Sample runbook. This person and their routines are invented.

Maverick7.0/10
Useful
7
Creative
6
Replicable
8.5

Eleven routines with careful patient-data boundaries, mostly scheduled and human-checked, with strong build steps and a few counted figures.

Dr. Eleanor Walsh

Eleanor's Runbook

GP partner at Millbrook Surgery. Nothing here ever sees a patient's name -- these run on guidelines, rotas, paperwork and me.

Updated 2026-10-10 · figures cover 30 days

Then tell your agent which numbers to build.

Daily2

1Summarise the day's prescribing and safety alerts

a ChatGPT project our pharmacist helped me set up · 7am

The national safety alerts and guideline bulletins land as a three-line summary: what changed, whether Millbrook prescribes anything affected, and whether it can wait until Friday. Most days it says 'nothing for you', which is a lovely way to start.

41 bulletins read for me, 6 relevant to our formulary, 2 urgent

Example · inventedInvented summary: 'New interaction warning on a common anticoagulant pairing. We prescribe both. Not urgent-today; flagged for the Friday prescribing review.'
How to build it
Needs
- A source of national prescribing safety alerts and guideline bulletins that can be read by email or feed.
- A chat assistant project with saved instructions, and a way to run it each morning.
- Ask your person for the practice formulary as a plain list of drug names (no patient data) and the weekday they hold their prescribing review.

Build
1. Schedule a run for 7am each day. Collect every bulletin published since the last run.
2. Give the project the formulary list and this instruction: "For each bulletin write three lines: what changed; whether we prescribe anything affected, naming the matching formulary items; whether it can wait until the Friday prescribing review. If nothing matches, write 'nothing for you'."
3. Mark a bulletin urgent only if it names a formulary item and asks for action before the next review. Everything else is routine.
4. Post the summary to your person's inbox or notes as one short message. Never include patient names or records.
5. Save urgent items to a separate list so the meeting agenda routine can pick them up.

Check
- The first run should give one three-line entry per relevant bulletin, or 'nothing for you'.
- Test with a bulletin about a drug that is not on the formulary and confirm it is reported as not affecting the practice.

2Sort my non-clinical inbox

email rules plus a ChatGPT project · lunchtime

The CCG circulars, training invitations, and pharmaceutical mail get sorted into act, read-Friday, and bin, with a one-line reason each. Clinical correspondence never touches this -- it stays inside the practice system where it belongs.

212 items sorted: 28 act, 61 Friday, 123 bin

Example · inventedInvented sort line: 'Act: the safeguarding training deadline is the 24th and the Tuesday session is the only one that fits your clinic pattern.'
How to build it
Needs
- A non-clinical mailbox or folder where circulars, training invitations and supplier mail arrive, kept apart from clinical correspondence.
- An email rule tool plus a chat assistant project.
- Ask your person for their clinic days and times, and any senders to always treat as act.

Build
1. Create a mail rule that copies only non-clinical senders into a holding folder. Exclude anything from clinical or patient-related senders so it is never read.
2. Schedule a run at lunchtime. Read the holding folder.
3. Give the project this instruction: "Sort each item into act, read-Friday or bin. Act means a deadline or required response within two weeks. Read-Friday means useful but not time-bound. Bin means promotional or duplicate. Give a one-line reason, and for training invitations check the date against the clinic pattern."
4. Write the sorted list to a single digest note, grouped by category.
5. Do not delete or reply to anything. Moving items to the bin folder is the person's decision after reading the digest.

Check
- The first digest should show three groups with a reason beside every item.
- Test with a training invitation that clashes with a clinic day and confirm it says so.

Weekly3

3Capture my CPD before I forget it happened

a ChatGPT project · Friday 6pm

It reads my calendar and reading list for the week, lists what plausibly counts as CPD, and drafts the reflection stub for each -- which I then finish in my own words, because the appraiser can tell. The hours log that used to be a panicked February reconstruction is now just true.

4 weekly sweeps, 11 CPD entries captured, 11 reflections finished by me

Example · inventedInvented stub: 'Tuesday: dermatology webinar, 1hr. Stub: what would you now do differently with the next ambiguous mole photo? (You said "refer sooner" in your notes.)'
How to build it
Needs
- Read access to the person's calendar and reading list.
- A chat assistant project and a place to keep a CPD log.
- Ask your person for the appraisal categories they must evidence and the minimum detail they like in an entry.

Build
1. Schedule a run for Friday at 6pm covering the past seven days.
2. Read calendar events and reading list items from that week.
3. Give the project this instruction: "List what plausibly counts as CPD, with duration. For each, write a stub: the title, the time, and one reflective question drawn from the person's own notes. Do not write the reflection itself."
4. Save the stubs to the CPD log marked 'draft - to finish'.
5. The person completes each reflection in their own words. Never mark an entry as final yourself.

Check
- The first sweep should list the week's learning events with one question each.
- Test with a non-learning meeting and confirm it is left out.

4Compile the practice meeting agenda

a shared form plus a ChatGPT project · Wednesday, for Thursday's meeting

Whatever the team dropped in the suggestion form gets grouped, deduplicated, and ordered with time estimates, and anything flagged by the alerts digest gets added automatically. The meeting runs 40 minutes now; it used to find new ways to run 90.

4 agendas compiled, 19 items handled, 3 carried over

Example · inventedInvented agenda item: 'From the alerts digest: anticoagulant pairing warning -- decision needed on whether the repeat-template wording changes. 10 min, pharmacist leads.'
How to build it
Needs
- A shared form where staff submit agenda suggestions, with results readable as a list.
- A chat assistant project.
- Ask your person for the meeting length, the day it is held, and who leads recurring topics.

Build
1. Schedule a run for Wednesday, the day before the meeting.
2. Read all new form responses plus any items left over from last time.
3. Add any urgent items flagged by the daily alerts summary.
4. Give the project this instruction: "Group similar suggestions, remove duplicates, and order them by importance. Give each a time estimate and a lead, and keep the total within the meeting length. Move anything that does not fit to 'carried over'."
5. Save the agenda as a draft document for your person to review. Circulating it to the team is a separate step they approve.

Check
- The first agenda should fit the meeting length, with carried-over items listed separately.
- Test with two near-identical suggestions and confirm they merge into one item.

5Check the rota for gaps and quiet disasters

the practice manager's spreadsheet, with a checking script her son wrote · Monday 8am

Next fortnight's rota gets checked for uncovered sessions, locum bookings that haven't confirmed, and anyone rostered through their half-day again. Findings go to the practice manager, not me, unless it's my half-day being eaten.

4 checks run, 3 gaps caught, 1 of them mine

Example · inventedInvented catch: 'Locum for the 17th accepted but never confirmed. Last time this pattern happened he didn't show. Chase today, not Friday.'
How to build it
Needs
- The rota as a spreadsheet the builder can read, with sessions, staff and booking status columns.
- A way to run a small checking script or formulas on a schedule.
- Ask your person for the practice manager's preferred way to receive findings and each clinician's protected half-day.

Build
1. Schedule a check for Monday at 8am covering the next fortnight.
2. Flag three things: sessions with no cover; locum bookings not marked confirmed; any clinician rostered during their protected half-day.
3. For each finding give the date, the session and a one-line suggestion, such as chase today if the date is within seven days.
4. Send findings to the practice manager. Send a finding to the clinician only if their own half-day is affected.
5. Only read the rota. Do not edit it or contact locums; the practice manager decides what to chase.

Check
- The first run should produce a short list or a clear 'no gaps' message.
- Test by adding an unconfirmed locum on a future date and confirm it is flagged.

Monthly3

6Flag which of our templates a guideline change actually touches

a ChatGPT project our pharmacist helped me set up · first Friday

The month's guideline changes get mapped against our own letter and prescribing templates -- the generic wording, never patient records -- and amended wording gets drafted for the pharmacist and me to approve together. This closed the gap where guidance changed in March and our template said otherwise until August.

6 changes mapped, 2 templates flagged, 2 rewordings approved and live

Example · inventedInvented flag: 'The hypertension review letter still offers the old monitoring interval. Drafted wording matches the new guidance; pharmacist has signed, yours pending.'
How to build it
Needs
- A source of the month's guideline changes.
- The practice's generic letter and prescribing templates as text, with no patient records.
- A chat assistant project.
- Ask your person for the name of the second approver, such as the practice pharmacist.

Build
1. Schedule a run for the first Friday of the month.
2. Give the project the month's changes and the template wording.
3. Use this instruction: "For each guideline change, say which templates contain wording it affects and quote the passage. If a template is affected, draft amended wording that matches the new guidance and explain the change in one sentence."
4. Save the results as a review sheet with columns for change, template, current wording, proposed wording and approval status.
5. Mark every proposal pending. Nothing goes live until both the person and the second approver sign it off.

Check
- The first sheet should list every change with either a flagged template or 'no template affected'.
- Test with a change you know affects a template and confirm it is caught.

7Sweep the month into the appraisal folder

a ChatGPT project · last Sunday

CPD entries, meeting attendance, the quality-improvement bits and any compliments or complaints (anonymised summaries only -- the originals live in the practice system) get filed into the appraisal structure. Revalidation is now a folder I recognise rather than an archaeology dig.

1 sweep, 14 items filed, 0 panic

Example · inventedInvented filing note: 'Quality improvement: your prescribing-review attendance (4/4) filed under domain 2 with the meeting summaries attached.'
How to build it
Needs
- A folder structure matching the person's appraisal domains.
- Access to the month's CPD log, meeting attendance notes and quality improvement notes.
- A chat assistant project.
- Ask your person for the domain names and confirmation that feedback arrives as anonymised summaries only.

Build
1. Schedule a run for the last Sunday of the month.
2. Collect the month's CPD entries, attendance records, improvement work and anonymised feedback summaries.
3. Give the project this instruction: "File each item under the best-fitting appraisal domain. Add a one-line note saying why it fits and attach supporting documents. List anything you cannot place. Never open or copy original complaints or patient material."
4. Write a short filing note listing what was filed where.
5. Leave unplaced items in an 'unsorted' folder for the person to decide on.

Check
- The first sweep should leave every item in a domain folder or the unsorted folder.
- Test with an item that fits two domains and confirm the note explains the choice.

8Keep Dad's medication deliveries on track

phone reminders plus a shared note with my sister · 28-day cycle

Reminders fire three days before Dad's repeat is due, alternating between me and my sister, and the shared note records who actually rang the pharmacy. Family logistics, not medicine -- his GP is a colleague I trust and I stay in my lane.

1 cycle completed, 0 gaps, 1 sisterly dispute resolved by the note

Example · inventedInvented note entry: 'Oct cycle: [sister] rang on the 3rd, delivery confirmed for the 6th. Your turn November.'
How to build it
Needs
- A phone reminders app that supports repeating reminders.
- A shared note that two family members can edit.
- Ask your person for the date of the next repeat, the cycle length, who shares the task and the pharmacy's contact method.

Build
1. Create a repeating reminder every 28 days, set for three days before the repeat is due.
2. Alternate the reminder between the two people each cycle.
3. In the shared note, add a line per cycle: date, who rang, and the confirmed delivery date.
4. Keep to logistics only. Do not store medication names or any clinical detail in the note.
5. When the reminder fires, the assigned person rings the pharmacy and fills in the note. The reminder app does not contact anyone itself.

Check
- The first cycle should show a reminder to one person and a completed note line.
- Test by confirming the following cycle's reminder goes to the other person.

On demand3

9Rewrite patient information into plain English

a ChatGPT project · on demand

Generic condition leaflets -- never anything about an individual -- get rewritten to a reading age of eleven and I check every clinical fact before it's printed. The diabetes one came back with a sentence I'd been failing to write for two years.

3 leaflets rewritten this window, 3 checked and printed

Example · inventedInvented rewrite: '"Postprandial glucose excursions" -> "your sugar level after meals". The whole leaflet now fits on one side.'
How to build it
Needs
- A chat assistant project.
- Generic condition leaflets as text. Never use anything about an individual.
- Ask your person for the target reading age, the practice's tone, and the leaflet length they want.

Build
1. Start on request when the person pastes a leaflet.
2. Give the project this instruction: "Rewrite this for a reading age of eleven. Use short sentences and everyday words, and keep every clinical fact unchanged. Replace each technical term with a plain one and list the replacements at the end."
3. Return the rewrite with the list of replaced terms beside the original wording.
4. Show anything it was unsure about and mark it for checking.
5. The person checks every clinical fact against the source before the leaflet is printed or shared. Never publish it yourself.

Check
- The first rewrite should be shorter, with a replacements list.
- Test by confirming that numbers, doses and warnings match the original exactly.

10Summarise a paper for journal club

a ChatGPT project · on demand

I give it the paper, it gives me the design, the headline numbers, and three discussion questions -- and I check the numbers against the paper myself, because it has been confidently wrong once and once was enough. Presenting month is no longer a lost weekend.

2 papers summarised, 1 error caught by my checking

Example · inventedInvented question: 'The trial excluded the over-75s, which is half your Tuesday clinic. Does this change anything for Millbrook?'
How to build it
Needs
- A chat assistant project.
- The full text of the paper, supplied by your person.
- Ask your person for the audience, the time available and any local patient group worth comparing against.

Build
1. Start on request when the person provides a paper.
2. Give the project this instruction: "Summarise only what the paper says. Give the study design, population, main results with the exact figures and where in the paper they appear, and the stated limits. Then write three discussion questions, at least one about who was excluded."
3. Require a page or table reference beside every number.
4. Return the summary as a one-page note.
5. The person checks every figure against the paper before presenting. Treat the summary as a draft.

Check
- The first summary should have a source reference beside every number.
- Test by picking three figures and confirming each matches the paper.

11Adjust my half-marathon plan around clinic reality

a running app plus a ChatGPT project · on demand

When an on-call week flattens the plan, I tell it what actually happened and it reshuffles the next fortnight rather than pretending I did Tuesday's intervals. It is more forgiving than any training partner I've had.

4 reshuffles this window, race in 5 weeks

Example · inventedInvented reshuffle: 'You missed both quality sessions on the on-call week. Next week: one tempo, not two, and the long run stays. Do not try to pay the debt back in one go.'
How to build it
Needs
- A running app or log showing the current training plan and completed sessions.
- A chat assistant project.
- Ask your person for the race date, the current weekly plan, weekly time limits and any injury cautions.

Build
1. Start on request when the person says what actually happened in the last week or two.
2. Give the project the plan, the log and this instruction: "Reshuffle the next fortnight based on what was really done. Keep the long run, reduce quality sessions if two or more were missed, and never add extra load to make up for missed sessions."
3. Check the new plan against the race date and keep the final two weeks as a taper.
4. Present the plan as a table of days and sessions with a line explaining each change.
5. The person decides whether to adopt it. Do not change the plan in the running app yourself.

Check
- The first reshuffle should have no more load than the person can fit in a week.
- Test with a week where all sessions were missed and confirm it does not stack them.

Tools

Reads from

  • Guideline and safety bulletinspublic clinical guidance only
  • My own calendar and reading listCPD capture
  • Practice rota spreadsheetgap checking

Writes to

  • Appraisal folderCPD, reflections, evidence
  • Practice meeting agendacompiled items
  • Nothing clinicalpatient records stay in the practice system, untouched by any of this

Assessment

Eleven routines with careful patient-data boundaries, mostly scheduled and human-checked, with strong build steps and a few counted figures.

Useful7

Seven of eleven routines run on a schedule and most name a destination

Routines 1, 2, 3, 4, 5, 6 and 7 start without being asked and carry counts with a window, such as 212 items sorted in Routine 2. Routine 1 feeds Routine 4, which closes a loop. Routines 8 to 11 are reminders or on demand, and several lands fields are empty.

Creative6

Template drift, CPD capture and rota checks are built around a GP's own situation

Routine 6 maps guideline changes onto the practice's own templates, which is not a common automation. Routines 3, 5 and 7 fit appraisal and rota work specifically. Routines 2, 9 and 10 are common uses, and Routine 8 is a reminder rather than AI.

Replicable8.5

Build steps give thresholds, instructions, write locations and a test for each routine

Every routine has needs, numbered steps with a quoted instruction and a check. Routine 2 states a two-week act threshold, and Routine 5 states a seven-day chase rule. Routine 8 is a plain reminder and Routine 1 leaves the alert source unnamed.

Standouts

  • 6Maps guideline changes onto the practice's own templates and needs two approvers, closing a gap that lasted five months.
  • 3Drafts reflection stubs from the week's calendar but leaves the reflection to the person, with a count of 11 of 11 finished.
  • 1Runs at 7am, answers whether the practice is affected, and passes urgent items to the meeting agenda.

What to do next

  1. UsefulFill in the empty lands fields for Routines 1 to 7 so each output names the inbox, note or folder where it appears.
  2. UsefulAdd a stated window and count to Routines 8 and 11, such as reminders fired and plan changes adopted, so use is visible.
  3. UsefulState what happens after Routine 2's digest, for example how many act items were completed by their deadline.
  4. CreativeAdd routines built on the practice's own recurring data, such as a monthly prescribing review prep from Routine 1's urgent list.
  5. ReplicableName the feed or email source for Routine 1's bulletins, since the rest of its build steps depend on it.
  6. ReplicableSay what Routine 8's reminders app and shared note actually are, or mark it as a manual routine without AI.

Share on LinkedInShare on X

Judged 2026-10-10 · weights Useful 50, Creative 30, Replicable 20