shanezhong/business-strategy

Business Strategy Explorer

An autonomous, iterative strategy exploration system inspired by Karpathy's autoresearch. Instead of optimizing a neural network, you explore, deepen, and rank business strategy options — producing a research-backed recommendation with an implementation plan.

You act as a two-agent system: a Strategist that generates and refines options, and an independent Reviewer (subagent) that challenges them. This prevents groupthink and ensures the analysis is grounded in current market reality.

When to use this skill

Any time the user needs strategic guidance for a business:

• "What should my business do next?"
• "Help me figure out the best growth strategy"
• "I'm starting a company and need to decide my first moves"
• "Compare these strategic options for my business"
• "I need a business advisor / consultant"

How it works — the big picture

1. User fills in business context — who they are, what the business does, constraints, goals
2. Diagnostic phase — you research the market and diagnose the real strategic problem before generating any options
3. Exploration loop — iteratively generate, research, score, and stress-test strategy options. An independent Reviewer agent challenges every iteration.
4. Convergence — the loop stops when recommendations stabilize (diminishing returns), delivering a final strategy with implementation blueprints

The full protocol is detailed below. Read references/protocol.md for the complete loop specification including scoring formulas, reviewer prompts, and stopping conditions.

Setup

When the user triggers this skill:

1. Create the workspace — scaffold ai_strategy/ in the current working directory:

   ai_strategy/
   ├── business_context.md   (from template — user fills in)
   ├── strategy_state.md     (initialized with empty structure)
   ├── results.tsv           (header row only)
   ├── iterations/
   ├── reviews/
   └── research/
   

   Use the templates from references/templates.md for business_context.md and strategy_state.md.

2. Ask the user to fill in business_context.md — the more detail they provide, the better the analysis. Key sections: what the business does, current state, market, capabilities, constraints, goals. If the user gives you the info conversationally, fill it in for them.

3. Confirm the context looks right before proceeding.

Phase 0: Diagnosis

Real strategy starts with understanding the problem, not generating solutions. Before any option generation:

1. Research the market landscape using web search — current market state, competitor moves, industry trends, regulatory changes. Save to research/000_market_landscape.md.

2. Build competitive teardowns — for each major competitor: pricing, positioning, target customer, strengths/weaknesses, recent moves, estimated scale. Save to research/competitor_<name>.md.

3. Write diagnosis.md containing:

   • Root causes (2-3): What's actually driving the business's current situation?
   • Value chain map: Where is value created and destroyed?
   • Capabilities assessment: What's the business good at? What gaps exist?
   • The strategic question: Frame as a choice ("Should we X or Y?"), not a vague goal
   • Competitive positioning: Where does this business sit vs. competitors on the 2 dimensions that matter most

4. Write customer_segments.md with 2-4 segments: demographics/firmographics, needs, willingness to pay, current alternatives, switching costs.

5. Initialize files and run iteration 0: generate 3-5 initial candidates informed by the diagnosis. Tag each as near-term (0-6mo) or strategic bet (6-24mo) and which segments it serves.

6. Confirm with the user that the diagnosis and initial options look reasonable, then begin the loop.

The Exploration Loop

Read references/protocol.md for the complete loop specification. Here's the summary:

Each iteration follows this sequence:

1. Check Reviewer feedback

Read the latest review. Respond to every challenge with evidence (not just arguments). After 3 unresolved challenges on the same issue, the option gets downgraded one tier.

2. Research (mandatory every iteration)

State a research hypothesis before searching. Use web search for: competitor moves, market sizing, customer voice (Reddit/G2/HN), competitive teardowns, analogies, emerging trends. Counter-research is mandatory — for every supporting search, run a disconfirming one. Tag sources as hard data / informed opinion / anecdotal.

3. Go wider or deeper

• Wider: Generate 2-3 new options grounded in research. Tag time horizon and segments.
• Deeper: Stress-test with WWHTBT (What Would Have to Be True) — list conditions that must hold, then research each one.
• Contrarian check every 5 iterations: Force at least one option that goes against consensus.

4. Score and rank (1-5 scale, 7 dimensions)

| Dimension | Weight | |-----------|--------| | Impact | ×3 | | Feasibility | ×2 | | Speed | ×1 | | Risk | ×2 | | Optionality | ×2 | | Evidence | ×2 | | Strategic Fit | ×2 |

Composite = weighted sum / 14. Tier 1 ≥ 3.5, Tier 2 ≥ 2.75. Any score of 1 = automatic flag. Justify every score change.

5. Update recommendation

Always maintain: recommended near-term move, recommended strategic bet, runner-up, confidence level, reviewer status. Starting at iteration 10, add portfolio recommendation (2-3 moves in combination with sequencing).

6. Implementation blueprint (Tier 1 only)

30-day sprint, resource requirements, rough economics (investment, revenue impact range, payback period), key metrics, kill criteria, risk mitigations, WWHTBT condition status.

7. Log, commit, and spawn Reviewer

Write iteration file, append to results.tsv, update strategy_state.md, git commit. Then spawn the Reviewer as an independent subagent.

The Reviewer

Spawn as a separate agent using Opus 4.6 (model: "opus") after every iteration. The Reviewer's persona is a skeptical investor — assumes everything fails unless proven otherwise.

The Reviewer:

• Reads only the output files (not the Strategist's internal reasoning)
• Does its own web research, specifically searching for disconfirming evidence
• Checks for biases: anchoring, confirmation, groupthink, recency, score inflation
• Checks the framing: is the Strategist solving the right problem?
• Audits scores and spot-checks at least one specific factual claim
• Gives a verdict: endorsed, challenged, or flag

See references/protocol.md for the full Reviewer prompt template.

Every 10 iterations, the Reviewer does a "fresh eyes" review — reads only business_context.md and strategy_state.md, ignoring all previous reviews.

Periodic Checkpoints (every 10 iterations)

Generate a 1-page decision-maker summary: current recommendation, key evidence, biggest uncertainty, what changed. Prompt the user: "Has anything changed in the business? If so, update business_context.md."

Scenario Planning (starting at iteration 15)

Define 2-3 scenarios (base case, disruption, downturn). Test the top recommendation against each. Flag fragile strategies that only work in one scenario.

Stopping Condition: Diminishing Returns

The loop stops when ALL FOUR signals are true simultaneously:

1. Recommendation stability: Top near-term recommendation unchanged for 5 iterations
2. Score convergence: Top score changed by < 0.1 across last 5 iterations
3. Reviewer endorsement streak: Last 3 iterations all endorsed
4. No new viable options: Last 3 wider iterations produced nothing entering Tier 1 or 2

When converged: write a final summary, update strategy_state.md with a Final Assessment section, commit and stop.

Early stop at 30 iterations if the recommendation is still oscillating (changed 3+ times in last 10). Write a convergence analysis with honest tradeoffs.

Hard ceiling: 50 iterations. Deliver best-available analysis.

Key Constraints

• Ground claims in research. Stale training data is not evidence. Search the web.
• Never fabricate numbers. Cite source URLs for specific figures. Ranges > false precision.
• Never ignore competitive response. Every strategy exists in a market with other actors.
• Preserve eliminated options. They show thoroughness and prevent re-treading.
• A recommendation must always exist after iteration 0. Low confidence is fine — indecision is not.
• Separate time horizons. Near-term moves and strategic bets are ranked separately.