Practical AI for Executives

Analysis of an European market entry strategy

Development of profile preferences for consultants

Use of built-in styles for business writing

Development of financial briefs

Development of data center energy briefs

Managing a SaaS product launch

Building a ROI calculator for software investment

Delegating tasks within an HR hiring workflow

Summarizing earnings reports

Instructor introduces a concept or technique

Instructor demonstrates the concept in Claude Chat

You apply the concept in Claude through an exercise or lab

Note: slides and student guide do not include screen shots

Not just a chat window

Structured environment for producing professional assets

Part of a family of apps and extensions that surface* Claude models

Built for how knowledge workers work:

Central threaded interface — user prompts and Claude responds sequentially

Each exchange is a "turn":

Claude’s responses vary in format depending on the task:

Conversations saved to chat history automatically — except in Incognito mode

"Context" is the information Claude can "keep in mind" at once to generate a response.

Context window size varies by model:

In very long conversations:

Dynamic space for Claude to surface details, visuals, files, and tools

Appears when Claude determines it would enhance the response

Shows Project and Artifact details when using those features (covered in later modules)

Ask Claude to use the right panel for specific outputs:

Memory: role, professional context, communication preferences, technical style, ongoing project themes

Chat history: verbatim conversation content — decisions, artifacts, specific facts discussed

Neither system stores: incognito chats — fully excluded from both memory and chat history

Memory does not store: conversations held while memory is paused

Chat history search and memory:

Independent controls for two independent systems:

View and edit your memory profile: Settings > Capabilities > View and edit memory

Pause memory: preserves existing memory but stops all new memory creation

Reset memory: permanently deletes all memory including project memories — cannot be undone

Import or export memory: between Claude and other AI services (experimental feature)

Claude Opus 4.7 — Most Capable

Claude Sonnet 4.6 — Best Balance

Claude Haiku 4.5 — Fastest

Quality of model output depends on the quality of the input prompt

Prompting: a skill that improves with a structured approach and practice

C-L-E-A-R: a structured methodology for LLM prompting

Designed to improve accuracy, relevance, and reliability

Checklist to ensure prompts are comprehensive and effective

Concise-Logical-Explicit-Adaptive-Reflective

Use only essential words to minimize noise.

Avoid "filler" language like "please" or "thank you,"

"Fillers" distract the model from the core task.

Bad Example: "Can you please help me summarize this report? I would really appreciate it. Thanks!"

Good Example: "Summarize this report in three bullet points."

Organize information in a clear, logical order.

Don’t use two different words for one concept (e.g., "short phrases" vs. "bullets")

Claude 4.x models pay close attention to details - inconsistent terminology will surface in their response

Bad Example: "Summarize this report in three short phrases, and make sure to use bullets."

Good Example: "Summarize this report in three bullet points."

Be specific about what you want.

Claude mirrors your instructions precisely on format, length, and structure

Assign a role when expertise matters:

Avoid vague language that can be ambiguous

Bad Example: "Summarize this report in a way that’s easy to understand."

Good Example: "Summarize this report in three bullet points that a non-technical stakeholder would understand."

These two components are applied together and drive the iterative process of prompt refinement.

You are the human-expert-in-the-loop: apply your SME.

Bad: You accept the first response without revision.

Good:

Claude can help improve prompts — ask it directly:

PDF, DOCX, TXT, CSV, HTML, RTF, ODT, EPUB, JSON

XLSX (requires Code Execution to be enabled in Settings → one-time setup)

JPEG, PNG, GIF, WebP

How It Works

Enabling Web Search

Tips for Effective Use

Limitations or Warnings

An agentic feature that:

Research Mode goes deeper than a single response in a chat

Searches both the public web and any connected internal sources

Delivers thorough output in 5 - 30 minutes with citations

Output is typically:

Not available in free tier accounts

Find the "+" in the chat input box and toggle on the Research Magnifying Glass icon.

A white icon means Research Mode is disabled; click it to enable and turn it blue

Web Search must be enabled for Research Mode to function

Once enabled, Research Mode triggers automatically when prompt requires deep research

Also, explicitly ask Claude to "Research [topic]" to trigger it on demand

Research Mode consumes usage limits faster than standard conversations

Progress indicator shows what research task is running

"Research complete" message appears when done

An expandable chain-of-thought or "thinking" log shows:

Review log during and after response delivery:

State research question clearly and with as much context as possible

Upload any relevant documents or data to ground the research

Front-load constraints in the prompt:

Interject new instructions or constraints if Claude goes off track during research

Ask follow-up questions after the initial research is delivered:

An internal reasoning process Claude runs before producing a response

All major frontier models are being trained to think before they respond

"Thinking" and "reasoning" are used interchangeably.

Available with or without Research Mode

You can toggle thinking on or off, but the newest Claude models controls how deeply it thinks

Thinking produces:

Claude models are migrating to Adaptive Thinking

When to Toggle ON Thinking

When to Toggle OFF Thinking

Switch mid-conversation

Switching by starting a new chat

Claude’s chat search and memory: https://support.claude.com/en/articles/11817273-use-claude-s-chat-search-and-memory-to-build-on-previous-context_

Up-to-date model information: https://platform.claude.com/docs/en/about-claude/models/overview

Up-to-date information on freshness of training data: https://support.claude.com/en/articles/8114494-how-up-to-date-is-claude-s-training-data

Using incognito chat: https://support.claude.com/en/articles/12260368-using-incognito-chats

Link to all Claude support articles: https://support.claude.com/en/collections/4078531-claude

Managing data privacy: https://support.claude.com/en/articles/8325621-i-would-like-to-input-sensitive-data-into-my-chats-with-claude-who-can-view-my-conversations

C-L-E-A-R Framework: Developed by Dr. Leo S. Lo, University of New Mexico, https://digitalrepository.unm.edu/cgi/viewcontent.cgi?article=1214&context=ulls_fsp

Profile preferences are account-wide settings

Help Claude understand your general context and communication needs:

Find Profile Preferences in bottom left Settings menu

What to include:

Styles customize Claude’s tone, format, and delivery

Switch styles via "+" in the chat interface then Use Style

Choose from five built-in presets to match immediate need:

Selection applies to new messages, edits, and retries within that chat

Selection also applies to any new chat you start subsequently until you change it again

Create a custom style via "+" in the chat interface 3 ways:

Manage styles from the style library — rename, reorder, preview, and delete

Edit styles instructions directly or collaborate with Claude to edit

Select styles strategically — match style to audience and purpose for individual conversations

Align your work — Keep all relevant context, files, and instructions together

Retain context — Claude carries knowledge of your engagement forward

Reuse instructions — Write once, apply to every conversation automatically

Persist files — Uploaded files stay available across all project chats

Enable collaboration — Share context with team members on paid plans

Build continuity — Claude grows more useful as the project deepens

Use clear, consistent naming conventions

One project per client, initiative, program, or function

Archive projects when work is complete

View all chats associated with the project

Search chats for specific topics to bring them into the current conversation

Upload additional files that Claude can reference

Edit project instructions at any time

Share the project with collaborators (Team and Enterprise plans)

Archive or delete the project when complete

Reusable templates — structured intake forms, meeting prep guides, or project checklists

Data visualizations — built from your data, exportable and shareable via link

Custom calculators — ROI estimators, pricing tools, or budget planners

Training quizzes — interactive assessments for onboarding or compliance training

Reference documents — standalone deliverables like reports, specs, or policy summaries

Interactive dashboards — KPI trackers, budget monitors, or project status tools with live filters

Knowledge bases — searchable reference tools for policies, procedures, or product specs that respond to natural language queries

Live data tools — Artifacts that pull from Google Drive, Slack, or other connected systems in real time

Workflow automations — multi-step business processes triggered from an Artifact

Find "Artifacts" in the sidebar:

Describe what you want

Claude writes the code

Review it in the side panel

Iterate in conversation

Switch between versions

Share or publish

Publish (Free, Pro, Max) — makes the Artifact publicly accessible to anyone with the link

Share (Team, Enterprise) — restricts access to authenticated members of your organization

How to publish or share:

To customize an Artifact:

Set profile preferences to give Claude …​ consistent account-wide context about who you are and how you work

Use styles to control …​ tone and format per conversation — switch between built-in presets or create custom styles from your own writing samples

Organize work in projects to …​ apply shared context, files, and instructions to every conversation within that workspace

Write project instructions to …​ define Claude’s role and behavior for an entire project — applied automatically without repeating yourself

Build artifacts as …​ reusable, shareable tools that live in your sidebar.

The personalization stack:

Using Styles: https://support.claude.com/en/articles/10181068-configure-and-use-styles

Creating and Managing Projects: https://support.claude.com/en/articles/9519177-how-can-i-create-and-manage-projects

Examples of Projects: https://support.claude.com/en/articles/9529781-examples-of-projects-you-can-create

Building Artifacts: https://support.claude.com/en/articles/9487310-what-are-artifacts-and-how-do-i-use-them

Generates text by predicting likely next words

Does not “know” facts

Produces statistically plausible responses

Confidence in tone does not indicate correctness

Strong at mimicking reasoning, not guaranteed to actually reason

Does not retain or use knowledge like humans do

AI is one component within a workflow, not the final authority

Use human checkpoints for review and output validation

Separate content generation from verification to reduce bias

Use structured inputs and outputs to reduce ambiguity

Refine AI workflows (and artifacts, styles) based on observed performance

Generate first drafts:

Summarize large volumes of information into digestible formats

Identify patterns or themes across datasets or documents

Create multiple options or scenarios for consideration

Final decision-making authority and accountability for outcomes

Interpretation of context, nuance, and stakeholder intent

Evaluation of ethical, legal, and reputational implications

Resolution of ambiguity where multiple interpretations exist

Validation of high-impact or externally facing deliverables

Start with AI assisting humans → progress to AI augmenting workflows → selective automation

Increase AI responsibility only after consistent, validated performance

Break complex work into smaller tasks to enable safe delegation

Ensure decisions can be reversed if AI output is incorrect

Avoid automating tasks prematurely without understanding failure risks

Trust boundaries:

Trust calibration:

Trust boundaries define where AI can and cannot be trusted

Define trust at the task level (not globally)

Determine acceptable error tolerance

Identify areas where AI should NOT be used (regulatory claims, legal advice)

Require evidence for critical outputs (citations, explicit assumptions)

Build trust gradually through repeated evaluation and experience

Compare outputs from rephrased prompts to assess consistency

Ask the model to explain assumptions and reasoning to expose gaps

Test with edge cases or adversarial inputs to identify weaknesses

Cross-check results against independent, reliable sources

Track patterns of accuracy and failure over time

Understand how AI systems fail to calibrate trust

Most AI failures follow consistent and recognizable patterns

Patterns that can be anticipated, detected, and mitigated

Common failure modes:

AI outputs that are plausible and fluent but not grounded in reality

Often occurs when:

Model cannot reliably determine when it is producing incorrect information

External validation is required to confirm accuracy

AI presents specific facts, statistics, or assertions that are untrue

Generates precise-sounding numbers or statements without supporting evidence

Often occurs when asked for data, metrics, or claims

Appears credible due to specificity and confident tone

May be difficult to detect without independent verification

Requires validation of all factual claims before use

References may be fabricated, misattributed, or inaccurately described

Real sources may be cited but don’t support the stated claim

Formatting often appears credible, increasing risk of misuse

Common in research-style or citation-heavy prompts

Requires verification of both source existence and relevance

Never rely on citations without independent confirmation

Treat all AI output as a draft requiring validation

Separate fluency from correctness

Assume failure modes may be present unless explicitly ruled out

Focus attention on claims, evidence, and reasoning

Don’t be fooled by high-quality presentation or confident tone

Use domain expertise to challenge assumptions and conclusions

Let’s distinguish between two different AI failures or risk types

Failures due to model limitations

Failures or risks due to input manipulation

Malicious or misleading instructions embedded within input data

Lays dormant until triggered in:

Attempts to override the intended task or introduce harmful actions

Exploits implicit trust in input sources

Clearly define which instructions the model should follow and which to ignore

Constrain model behavior through system-level prompts or policies

Filter or preprocess inputs to remove or flag embedded instructions

Claude’s "model constitution" is an example of this approach and sets guidelines on:

Limit model or agent actions to predefined capabilities and safe operations

Log inputs and outputs to enable auditing and traceability

Treat all external content as untrusted, regardless of source or format

Tell AI to ignore directives embedded within source material

Reinforce in your prompting what the model should and should not do

Question outputs that seem unusually biased, constrained, or incomplete

Do not rely solely on AI interpretation of external content (validate)

So far: recognized AI failures and protecting against input manipulation

Next: expanding AI use as evidence of AI reliability accumulates

Trust calibration produces the evidence that informs delegation decisions

Risk management and autonomy growth are two sides of the same discipline

Preserve human override mechanisms at all workflow stages

Apply structured review for high-risk or externally facing outputs

Monitor for failure modes, input risks, and unexpected behavior

Reassess trust boundaries as models, inputs, and use cases evolve

Avoid full autonomy where failure cannot be reliably detected or mitigated

Human role is shifting to orchestration, validation, and synthesis — not execution

Capability and reliability are different — delegate based on reliability, not capability

Trust is designed (boundaries) and measured (calibration), not assumed

AI failure modes — hallucination, fabrication, incorrect citations — are predictable and reviewable

Input manipulation risks require treating external content as untrusted by default

Expand AI autonomy gradually, supported by task-level evidence of reliability

Anthropic Trust Center and model cards: https://trust.anthropic.com/

Claude constitution announcement: https://www.anthropic.com/news/claude-new-constitution

Full Claude constitution: https://www.anthropic.com/constitution

Accenture’s "The age of co-intelligence": https://www.accenture.com/us-en/insights/strategy/age-of-co-intelligence

"AI Capability Isn’t Permission": https://www.linkedin.com/pulse/ai-capability-isnt-permission-framework-growing-autonomy-marcia-price-mtb3e

From one-off tasks to repeatable systems

Cowork Projects as persistent workspaces

Scheduled tasks and operational constraints

Plugins and skills for reusable capabilities

Quality, reliability, and iterative improvement

Building your first workflow

Why one-off prompting fails for recurring tasks and how standardization addresses it

Projects as persistent workspaces storing instructions, templates, memory, and connectors

Scheduled tasks and how to plan around the awake constraint

Skills as reusable instruction sets and plugins as bundled tool capabilities

Quality assurance practices including acceptance criteria, output checklists, and the iterative improvement cycle

A step-by-step process for building and graduating to scheduled workflows

What recurring task in your work is the best candidate for your first Cowork Project?

How would you write the acceptance criteria for your most common delegated output?

What is your plan for keeping your machine awake during a critical scheduled task window?

Which of your organization’s standards would benefit from being encoded as a shared skill?

Cowork Documentation

Scheduling Tasks in Cowork

Working with Cowork Projects

Creating and Sharing Skills in Cowork

Match the model and tool to each task

Prompt with the C-L-E-A-R framework

Use memory and chat history intentionally

Stack preferences, project instructions, and styles

Organize ongoing work into projects

Build artifacts as reusable, shareable tools

Distinguish AI capability from AI reliability

Recognize hallucinations, fabrications, and faulty citations

Treat external inputs as untrusted by default

Grow AI autonomy through demonstrated reliability

Apply the C-L-E-A-R framework to your next real task

Set up profile preferences and one project this week

Build a simple artifact you’ll actually reuse

Calibrate trust through low-risk delegations first

Stay grounded — you are the human-in-the-loop

Track Anthropic updates — Claude evolves rapidly