Systematic Debug

Systematic Debugging for Analytics

A 4-phase structured debugging methodology. Follow these phases in strict order. No shortcuts. No cargo-cult debugging. No "just try this and see."

---

Ground Rules

Read these first. They are non-negotiable.

1. NEVER skip Phase 1. Do not attempt a fix before reproducing the error and

reading the full output. No exceptions.

1. NEVER make multiple changes at once. One variable at a time. If you change

two things and the problem goes away, you do not know which one fixed it.

1. 3-Strike Rule. After 3 failed fix attempts, STOP. Summarize what you have

tried, what you have learned, and escalate to the user. Do not keep guessing.

1. State your hypothesis. Before every fix attempt, tell the user what you

believe is wrong and why. "I think X because Y, so I will try Z."

1. Clean up after yourself. Remove all debug instrumentation (extra logging,

LIMIT clauses, temp tables, print statements) before declaring the issue resolved.

---

Phase 1: Reproduce and Gather Evidence

Goal: See the failure with your own eyes. Understand exactly what is happening before forming any opinion about why.

Steps

1. Run the failing query or script exactly as reported.

• Use the same database, schema, and role if possible.
• Do not modify the query before running it.
• Capture the full error output — not just the first line.

1. Read the complete error message.

• Copy the full stack trace or error output.
• Identify the specific line, column, or object that failed.
• Note the error code if one is provided.

1. Check the environment.

• What database/warehouse is this running against?
• What schema or dataset is active?
• Are there any session-level settings (timezone, role, warehouse size)?

1. Check data freshness. (Analytics-specific)

• When was the source data last updated?
• Is the failure caused by stale or missing data rather than a code bug?
• Run SELECT MAX(updated_at) or equivalent on key source tables.

1. Check recent changes.

• git log --oneline -10 on relevant files.
• git diff HEAD~3 -- to see recent modifications.
• Were any upstream models or sources changed recently?

1. Record your findings. Write a brief summary:

• What is the exact error?
• When did it start?
• What is the scope (one model, one query, everything)?

Do NOT proceed to Phase 2 until you can reliably reproduce the failure.

---

Phase 2: Analyze and Hypothesize

Goal: Understand the root cause. Narrow down from "something is broken" to "this specific thing is wrong because of this specific reason."

Steps

1. Trace data lineage.

• If this is a dbt model, trace upstream with ref() and source().
• Identify which input tables feed into the failing query.
• Check whether upstream models built successfully.

1. Find a working example.

• Was this query working before? When? What changed?
• Is there a similar query or model that still works?
• Compare the working version to the failing version line by line.

1. Check for common analytics gotchas. (Analytics-specific)

• NULL aggregation: Is SUM() or COUNT() silently dropping NULLs?
• Fan-out joins: Is a JOIN producing more rows than expected? Check

COUNT(*) before and after the JOIN.

• Timezone mismatches: Are timestamps being compared across different

timezones? Is created_at in UTC but filtered with local time?

• Integer division: Is revenue / quantity doing integer division

instead of decimal?

• Implicit casting: Is a string being compared to a number?
• Schema drift: Did a column get renamed, removed, or change type upstream?
• Duplicate keys: Is the assumed primary key actually unique?

1. Form a specific hypothesis.

• Write it down: "The query fails because [X] causes [Y]."
• The hypothesis must be testable with a single change.
• If you cannot form a hypothesis, you need more evidence — go back to Phase 1.

1. State your hypothesis to the user before proceeding.

---

Phase 3: Test and Fix

Goal: Validate your hypothesis with a minimal, targeted change.

Steps

1. Make the smallest possible change.

• Change one thing. Run the query. Observe the result.
• If you need to test a theory, use a SELECT statement or CTE — do not

modify the production model until you have confirmed the fix.

1. Validate the fix.

• Does the query run without error?
• Does it return the expected number of rows?
• Do the values look reasonable? Spot-check key metrics.

1. Track your attempts.

• Attempt 1: Hypothesis — [X]. Change — [Y]. Result — [Z].
• Attempt 2: Hypothesis — [X]. Change — [Y]. Result — [Z].
• Attempt 3: Hypothesis — [X]. Change — [Y]. Result — [Z].

1. 3-Strike Escalation. If three attempts have failed:

• STOP. Do not try a fourth fix.
• Summarize all three attempts and their results.
• Present your findings to the user.
• Ask for additional context or suggest pairing with someone who has

domain knowledge of this part of the data.

• Only continue after the user provides new direction.

---

Phase 4: Verify and Clean Up

Goal: Confirm the fix is complete, correct, and leaves no mess behind.

Steps

1. Confirm the fix resolves the original error.

• Re-run the exact command from Phase 1 Step 1.
• The error must be gone — not just different.

1. Check for regressions.

• Run downstream models or queries that depend on the fixed model.
• If dbt is available, run dbt test on the affected models.
• Verify that fixing this did not break something else.

1. Validate row counts and metrics. (Analytics-specific)

• Compare row counts to a known baseline or previous run.
• Spot-check 2-3 key metrics against expected values.
• If the fix changed output values, confirm the new values are correct

rather than just "not erroring."

1. Remove all debug instrumentation.

• Delete any temporary tables you created.
• Remove added LIMIT clauses, WHERE 1=0 filters, or debug SELECTs.
• Revert any print() or logging statements added for debugging.
• Confirm the final code is clean and production-ready.

1. Write a summary. Provide the user with:

• Root cause: One sentence explaining what was wrong.
• Fix: One sentence explaining what you changed.
• Verification: Confirmation that the fix works and regressions were checked.
• Prevention: If applicable, suggest a test or check that would catch

this issue earlier in the future (e.g., a dbt test, a CI check, a data quality assertion).

---

Quick Reference Card

Phase 1: REPRODUCE    — Run it. Read the error. Check the data. Check recent changes.
Phase 2: ANALYZE      — Trace lineage. Find working examples. Check gotchas. Hypothesize.
Phase 3: FIX          — One change at a time. Three strikes and escalate.
Phase 4: VERIFY       — Re-run original. Check regressions. Validate metrics. Clean up.

Remember:

• Reproduce before you hypothesize.
• Hypothesize before you fix.
• Fix one thing at a time.
• Three strikes means stop and ask for help.
• Clean up before you declare victory.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: adityawrk
• Source: adityawrk/analytics-with-claude-code
• License: MIT

Install and usage instructions live in the source repository linked above.