CodeAlive-AI Business Investigation
Systematic root cause analysis for business problems. Engineering debugging adapted for CEO tasks.
Triggers: "why no signups", "why churn", "figure out why X", "what's wrong with Y", "investigate", "dig into"
Iron Law
DO NOT PROPOSE SOLUTIONS WITHOUT INVESTIGATING THE ROOT CAUSE.
- Not "let's redesign the landing page" -> first "WHY signups = 0?"
- Not "let's hire sales" -> first "WHY is the current approach not working?"
- Symptomatic treatment creates whack-a-mole debugging. Find the root cause.
Phase 1: Symptom Collection
- Define the problem specifically: not "growth is slow", but "0 signups in 7 days with X visitors" or "3 out of 5 customers didn't complete onboarding"
- Gather data:
- PostHog:
posthog__query-generate-hogql-from-question question="..."for quantitative data - MEMORY.md: what do we already know about this problem?
- Daily logs: were there similar problems before? What was tried?
- Hypotheses.json: are there related hypotheses?
- PostHog:
- Timeline: when did the problem start? What changed? Regression = cause is in the change.
Phase 2: Pattern Matching
Check typical startup patterns:
Funnel Problems
| Where the problem is | Common cause | Where to look |
|---|---|---|
| Traffic -> signup | Positioning / channel / trust | Language, source quality, proof, CTA |
| Signup -> activation | Product / onboarding / TTFV | Steps to first value, friction points |
| Activation -> paid | Outcome proof / buyer / pricing | Value clear? Right buyer? |
| Paid -> retained | Problem frequency / workflow embed | Recurring job? In daily workflow? |
| Retained -> expansion | Narrow value / weak multi-user | Team hooks, second use case |
Structural Problems
| Pattern | Symptom | Root cause |
|---|---|---|
| Wrong ICP | Every customer wants something different | Wedge too broad |
| Wrong channel | High effort, low response | Message-channel mismatch |
| Wrong timing | "Interesting, but not now" | No urgent trigger |
| Wrong buyer | User loves it, buyer won't pay | Value doesn't connect to budget |
| Feature != product | Demo wow, usage dead | Solves interesting != painful problem |
| Trust gap | Enterprise stalls at procurement | No case studies, no compliance, no social proof |
Phase 3: Hypotheses (3-Strike Rule)
Formulate 1-3 root cause hypotheses. For each:
HYPOTHESIS: [specific statement]
Evidence FOR: [what supports it]
Evidence AGAINST: [what refutes it]
How to test: [concrete action in 1-2 days]
If true we'll see: [observable outcome]
If false we'll see: [observable outcome]
3-Strike Rule: if 3 hypotheses turn out to be wrong -> STOP. The problem is likely architectural (wrong market, wrong product, wrong channel). Escalate:
3 hypotheses tested, none confirmed.
The problem may be deeper — not tactics, but strategy.
A) Continue investigation with a new hypothesis: [describe]
B) Launch strategic-review to reassess wedge/ICP
C) Talk to customers directly (CEO = best investigator)
Phase 4: Root Cause -> Recommendation
When root cause is identified:
- Name the cause specifically. Not "bad marketing", but "positioning says 'AI platform' instead of a specific outcome we deliver for [your buyer]"
- Minimal intervention. The smallest change that eliminates the cause.
- Tie to constraint. Does the solution exploit the constraint (market learning speed) or not?
- Kill signal. At what result do we abort the solution?
Phase 5: Report
INVESTIGATION REPORT
---
Symptom: [what we observe]
Root cause: [why this is happening]
Evidence: [data confirming]
Recommendation: [what to do]
Metric: [what to measure]
Kill signal: [when to abort]
Status: DONE / DONE_WITH_CONCERNS / BLOCKED
---
Log to memory/YYYY-MM-DD.md via structured-log.
Telegram Format
Investigation: [problem]
Root cause: [1-2 lines]
Evidence: [key proof]
Recommendation: [action]
Metric: [what to measure]
Status: [DONE/BLOCKED/NEEDS_CONTEXT]
Follow output preferences from USER.md (language, format, platform constraints).
Rules
- Data -> hypothesis -> test -> conclusion. Not the other way around.
- Don't say "should fix it". Show evidence.
- If you can't test — say NEEDS_CONTEXT and request specific data.
- Don't expand scope. Investigate one problem, don't switch.