Step By Step Debug

/step-by-step-debug Mode

A structured debugging companion that refuses to guess. Before any code changes, Claude understands the bug, verifies it against evidence, and agrees on root cause with the user. Only then does it shift into the disciplined fix workflow. Six phases:

1. Interview — question the user until the bug is fully understood
2. Screenshot — request UI evidence if the issue is visual
3. Verify — confirm the evidence matches the described symptoms
4. Analyze — identify root cause, present findings, agree with the user
5. Fix — plan → execute → optional review → test → summary
6. (Fix reuses the full step-by-step workflow; see below.)

Activate: /step-by-step-debug (alone or with a bug description). Deactivate: /done, "exit debug mode", "just fix it". On activation, confirm briefly, run the glossary check, then start Phase 1: "Debug mode on. Let me get oriented before we dig in." When the fix begins (Phase 5), ask for a pace exactly as /step-by-step does — Fine (every step), Balanced (every 3), or Fast (every 5); steps stay atomic, pace only sets how often Claude stops for sign-off. Default Balanced.

---

Bundled Reference Files

• references/ubiquitous-language.md — DDD glossary extractor. Load at activation.
• references/execution-and-review.md — fix-phase mechanics: atomic steps, optional reviewer

subagent, broken-build handling, parallelism, mid-execution feedback. Load when entering Phase 5.

• references/context-management.md — leanness rules, SESSION STATE, /compact prompting.

---

Ubiquitous Language (at activation, before Phase 1)

Same as /step-by-step: check for UBIQUITOUS_LANGUAGE.md; if missing, generate it from available context using references/ubiquitous-language.md and tell the user. Use canonical terms in all phases, correct aliases gently, and note glossary updates in the summary.

---

Phase 1: Interview — Understand the Bug

Never assume you understand the bug from the initial description. Interview until every item is resolved. Ask 2–4 focused questions per round, wait, follow up on anything unclear. Be direct about vagueness. Escape valve: if the opening message already answers most items, only ask about genuine gaps.

• Symptom — exact observed behaviour (error/output/glitch/crash); expected behaviour; when it first appeared
• Reproduction — reliable repro steps; always/sometimes/conditional; environment (browser, OS, device, API version, data state)
• Scope — isolated or multiple places; all users/data or specific; reported by others or one-off
• Context — recent changes (deploys, config, deps, migrations); available logs/traces; prior fix attempts

Close by summarising: symptom, expected, reproduces, first seen, scope, key log/trace info. Get confirmation before proceeding. Prompt /compact after the interview.

---

Phase 2: Screenshot — Visual Evidence

If the bug has any visual component (layout, styling, interaction, rendering, form behaviour, navigation, visual state), ask for a screenshot showing the problem — and a working "before" if one exists. Wait for it. If the user can't provide one, note the caveat and proceed.

If the bug is purely backend/logic/data, skip this phase silently.

---

Phase 3: Verify — Confirm the Issue

Before theorising about root cause, read the relevant code, logs, and any screenshots. Confirm: the code path that would produce this symptom exists; the described repro actually reaches the broken code; the screenshot matches the described symptom. If there's a discrepancy between what was described and what the evidence shows, surface it and resolve it before analysis:

⚠️ Before I dig into root cause: [discrepancy]. Could you clarify? I want to
make sure I'm investigating the right thing.

---

Phase 4: Analyze — Root Cause and Findings

Investigate systematically; don't jump to the obvious explanation. Then present findings and get agreement before proposing a fix:

🔍 Root cause analysis:
**Immediate cause:** [the specific thing that directly produces the symptom]
**Underlying cause:** [why that thing exists / what allowed it]
**Contributing factors:** [anything compounding it, if any]
**Blast radius:** [other areas that may be affected]
**Proposed fix direction:** [high-level approach — not the implementation yet]

Does this match your understanding? Anything to add or correct before I plan the fix?

Do not move to the fix until root cause is mutually agreed. Prompt /compact after agreement.

---

Phase 5: Fix — Step-by-Step Execution

From here the workflow is identical to /step-by-step. Load references/execution-and-review.md and references/context-management.md, then:

• Plan — read the codebase for conventions, produce a full numbered step breakdown

(title + one-sentence rationale), run parallelism analysis, get explicit approval before coding.

• Execute — one atomic change per step; announce → change → report. Checkpoint at the

chosen pace (Fine/Balanced/Fast): in Balanced/Fast, run the batch then present one consolidated checkpoint with the optional review. Pause a batch early on a broken build, a reviewer flag, or a step needing a user decision. Keep responses lean; follow the SESSION STATE / /compact rhythm per checkpoint.

• Non-functional checkpoint — after the fix steps, ask whether error handling, logging,

analytics/telemetry, or performance need attention for the changes made; add atomic steps for any the user picks.

• Test — derive scenarios from the actual fix (include a regression test for the bug itself),

get scenario approval, ask review mode (per-test vs all-at-end), then write.

• Summary — full change summary including a one-line note on the original bug and its root

cause, plus Deferred Review Items and Glossary updates if any.

---

Hard Rules

Always: check/create UBIQUITOUS_LANGUAGE.md before Phase 1; complete each phase before the next (interview → screenshot → verify → analyze → fix); ask for a screenshot for any visual bug; surface evidence discrepancies before theorising; get explicit agreement on root cause before planning; include a regression test for the bug; apply the full step-by-step fix workflow without shortcuts; keep steps atomic; get sign-off before advancing.

Never: skip the interview because the bug "seems obvious"; theorise about root cause before verifying evidence; propose a fix before root cause is agreed; skip the screenshot for a visual bug; bundle multiple concerns into one step; proceed without confirmation; compact mid-step.

---

Example (abridged)

User: /step-by-step-debug The checkout button stops working after applying a discount code

Claude: Debug mode on. Let me get oriented.
[glossary check] A few questions:
1. Does the button do nothing, show an error, or fail silently?
2. All discount codes or specific ones?
3. Anything in the console or network tab?

User: Goes grey and unclickable. All codes. Console: "TypeError: Cannot read
properties of undefined (reading 'total')". Started after yesterday's deploy.

Claude: [summarises bug, confirms] Since this is a UI issue, could you share a
screenshot of the greyed-out button?

User: [attaches screenshot]

Claude: [verifies code path against screenshot]
🔍 Root cause analysis:
**Immediate cause:** applyDiscount() omits `total` when a discount zeroes the cart.
**Underlying cause:** yesterday's refactor strips zero-value fields for payload size,
but checkout expects `total` to always exist.
**Contributing factors:** no type guard on cart.total; no 100%-discount test.
**Blast radius:** any component reading cart.total; order summary display.
**Proposed fix direction:** always return `total` (default 0), add a defensive guard.
Does this match your understanding?

User: Yes, let's fix it.

Claude: Here's the plan:
1. **Fix applyDiscount() response** — always include `total`, default 0
2. **Add guard in checkout component** — treat undefined total as 0
3. **Check order summary component** — apply the same guard if needed
⚡ Steps 2 & 3 can run in parallel (different components, no shared deps).
Does this breakdown look right?

Source & license

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

• Author: arvindraghupathy
• Source: arvindraghupathy/claude-skills
• License: MIT

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