Claude for US Knowledge Workers (Not Devs): The 2026 Skills + Projects Playbook

#What we're covering

1. Skills and Projects — the parts nobody explains clearly
2. Playbook 1: M&A due diligence assistant
3. Playbook 2: SEC filing summarizer
4. Playbook 3: HIPAA-safe clinical documentation
5. Playbook 4: MBB-style consulting frameworks
6. Playbook 5: Executive briefing prep
7. Pitfalls, gotchas, and what NOT to trust Claude with
8. FAQ

#Skills and Projects — the parts nobody explains clearly

Anthropic renamed things a few times, so here is the current (mid-2026) map.

Skills

A Skill is a reusable instruction set that tells Claude how to behave for a specific type of task. Think of it as a saved prompt template combined with reference documents. You create a Skill once ("Act as a senior M&A analyst..."), attach any reference material (comparable deal terms, standard NDA templates, valuation methodologies), and then invoke it whenever you start a related task.

What makes Skills genuinely different from "just save a prompt": they can include structured reference files that Claude reads every time, and they can be shared with your team or organization. If your firm has a house style for M&A memos, one person builds the Skill, then every analyst uses it and produces consistent output.

Projects

A Project is a persistent workspace with its own conversation history, uploaded documents, and system instructions. Think of it as a "matter" (if you are a lawyer) or an "engagement" (if you are a consultant). You upload the actual work product for a specific client or deal, and Claude keeps context across conversations. When you close your laptop Friday and come back Monday, the Project still knows you are working on the Delaware acquisition, not the Texas one.

The key difference from just pasting things into ChatGPT: Projects have persistent memory for the work you are doing. That is what makes them useful for actual professional work, where a single question can span weeks.

How they work together

The pattern that actually scales: build a general-purpose Skill for your role, then attach it to Project after Project as you take on new work. The Skill provides the how (methodology, output format, ethical guardrails). The Project provides the what (the specific documents, entities, and history for this engagement). This is a much cleaner separation than "one giant prompt per task" and it survives the way real knowledge work actually happens.

#Playbook 1: M&A due diligence assistant

You are working on a mid-market acquisition. The target's data room just dropped: 47 folders, 800 documents, financials, contracts, employee files, litigation records. Traditional first-pass review is a week of associate hours. Claude can turn it into a day, provided the Skill is set up correctly.

The Skill

Create a new Skill called "M&A Due Diligence Analyst" with this system prompt:

You are a senior M&A due diligence analyst for a US middle-market private equity firm. Your job is to review target company documents and produce findings that flag risks, opportunities, and information gaps in the standard format expected by an investment committee memo.

For every document I share, produce a structured output with these sections:

1. Document classification (contract type, financial statement, HR record, litigation, IP, etc.)
2. Key facts (parties, dates, dollar amounts, jurisdictions)
3. Red flags (change-of-control provisions, non-compete violations, related-party transactions, undisclosed liabilities, tax exposures, IP ownership gaps, key-person dependencies, unusual accounting treatments)
4. Yellow flags (issues that are not disqualifying but require further diligence or negotiation)
5. Green flags (favorable terms, strong protections, competitive advantages)
6. Information gaps (what should exist that I do not see — always include this)

Never speculate about legal conclusions. Flag issues, do not opine on them. Cite the specific document text supporting each finding. If you are not confident, say so.

Output format: markdown, use tables where they aid clarity, keep findings concise but specific.

Attach reference files: your firm's standard rep-and-warranty template, a redacted example of a well-done DD memo from a past deal, and any industry-specific term sheets. Claude will use these to calibrate output.

The Project

Create a Project for the specific deal (e.g., "Project Sequoia — Acquisition of Redwood Manufacturing"). Upload the data room in batches: financials to their own conversation, contracts to another, HR/legal to a third. Ask Claude to work through each batch, then produce a consolidated memo pulling the top findings across all conversations.

The critical piece: at the end, ask for what is missing. Claude is unusually good at flagging "you would expect to see the 2023 tax returns but they are not in this folder." That is often the most valuable output.

What this saves

A first-pass DD review that would take a junior associate 40-60 hours drops to 5-8 hours of analyst time with the Skill+Project setup. You still need a human reviewer for the final memo — Claude is the associate here, not the partner. But the leverage is real.

#Playbook 2: SEC filing summarizer

If you cover public companies as an analyst, banker, journalist, or investor, you are reading 10-Ks, 10-Qs, 8-Ks, and proxies constantly. Each of these is 100-400 pages of boilerplate wrapped around 20-40 pages of actual content. Claude excels at this specific extraction task.

The Skill

Create "SEC Filing Extractor" with this instruction:

You are a securities analyst who reads SEC filings for a living. When I share a filing, produce a structured summary in this exact format:

1. Filing type and filer (issuer name, ticker, filing type, period covered, filed date)
2. The story (1-2 paragraphs: what is this filing telling us that was not obvious before?)
3. Key financial changes vs. prior period (revenue, EBITDA, margins, guidance, cash position — with the actual numbers and % changes)
4. Material developments from the Business/Risk Factors/MD&A sections (only truly new items, not repeated boilerplate)
5. Litigation and regulatory (new matters, updates to existing, settlements)
6. Related-party and executive compensation notes (anything unusual)
7. Auditor's opinion (unqualified, qualified, going concern, or PCAOB items of note)
8. Buried leads (items in the back of the filing or in footnotes that would matter to an informed reader)

Cite the specific section (item number, page number, or footnote) for every finding. Never guess at forward-looking implications. If the filing contains restatements or accounting changes, flag them prominently in section 2.

The Project

One Project per company you cover. Upload every filing as they come out. Over time, Claude builds context across quarters — it will notice things like "guidance has been revised downward three quarters in a row" or "the same risk factor was expanded in this 10-Q vs. the 10-K." Persistent context is what makes this valuable beyond a one-off summary.

Bonus workflow: when you are prepping for an earnings call, ask Claude "based on the last 8 quarters of filings and calls in this project, what are the three questions I should ask that management is unlikely to have pre-scripted answers for?" It is very good at surfacing the topics executives have been dodging.

#Playbook 3: HIPAA-safe clinical documentation

This one comes with the biggest caveats, so lead with them: Claude via Anthropic's public API is NOT HIPAA-compliant. If you are a clinician handling actual patient PHI, you need either a Business Associate Agreement (BAA) with Anthropic via their enterprise tier, or you use Claude through a HIPAA-compliant intermediary like Amazon Bedrock (which does offer a BAA for Anthropic models). Do not paste real patient identifiers into consumer Claude.

With that said, once you have the BAA layer in place, Claude is genuinely useful for the drudgery of clinical documentation: SOAP notes, discharge summaries, referral letters, and pre-authorization prose for insurers.

The Skill

Create "Clinical Documentation Assistant" with this instruction:

You are a clinical documentation assistant for a US physician. You produce well-formatted, factually accurate notes based on my dictated or typed encounter summaries. You NEVER add clinical findings that were not in my input — hallucinated diagnoses in patient records can cause real harm.

Output formats I use:

• SOAP note: Subjective / Objective / Assessment / Plan, each section clearly labeled, using standard clinical abbreviations.
• Discharge summary: Reason for admission, hospital course, procedures, medications at discharge, follow-up, patient education completed.
• Referral letter: Reason for referral, relevant history, current medications, specific question I want the specialist to address.
• Prior authorization letter: Patient info, diagnosis with ICD-10, specific therapy being requested, medical necessity justification citing clinical guidelines when applicable.

When I share encounter details, ask which output I want if it is ambiguous. Use only my clinical findings — do not add PMH, allergies, or medications I have not mentioned. Flag any inconsistencies in the input (e.g., "the note says the patient is allergic to penicillin but you have prescribed amoxicillin — please confirm").

The Project

Create a Project per patient (if allowed by your organization's data handling policy) or per encounter type. Skill provides the format; Project provides the patient-specific history. Over time Claude can flag things like "you have documented this patient's HbA1c for the last four visits, this one is significantly higher — do you want to note that trend in the Assessment?"

The productivity gain here is well-documented — clinicians using AI documentation assistants report 30-60 minute reductions in after-hours "pajama time" charting per day. That is a meaningful clinical and personal outcome.

#Playbook 4: MBB-style consulting frameworks

If you are a consultant at a strategy firm — or trying to think like one — Claude is a genuinely capable framework partner. It knows the standard toolkit (Porter's Five Forces, McKinsey 7S, BCG growth-share matrix, jobs-to-be-done, MECE analysis, hypothesis pyramid) and can apply them structurally to a real business situation.

The Skill

Create "Strategy Consultant" with this instruction:

You are a senior strategy consultant at a top-tier firm (think McKinsey / BCG / Bain). Your job is to apply structured frameworks to business problems and produce output that a client-facing partner would be comfortable presenting.

Your operating principles:

1. Frameworks first, opinions second. When I describe a client situation, propose 2-3 frameworks that fit before diving into analysis. Explain why each fits.
2. MECE always. Any list, breakdown, or analysis must be Mutually Exclusive and Collectively Exhaustive. If it is not, flag the gap.
3. Hypothesis pyramid. Every recommendation traces up to a governing hypothesis. Sub-arguments support the main claim. State the pyramid explicitly.
4. Data-driven where possible, humble where not. If you do not have the specific data, say "would need X data to test this" rather than guessing.
5. Output format: Executive summary (3-5 bullets), then structured analysis, then recommendations with prioritization, then risks and mitigations, then required next steps.

Never invent client data, market sizes, or benchmarks. When you cite a benchmark, state its source (public reports, well-known consulting surveys, or specifically flag as "illustrative — needs sourcing"). If I ask for something that would require actual client data I have not shared, tell me what data you would need.

The Project

One Project per engagement. Upload the client's public filings, industry reports, executive interview notes (with names redacted if needed for confidentiality). Claude can then produce the working analysis, hypothesis testing, and slide-outline drafts throughout the engagement while maintaining context.

The specific technique that produces the best output: ask for "the wrong answer first." Prompt Claude to argue the opposite of your working hypothesis, then defend against those counter-arguments. It surfaces the weaknesses in your own reasoning faster than any solo strategy session.

#Playbook 5: Executive briefing prep

If your job involves preparing briefings for executives — a board packet, a CEO's daily read, a Congressional testimony prep — you are essentially in the compression-and-synthesis business. Claude is extraordinarily good at this specific task.

The Skill

Create "Executive Briefing Writer" with:

You are an experienced executive briefing writer preparing materials for a busy US executive audience (CEO, board, senior government official, or equivalent). Your outputs must respect the reader's time and cognitive load.

Style rules:

• Lead with the bottom line — the single sentence the executive would tell their boss.
• Structure: Bottom Line Up Front (BLUF), then Context (max 3 sentences), then Key Facts (bulleted, no more than 5), then Recommendation (with 2-3 alternatives if applicable), then Timeline / Action Required.
• Length target: one page equivalent unless specified otherwise. Every additional page reduces the probability the reader will finish.
• Never bury the ask. If action is required, state it explicitly.
• Attribute claims to sources. Executives get burned when a fact turns out to be wrong.
• Anticipate the follow-up questions in a separate "Q&A" section.

When I share source materials, produce the briefing in the specified format. When I do not specify a format, default to BLUF-Context-Facts-Recommendation-Action.

The Project

One Project per recurring briefing recipient (e.g., "Weekly Briefings — CEO," "Board Prep — Q3 Meeting"). Over time the Project learns the executive's preferences, the topics they follow, the phrasing they prefer. This is genuinely where Claude's persistent context shines.

The compounding effect: three months in, when you upload a new source document and ask for a brief, Claude produces something that sounds like your voice on that executive's topics. That is not a party trick — that is a real productivity gain.

#Pitfalls, gotchas, and what NOT to trust Claude with

Confidentiality

Claude via the Anthropic consumer product does not train on your data (per their policy), but your data is transmitted to their servers. For truly confidential work — attorney-client privileged material, regulated financial data, PHI — use the enterprise tier with a Data Processing Agreement, or use Claude through Bedrock/Azure/GCP where you can constrain the data flow.

Hallucinated citations

Claude is generally better than GPT-4 at citation accuracy, but it still hallucinates. Any citation to a legal case, SEC filing item number, or regulatory reference must be verified. The specific failure mode: Claude confidently produces a case name and citation that looks entirely real but does not exist. If you use it in a court filing, you can be sanctioned. Lawyers have been.

Numbers

Claude has gotten better at arithmetic since 2024, but for anything involving multiple calculations, you should either (a) have Claude use its analysis tool (Python execution) which is much more reliable, or (b) recompute the numbers yourself. Do not put Claude-computed financial totals into a client deliverable without verification.

Overconfidence framing

Claude will produce output that sounds authoritative even when it is guessing. The mitigation: explicitly instruct it to "flag any claim you are not confident about" and "if you would need specific data to answer, say so and tell me what data you would need." Do this in every Skill you create.

Getting locked into one platform

Both Skills and Projects live in Claude specifically. If Anthropic changes their pricing or feature set (they will), you may need to rebuild the same workflows in another product. Keep your Skill definitions in your own notes as well — they are essentially prompt templates, not proprietary to Anthropic.

#FAQ

#Bottom line

Claude is not built for developers. It is built for knowledge workers whose actual output is prose, analysis, and structured judgment — and it happens to be free-to-cheap and available today. Every playbook above will save you real hours per week if you invest an afternoon setting up the Skill and Project correctly.

Two rules to close on: keep humans in the loop for anything with real consequence (nothing goes to a client, court, or regulator without your review), and be honest with yourself about what Claude is producing versus what your firm's methodology actually requires. The AI is fast, but only you are accountable.

For more prompt-engineering guides, browse PromptSpace, use our AI prompt generator to build custom Skills prompts, and read the 50 Gemini prompts for business companion. If you also touch code, our Claude Code skills for solo developers covers the developer side.