AI Leadership Weekly · Issue #54 · Tuesday 21 October 2025 · 08:00 BST
Good morning.
Claude now has a way to reuse instructions rather than starting every task from scratch. Anthropic also released a smaller model, while Microsoft announced more ways to speak to Copilot and share what is on screen. This is a week of practical changes to familiar tools, with an important distinction between features available now and those still heading into testing.
IN 60 SECONDS
Claude gains reusable Skills. Anthropic introduced Skills: packages of instructions and resources that Claude can use for recurring tasks, including document work.
A smaller Claude arrives. Anthropic launched Haiku 4.5, with API launch pricing of $1 per million input tokens and $5 per million output tokens.
Copilot expands on Windows. Microsoft announced more voice and screen-aware Copilot experiences.
CEO / COO / CXO CHECKLIST
CEO: Turn an experienced colleague’s approach into an owned, versioned procedure.
COO: Compare cost per acceptable result, not just the model’s advertised price.
CXO: Approve the application, account and data access separately from the device.
TOP STORIES
1. Claude gains reusable Skills
Anthropic · 16 October 2025
What happened. On 16 October, Anthropic introduced Skills: folders containing instructions, scripts and resources that Claude can load when relevant to a task. The launch covered Claude apps, Claude Code and the API. Skills can contain executable code; Anthropic advised using trusted sources.
Why it matters. Our take: The useful asset is your organisation’s working method: the approved template, the checks and the exceptions. Treat a skill as a small operating procedure, with a named owner and a review date, rather than a clever prompt everyone copies differently.
What to do. Choose a repeatable output, such as a customer briefing. Write down the required evidence, format, quality checks and escalation point. Test the procedure on three ordinary examples and one awkward case before sharing it more widely.
2. A smaller Claude arrives
Anthropic · 15 October 2025
What happened. Anthropic released Claude Haiku 4.5 on 15 October. Standard API launch pricing was $1 per million input tokens and $5 per million output tokens. Anthropic positioned the smaller model for fast, cost-sensitive work, including coding; its performance comparisons were vendor-reported.
Why it matters. Our take: The right question is not whether every task needs the most expensive model. It is which tasks can meet your quality standard with a less costly option. A saving disappears quickly when extra checking or repeated attempts absorb it.
What to do. Take a small set of routine requests and compare two model options without telling the reviewer which produced each answer. Include an “ask a person” route for uncertain cases. Count successful outputs, checking time and total usage before deciding what to switch.
3. Copilot expands on Windows
Microsoft · 16 October 2025
What happened. On 16 October, Microsoft announced opt-in voice activation for Copilot on Windows 11 and broader Copilot Vision availability. Experimental Copilot Actions for tasks involving local files were described as coming soon to Windows Insiders through Copilot Labs, not a general release.
Why it matters. Our take: Do not treat an AI feature appearing on a work laptop as an approved business service. Ask which application and account are involved, what information it can see and which actions it may take. Consumer features and your managed workplace setup need separate checks.
What to do. Review one proposed desktop use with IT and the process owner. Start with non-sensitive example files, read-only assistance where practical and a clear stop point before anything is sent, changed or deleted. Record what is actually available in your environment.
SIGNALS FROM THE LAST MONTH
6 October · Codex becomes generally available. OpenAI moved Codex into general availability, expanding its coding-agent offer beyond the earlier preview. Source
6 October · AgentKit packages agent-building tools. OpenAI introduced AgentKit, including a beta visual workflow builder, tools for evaluating agents and ChatKit for embedding them in applications. Source
6 October · Apps arrive in ChatGPT. OpenAI introduced conversational apps and an Apps SDK preview. The initial consumer rollout excluded the UK, EEA and Switzerland. Source
30 September · Sora 2 combines generated video and audio. OpenAI introduced Sora 2, combining video generation with synchronised audio. Initial app access was limited rather than a worldwide release. Source
IN BRIEF
More dated updates from the preceding 30 days.
13 October · OpenAI plans new accelerator capacity. OpenAI and Broadcom announced a 10-gigawatt collaboration, targeting deployment from the second half of 2026 through 2029. Source
9 October · Google launches Gemini Enterprise. Google introduced a workplace platform connecting company information, applications and agents, with tools for building and deploying additional agents. Source
7 October · IBM works with Anthropic. The companies announced an enterprise development partnership, beginning with an AI-assisted coding environment in private preview. Source
7 October · Gemini tries browser control. Google previewed Gemini 2.5 Computer Use, a model for interacting with browser interfaces rather than unrestricted desktop operation. Source
THE 15-MINUTE PLAYBOOK
Turn one recurring task into a reusable method
Minutes 0–4: Describe a good result. Select one frequent output and an example that an experienced colleague would accept. Explain what makes it good, not just how it looks.
Minutes 4–8: Capture the boundaries. List the approved inputs, the checks that cannot be skipped and the exceptions that require a person. Keep unnecessary data out of the trial.
Minutes 8–12: Design a fair test. Use the same cases for the current method and the AI-assisted version. Include an incomplete or contradictory input. Measure correction effort alongside completion time.
Minutes 12–15: Assign maintenance. Name who approves the instructions, updates them when the process changes and withdraws them when they stop working. Set a review date.
The output is a one-page procedure ready to test, not a library of unowned prompts.
DATA WAVE MOMENT
Build from the work your team already understands. Capture the useful judgement, connect the right information and test the result with the people who will use it. Start with one repeatable activity that matters.
QUESTION FOR READERS
Which process depends too heavily on someone knowing the right way to ask — and who should own a better method?
Brought to you by Data Wave — your AI & Data Team as a Subscription.
