One skill. One book. One framework. Drop-in rigor for Claude Code, Cursor, or any AI agent that produces plans, analyses, or strategic recommendations.
git clone https://github.com/chefjefff/mckinsey-problem-solving.git
cp -r mckinsey-problem-solving/skills/problem-solving ~/.claude/skills/
# Paste mckinsey-problem-solving/CLAUDE.md into your project's CLAUDE.md
# Done. /problem-solving is now available in any Claude Code session.Jump to install → · See the 7 steps ↓ · Read the case study →
You're using AI agents — Claude Code, Cursor, Cline, something. They generate plans, analyses, and strategic recommendations 10x faster than you used to. The outputs look professional. But they're shallow.
- Analyses list 10 possible causes without ranking them
- Plans jump to solutions without defining the problem
- Strategic recommendations skip the "what would you have to believe?" stress test
- Every output is structured like work but reads like a Wikipedia summary
This isn't a model problem. It's a framework problem. Your AI agents don't have the analytical scaffolding that top consultants internalize over 5+ years of casework.
The fix: give them the framework. One skill file. 7 steps. From the best book ever written on the topic.
Bulletproof Problem Solving: The One Skill That Changes Everything — Charles Conn & Robert McLean (Wiley, 2019)
Conn and McLean spent decades at McKinsey. In 2019 they wrote the first public, systematic description of the 7-step problem solving method that's been used internally at McKinsey for 50+ years. The book has 30 worked examples from business, personal life, and public policy — everything from "where should I move?" to "how do we save Pacific salmon?"
This repo is a distillation of that method into a single Claude Code skill file. If the skill is useful to you, buy the book. The worked examples alone are worth 10x the price.
1. Define the problem ← SMART+, problem worksheet, "porpoise" to refine
2. Disaggregate ← MECE logic trees (component, deductive, hypothesis, decision, ROIC)
3. Prioritize ← 2x2 impact × ability-to-move matrix, prune the tree
4. Build a workplan ← hypothesis-driven, knock-out order, dummy the chart
5. Conduct analysis ← heuristics first (Occam, 80:20, 5 Whys) → big guns only if needed
6. Synthesize findings ← pyramid principle, visualize, surface cross-cutting insights
7. Tell the story ← SCR (Situation/Complication/Resolution), storyboard, lead with answer
Most "frameworks" are 3-5 steps because brevity is marketable. But real problems need all 7:
- Skip step 1 and you solve the wrong problem.
- Skip step 2 and you drown in complexity.
- Skip step 3 and you boil the ocean.
- Skip step 4 and your team runs in circles.
- Skip step 5 and you rely on gut when the data would've told you.
- Skip step 6 and you have facts but no insight.
- Skip step 7 and the best analysis dies in a deck nobody reads.
The 7-step method is comprehensive, iterative, and empirically tested on problems worth billions of dollars. The skill file in this repo preserves all 7 — plus the cleaving frames, bias-fighting team processes, and uncertainty tools from the book.
Beyond the 7 steps, the /problem-solving skill also includes:
Cleaving frames — ROIC trees, price/volume, principal/agent, assets/options, collaborate/compete, supply/demand, regulate/incent, incidence/severity, mitigate/adapt.
Bias-fighting team processes — obligation to dissent, multiple tree structures, "what would you have to believe?", pre-mortem analysis, distributed voting, perspective-taking, Red Team/Blue Team.
Heuristics toolkit — Occam's Razor, order-of-magnitude analysis, 80/20 rule, Rule of 72 (compound growth), S-curves, expected value, Bayesian thinking, break-even analysis, marginal analysis, distribution of outcomes.
Uncertainty levels & strategic responses — the 5-level uncertainty framework plus strategy-under-uncertainty responses (buy information, hedges, options, insurance, no-regrets moves, big bets).
Wicked problems — how to apply the 7-step method to problems with multiple causes, major stakeholder disagreements, and unintended consequences (obesity, overfishing, climate change).
All in one markdown file. ~480 lines. No dependencies.
The setup. A two-person B2B sales team had to decide: "Should we sponsor this $8K healthcare conference?" Their initial plan was a well-formatted checklist: venue logistics, outreach sequences, booth staffing, follow-up bullets.
Running step 1 alone — define the problem — reframed the whole thing. The real problem wasn't "how do we execute this event?" It was "what pipeline outcome does this event solve for, at what audience, at what cost, against what knock-out check?"
Step 2 — disaggregate — revealed the binding constraint. Event ROI was a function of audience quality × attendance × at-event conversion × follow-up discipline. They had a plan for three of those four. The missing piece was the one that mattered most.
Step 3 — prioritize — identified the knock-out. Audience quality was both high-impact and knowable in advance. If the audience was <30% real buyers, the entire sponsorship was wasted. They emailed the conference organizer and got the answer in 48 hours.
Steps 4-7 turned that insight into a workplan with named DRIs, a funnel with explicit conversion math, and a synthesis that the team could present to leadership in 60 seconds.
Full before/after in examples/case-study-event-plan.md.
# 1. Clone this repo
git clone https://github.com/chefjefff/mckinsey-problem-solving.git
# 2. Install the skill (global, works in every Claude Code session)
cp -r mckinsey-problem-solving/skills/problem-solving ~/.claude/skills/
# Or project-local (just this one repo):
# mkdir -p .claude/skills && cp -r mckinsey-problem-solving/skills/problem-solving .claude/skills/
# 3. Paste CLAUDE.md content into your project's CLAUDE.md (at the top)
cat mckinsey-problem-solving/CLAUDE.mdDone. Next Claude Code session, /problem-solving is available.
Works with: Claude Code, Cursor (paste into .cursor/rules/), or any AI IDE that reads markdown skill files.
Verify it works: ask Claude "How should I diagnose why our CAC went up 30% this quarter?" — a correctly-installed response will build an issue tree, apply 5 Whys, and use "what would you have to believe?" testing. A default (uninstalled) response will list 10 possible causes without ranking.
Full install guide (including persistent memory + agent wiring): INSTALL.md.
- Founders and operators using Claude Code / Cursor for strategic work
- Ex-consultants who recognize the framework and want it codified for AI
- Product, growth, sales, and ops teams that ask AI for analysis
- AI builders tired of their agents listing causes without ranking them
Framework source: Charles Conn & Robert McLean, Bulletproof Problem Solving: The One Skill That Changes Everything (Wiley, 2019). If you find the framework useful, buy the book.
Packaging: Jeffrey Zheng. Ex-McKinsey Sydney, now running B2B enterprise sales at Superpower. Built this to make a 2-person sales team + a fleet of Claude Code agents operate like a McKinsey study team.
License: MIT — see LICENSE and ATTRIBUTION.md.
Contributions welcome: see CONTRIBUTING.md.
If this is useful, ⭐ the repo and share it with someone whose AI outputs could use more rigor.