Learning

Building a weekly Civil Service Leadership Coaching workflow with AI

Robin Riley

Robin Riley

— 3 min read
Share

Earlier this year I wrote about using NotebookLM as a personal leadership coach, drawing on my own learning blog posts as the source material. That proved useful, but a it was ad-hoc, not a habit.

Over the past few months I've built something more deliberate: a weekly leadership coaching workflow, using Claude, that compounds week on week.

How it works

The workflow has a simple weekly cadence:

  • Thursday afternoon Capture (~10 min, work device). I use Microsoft Copilot to summarise my week from my calendar and inbox, paraphrased and suitably anonymised - using a weekly prompt that Claude helped me develop. I add personal reflections and three personal priorities for next week, and email it to my non-work self.
  • Friday morning Coaching (~40 min, personal device). I start a new chat in a Claude Project called "Leadership Coach", paste in my Thursday summary, and run a structured session, consisting (1) an initial reflection report from the coach, (2) a free-form Q&A conversation in which Claude challenges and responds, then (3) a wrap-up that captures commitments for the coming week - things I am going to do, or do differently - and an update to a "coaching log" which evolves from week to week. The coaching log is vital,
  • Panel review (~10 min, personal device, straight after Coaching). A separate Claude chat convenes a four-person expert panel to review the coaching session. The four are personas based on real people with very different leadership expertise, but all relevant to me and my aspirations. They are also personal to me, hence I am not sharing them here. One of the panel acts as chair.
  • Monthly iteration (last Friday of the month). The normal Friday session is replaced with a structured review of the workflow itself, what's working and what should change.

The coaching log is vital, as it records recurring themes, working hypotheses about my leadership, and things to watch out for in my own behavour. Without it, every Friday would start from scratch.

Setting it up

Creating this workflow involved a lot of iteration between Claude and I. It also involves a set of around a dozen source documents which themselves took some time to develop. These include my previous behaviour examples (from past Job Applications), the public SCS leadership behaviours, my current role and objectives, the 4x panel personas and more.

Part of the design challenge was handling the gap between my work and personal IT. Nothing OFFICIAL-SENSITIVE or above can leave the work device, so the Thursday summary is my own paraphrase, and any external parties (suppliers, partner organisations, individuals) are anonymised generically. The Copilot prompt enforces this but I verify before emailing to myself.

Early reflections on the coaching

It can be intense, at times exhausting. The AI coach has explicit permission to challenge, producing sessions that genuinely make me think. It is not a sycophant; it pushes back, names patterns I'm avoiding, and refers back to commitments I made the previous week. I have more than once found myself calling an end to a session before reaching the full 40 minutes, not because the session was failing but because it was working too well. I'd had enough!

The advice has translated into real impact. Specific suggestions from the coach — about how I show up in team meetings, when to delegate rather than take work on, how to approach senior conversations — have landed in my actual work.

The panel is useful but mainly as an addition to the coaching. The panel surfaces things the main coach misses, particularly around institutional politics and the wider AI landscape.

Sharing this recipe

I have published the whole workflow as a public GitHub repository at https://github.com/rr-coding/ai-leadership-coach. This is pitched primarily at UK civil servants but should be adaptable for any senior leader who wants an AI coaching workflow with some structure and bite.

Note that these files have been anonymised to remove the details personal to me; you'll need to personalise them to you. The README explains the design; start-here.md walks you through setup using Claude.ai. If you are not technically confident setting this up (I wasn't) then simply paste the text of thw start-here.md file into Claude.ai to get going. Setup takes about 90 minutes (or longer, depending on how much effort you need/want to spend on the underpinning documents) but you can split this time across multiple sittings.

I'll keep running and refining the workflow, and write a follow-up post in a few months once I have more lived experience. For now, I'm genuinely surprised by how valuable it has been.

Read more