft must answer something specific in its comment, and none may contain a banned phrase.

In my version

• Claude with Chrome, running a custom skill on weekdays.
• Shares its voice guide with my post-drafting routine.

Ask first

• Who gets it, and at what time? (default: one person, 7am)
• What topic? (default: ask them for one sentence)
• How many stories? (default 3 to 5)
• Sources to prefer or avoid? (default: major outlets, no paywalls)

Needs

• Web search.
• A way to send email on a schedule, or a draft you forward yourself.
• If your assistant can only chat: ask for the digest each morning and paste it into an email.

Build

1. Search news from the last 24 hours on the topic.
2. Drop duplicates, opinion pieces and anything older than 48 hours.
3. Rank by how directly each story hits the topic and keep the top 3 to 5.
4. Write one line on what happened and one on why it matters, each with a link.
5. Send the email. On a quiet day, send two lines saying so instead of padding.

Check

• Run it now with sending off. Every story must have a working link dated within 48 hours.

In my version

• A scheduled Claude task, one recipient, one topic she chose.

Ask first

• Where should the brain live? (default: a folder in your home directory, never a shared repository)
• Which projects are live right now? (default: your top 5)
• Where do meeting notes come from? (default: you drop transcripts and decks into a raw folder)

Needs

• A folder of plain text files: a constitution file describing your role, plus projects, frameworks, raw, processed and daily-logs folders.
• A coding agent that reads a folder and runs saved commands.
• If your assistant can only chat: keep one long note per project, paste in the day's notes and ask it to update the right notes and write today's log.

Build

1. Write the constitution file: your role, priorities and how you like answers written.
2. Make an ingest command: read each new file in raw, decide which project it belongs to, append the facts with source file name and date, then move the file to processed.
3. Never overwrite a project note. Append, and mark facts that changed.
4. Make a shutdown command: take one line from you about the day, list what moved per project and any decisions, and write open items under Rolled to tomorrow in today's log.
5. Make a today command that reads yesterday's log first.
6. Keep everything local. Confidential files are never uploaded anywhere.

Check

• Drop one old transcript in raw and run ingest. Every new line in a project note must name its source file.

In my version

• Claude Code on my laptop, a markdown vault, saved commands for ingest, shutdown, brief and today.
• Run by me each day. Next step is connecting Salesforce, Gong and an internal data agent.

Ask first

• What is your field, in one line? (default: ask)
• Which sources? (default: LinkedIn, two subreddits, X and three newsletters you name)
• Scoring weights? (default: timeliness 40%, fit with what you know 40%, not posted before 20%)
• How many ideas? (default 5)

Needs

• Web search or a browser agent.
• A list of your recent post topics so it does not repeat you.
• If your assistant can only chat: paste five links you saw this week and ask for scored angles.

Build

1. Search the last 7 days of discussion in your sources.
2. Group posts by theme and keep themes with at least 3 independent mentions.
3. Drop themes you posted on in the last 30 days.
4. Score each theme out of 10 on the weights and keep the top 5.
5. For each, write a hook line, why now with links, and the personal angle you could add.
6. Save to the ideas file. Never post anything.

Check

• Run it now. Every idea must link at least 3 sources from the last 7 days.

In my version

• A Claude skill on a Tuesday schedule, feeding my post-drafting routine.

Ask first

• Which network, and how far back? (default: LinkedIn, last 7 days)
• Which numbers? (default: impressions, reactions, comments, reposts)
• Where should the log live? (default: a spreadsheet, one row per post)

Needs

• Read access to your post analytics (a browser agent signed in as you, or the platform's export).
• A spreadsheet.
• If your assistant can only chat: paste a screenshot of each post's analytics and ask for the row and the note.

Build

1. Open each post from the last 7 days and read the four numbers.
2. Append one row per post with date, topic, first line and numbers. Never edit old rows.
3. Compare each number with your 4-week median.
4. Mark a post Landed above 1.5x the median and Flopped below 0.5x, and name the visible difference (first line, format, timing).
5. Post a short note in chat. Read-only on the platform.

Check

• Run it today. Each row must match the numbers shown on the post itself.

In my version

• Claude with Chrome on a Friday schedule, writing to a sheet my post-drafting routine reads.

Ask first

• Who is it for? (default: just you)
• Which test results do you have? (default: upload the latest lab reports)
• Which day for the check-in? (default: Sunday evening)
• Which health app numbers? (default: steps, sleep, resting heart rate)

Needs

• Lab results as PDFs or photos.
• An export from a health app, or the weekly numbers typed in.
• A place to keep one tracker page per person.
• If your assistant can only chat: keep one chat per person, paste the results once, then answer three questions each week.

Build

1. Read each test result and list every marker outside or near the edge of its reference range, with the date.
2. Record what the doctor said to do for each. Never invent medical advice; mark unknowns as questions for the doctor.
3. Each week, ask three questions: what did you do, any symptoms, any appointments.
4. Read the week's health app numbers and compare them with the 4-week average.
5. Update the tracker under Watch, Done this week and Next with the doctor.
6. Keep each next step open until you mark it done or a doctor changes it.

Check

• Load one past result. Every Watch line must cite the marker, value and date from the document.

In my version

• Claude with a visual tracker per family member, built from diagnostic tests, with Apple Health exports.

Ask first

• Which roles? (default: Director and above in RevOps, GTM operations and GTM engineering)
• Where? (default: Europe, plus companies from elsewhere opening there)
• Company size? (default: 100 to 2,000 people, B2B software)
• Which signals count? (default: funding round, new regional office, first regional hire, individual contributors hired with no manager)
• Who gets the outreach? (default: hiring manager, then CRO, then founder)

Needs

• Web search across job boards and news.
• A tracker with company, role, signal, date, contact and status.
• If your assistant can only chat: ask for the sweep three times a week and paste the results into your own sheet.

Build

1. Search roles posted since the last sweep that match title, region and size.
2. Search signals since the last sweep.
3. Drop anything already in the tracker and companies marked do-not-contact.
4. For each new row, find the likely hiring manager from public pages.
5. Draft a 4 to 6 line message that names the specific role or signal. No generic openers.
6. Add rows to the tracker with status New. Never send anything; you review and send.
7. Rebuild the board from the tracker.

Check

• Run it with sending off. Every row must have a link dated after the last sweep, with no duplicate companies.

In my version

• Two Claude skills, one for RevOps and GTM leadership and one for GTM engineering, scheduled Monday, Wednesday and Friday, each with its own tracker and board.

Ask first

• Which country's rules? (default: ask)
• Which returns? (default: personal, plus company if you own one)
• Do you use an accountant? (default: yes, for review)

Needs

• Bank and card statements, invoices, receipts and last year's return.
• The tax authority's published guidance.
• If your assistant can only chat: paste the year's figures by category.

Build

1. Gather every document and list what is missing.
2. Categorize each transaction. Flag anything unclear instead of guessing.
3. Reconcile income to bank deposits.
4. Fill each return line with the figure and its source document.
5. List questions for the accountant and every deadline.
6. You submit. Nothing is filed or paid for you.

Check

• Run it on last year's documents. The totals must match the return you filed.

In my version

• Claude, for Irish personal and company returns, reviewed before I submitted.

Ask first

• Who is the audience? (default: your boss)
• Which frameworks do you use for this kind of question? (default: write each one down once as a file)
• Option labels? (default: Conservative, Balanced, Bold)

Needs

• The private work brain from my daily meetings routine, or any folder holding your frameworks and project notes.
• If your assistant can only chat: paste the framework and the relevant project note, then ask the question.

Build

1. Read the question and pick the relevant frameworks and project notes.
2. Pull every fact with its source, and flag anything older than 30 days as possibly stale.
3. Write the answer in one sentence first.
4. Break it down by the framework's segments.
5. Give three options with cost, risk and time for each.
6. Recommend one and say why.
7. List 3 to 5 questions the audience is likely to ask, with short answers.
8. Save to drafts. Nothing is sent.

Check

• Ask a question you answered last month. Every number in the brief must trace to a named source file.

In my version

• Claude Code with a saved brief command, reading my markdown vault.
• Next: live numbers from Salesforce and an internal data agent.

Ask first

• Paste 10 posts you are proud of. (default: I pull your last 10)
• Phrases or patterns you never want? (default: em dashes, "it's not X, it's Y", lists of three for rhythm, closing questions that bait comments)
• Length? (default: 120 to 220 words)
• British or American spelling? (default: yours)

Needs

• A voice guide: sentence length, typical openers, banned patterns.
• Your idea in any form, including a voice note transcript.
• If your assistant can only chat: paste the voice guide at the top of each request.

Build

1. Read the idea and find the single concrete moment or number in it.
2. Open on that moment, with no warm-up line.
3. Write the body in short paragraphs, one idea each.
4. Run the banned list line by line and rewrite any match.
5. End on the point itself, with no question to the reader.
6. Save the post as a file and show it. Never post.

Check

• Draft from an old idea, then search the draft for each banned pattern. There must be zero matches.

In my version

• A Claude skill with a voice.md file, shared by all my writing routines.

Ask first

• What are you writing, and for whom? (default: ask)
• Past examples to match? (default: your last 3 of this type)
• Calendar dates or relative timing in schedules? (default: relative, such as week one, weeks two and three)

Needs

• Your voice guide and banned list.
• Notes or a call transcript for the piece.
• If your assistant can only chat: paste the voice guide and the notes together.

Build

1. Read the notes and pull out the client's problem in their own words.
2. Open with the answer or the number. No preamble.
3. Use headings that state a conclusion.
4. Write schedules in relative weeks so dates are agreed after signing.
5. Run the banned list and rewrite any match.
6. Hand back the draft. Nothing goes to a client without you.

Check

• Redo a proposal you already sent. It must contain no banned patterns and no calendar dates in the timeline.

In my version

• A Claude skill built on the voice guide from my post-drafting routine.

Ask first

• File type? (default: from the request, spreadsheet, deck or document)
• Minimum font size? (default: 14pt body on slides, 11pt in documents)
• Colors for inputs and formulas? (default: blue inputs, black formulas)

Needs

• Your data or notes.
• A short house style file.
• If your assistant can only chat: ask for the slide text and table layouts, then build the file yourself.

Build

1. Write the storyline first: one full-sentence headline per slide or section that states the conclusion.
2. Read the headlines alone. If they do not tell the story, fix them before building.
3. Models: one tab per job (inputs, calculations, outputs), a header block with source and version, no typed numbers inside formulas.
4. Decks: stat tiles for key numbers, one counter-case slide, a Note and Source line under every chart.
5. Memos: a short summary or the headline number first, a judgment column in tables, a list of data gaps, and a closing list of actions.
6. Check font sizes against the minimum. Add pages before shrinking type.

Check

• Build from sample data. Every chart must have a source line and no text may be below the minimum size.

In my version

• Three Claude skills (models, decks, documents) distilled from my years in strategy consulting.

Ask first

• The role in three lines: title, must-haves, nice-to-haves. (default: paste the brief)
• Target companies to source from? (default: 20 you name)
• Exclusions? (default: current clients and people already contacted)
• How many per run? (default 10)

Needs

• The role brief.
• Web search or a professional network search.
• A tracker per search.
• If your assistant can only chat: paste profiles you found and ask for scores and drafts.

Build

1. Search the target companies for matching titles.
2. Score each person on the must-haves and drop anyone missing one.
3. Dedupe against this tracker and any other list for the same role.
4. Write one line on fit and one on the main gap.
5. Draft a short message naming what makes the role worth a look.
6. Add to the tracker as New. You send every message.

Check

• Run with sending off. Every candidate must meet every must-have, and none may already appear in a tracker for the role.

In my version

• A Claude skill per search; one role has two lists deduped against each other.

Ask first

• Where is the tracker? (default: one spreadsheet, one row per opportunity)
• Which columns? (default: company, service, stage, value band, next step, date)
• Which views should be rebuilt? (default: a board by stage)

Needs

• A spreadsheet the assistant can write to.
• A way to publish a simple page or dashboard.
• If your assistant can only chat: paste the sheet and ask for the updated rows and a fresh board.

Build

1. Read the new opportunity details.
2. Look in the sheet for the same company and service. Update it if found, add a row if not.
3. Write the rows. The sheet is the source of truth; never edit a view by hand.
4. Rebuild every view from the sheet.
5. Compare counts between the sheet and the views and report any mismatch.

Check

• Add a test row, rebuild, then delete it and rebuild again. The views must match the sheet both times.

In my version

• A Claude skill, a Google Sheet, and published pipeline pages rebuilt on every run.

Ask first

• Client name, address and signatory? (default: ask, and leave marked blanks)
• Role and fee terms? (default: your standard terms)
• Any changes from standard? (default: none)

Needs

• Your contract template and two signed past examples.
• If your assistant can only chat: paste the template and the details and ask for the filled text.

Build

1. Read the template and past examples to learn which parts are fixed and which vary.
2. Fill only the variable fields. Never change clause wording unless told.
3. Mark any field with missing information as a highlighted blank.
4. List every non-standard change in a note above the document.
5. Save the file. You review it and send it for signature.

Check

• Generate one for a past client and compare it with the signed version. Only the variable fields may differ.

In my version

• A Claude skill and my firm's template, saved as a Word file in my clients folder.

Ask first

• Renewal base per quarter? (default: from your billing export)
• Up to three churn drivers and three expansion drivers, with rates? (default: last 4 quarters' averages)
• How to blend in actuals? (default: weight actuals by the share of the quarter elapsed)

Needs

• Renewal data by quarter, including slipped renewals.
• Quarter-to-date actuals.
• If your assistant can only chat: paste the numbers and ask for the table.

Build

1. Load the renewal base and slipped renewals by quarter.
2. Apply each churn and expansion driver to its segment.
3. Calculate net retention in money terms for the base case.
4. Make low and high cases by moving each driver to its worst and best value of the last 4 quarters.
5. Blend actuals into the current quarter by the share elapsed.
6. Show each driver as its own line so it can be changed.

Check

• Back-test last quarter using data as of week 6. The base case must land within 2 points of the actual.

In my version

• A Claude skill producing a branded spreadsheet that opens in Google Sheets, or an interactive page.

Ask first

• Do you have a series bible? (default: build one from book 1)
• Reading age? (default: 4 to 9)
• Which facts need checking? (default: every science claim)

Needs

• A series bible: characters, voice rules, world details.
• Drafts as text.
• A trusted reference for facts.
• If your assistant can only chat: paste the bible at the top of each request.

Build

1. Read the draft against the bible page by page.
2. Flag every character, voice or world detail that conflicts, quoting the bible line it breaks.
3. Check each factual claim against the reference and give a corrected line for the reading age.
4. List new details that should be added to the bible.
5. Return the review. The authors decide what changes.

Check

• Run it on a finished book. It should flag nothing; any flag means the bible needs an update.

In my version

• A Claude skill holding the bible for a multi-book series I co-write.