diff --git a/CLAUDE.md b/CLAUDE.md
index 6abc6a6..8725c7d 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -26,6 +26,8 @@ Standalone Skills:
 
 - **`/debrief`** — Produce a structured, blameless debrief document after an incident, failed release, or significant bug.
 
+- **`/debug`** — Structured, interactive hypothesis-driven debugging. Instruments code to test hypotheses, then removes all debug code after fixing.
+
 Each phase persists its output to `docs/` so the next phase can discover it from a cold start.
 
 **Fast path:** **`/hotfix`** — Streamlined workflow for emergency fixes. Skips brainstorm and planning but enforces review and testing. Use when speed matters but quality is still non-negotiable.
diff --git a/skills/debug/SKILL.md b/skills/debug/SKILL.md
new file mode 100644
index 0000000..f73d013
--- /dev/null
+++ b/skills/debug/SKILL.md
@@ -0,0 +1,244 @@
+---
+name: debug
+user-invocable: true
+description: Finds and fixes bugs through structured hypothesis testing and code instrumentation.
+when_to_use: Use when user says "debug", "debug this", "find the bug", "what's causing this", "troubleshoot", or "why is this broken".
+argument-hint: bug description, error message, or reproduction steps
+effort: high
+compatibility: Designed for Claude Code (or similar products with agent support)
+---
+
+# Debug — structured hypothesis-driven debugging
+
+Find and fix bugs through a structured cycle: hypothesize, instrument, reproduce, analyze, fix, verify. All debug instrumentation is removed at the end, leaving only the fix.
+
+## Bug Description
+
+<bug_description>$ARGUMENTS</bug_description>
+
+**If the bug description above is empty, ask the user**: "What's the bug? Describe the symptoms, paste an error message, or explain how to reproduce it. Include any hypotheses you have about the cause."
+
+DO NOT proceed until you have a bug description.
+
+## Phase 0 — Understand
+
+Summarize the bug in one sentence. Identify:
+
+- **Symptom**: what the user observes (error message, wrong behavior, crash)
+- **Suspected area**: which layer or component is likely involved
+- **User hypotheses**: any theories the user provided about the cause
+
+If the user did not provide hypotheses, use **AskUserQuestion**: "Do you have any guesses about what might be causing this? Even rough intuitions help narrow the search."
+
+1. **Yes** — let the user describe their hypothesis
+2. **No, just investigate** — proceed without user hypotheses
+
+## Phase 1 — Hypothesize
+
+### 1.1. Explore the codebase
+
+Run a focused codebase exploration to understand the area around the bug:
+
+- Task @codebase-review-agent("Locate the code responsible for this behavior. Focus on the symptom described — do not survey the entire codebase. Bug: <bug_description>")
+
+After the agent returns, read the identified files and their immediate context.
+
+### 1.2. Formulate hypotheses
+
+Based on the code exploration and any user hypotheses, formulate **2-5 concrete hypotheses** about what could cause the bug. Each hypothesis must be:
+
+- **Specific**: points to a concrete code path or condition
+- **Testable**: can be confirmed or ruled out with debug output
+- **Tagged**: assigned an identifier (H1, H2, H3, ...)
+
+Present the hypotheses to the user:
+
+```text
+## Hypotheses
+
+- **H1**: [description] — test by checking [what to observe]
+- **H2**: [description] — test by checking [what to observe]
+- **H3**: [description] — test by checking [what to observe]
+```
+
+Use **AskUserQuestion**: "Do these hypotheses look right? I'll instrument the code to test them."
+
+1. **Proceed (Recommended)** — instrument and test these hypotheses
+2. **Adjust** — modify or add hypotheses before proceeding
+
+### 1.3. Set up workspace
+
+Before writing any files, ensure the session is on a working branch:
+
+- Call /create-branch to check and optionally create a working branch or worktree.
+
+## Phase 2 — Instrument
+
+Add targeted debug instrumentation to test each hypothesis. Follow these rules:
+
+- **Tag every addition** with a comment containing `WINGSPAN-DEBUG` so it can be reliably found and removed later. Use the language's comment syntax (e.g., `// WINGSPAN-DEBUG`, `# WINGSPAN-DEBUG`, `<!-- WINGSPAN-DEBUG -->`).
+- **Label output by hypothesis**: each log statement must identify which hypothesis it tests (e.g., `[H1]`, `[H2]`).
+- **Log values, not just markers**: capture the actual state (variable values, conditions, return values) that confirms or rules out each hypothesis.
+- **Minimize invasiveness**: add logging only — do not change control flow, add dependencies, or modify behavior.
+- **Keep it simple**: use the project's existing logging mechanism or standard output. Do not introduce new dependencies.
+- **Prefer companion plugin guidance**: if a companion plugin provides debug or logging conventions for the project's stack, follow those over generic defaults.
+
+If the project uses a build or compilation step, run it to confirm the instrumentation compiles. Fix any build failures before proceeding.
+
+After instrumenting, tell the user:
+
+1. Exactly what steps to take to reproduce the bug
+2. Where the debug output will appear (console, log file, etc.)
+3. What to copy/paste back (or which file to point to)
+
+Use **AskUserQuestion**: "I've added debug instrumentation. Please reproduce the bug and share the output."
+
+1. **Here's the output** — user provides the debug output
+2. **Output is in a file** — user provides a path to read
+3. **Adjust instrumentation** — the instrumentation needs changes before reproducing
+
+**If "Adjust instrumentation"**: ask what needs to change, update, and re-present instructions.
+
+## Phase 3 — Analyze
+
+Read the debug output. For each hypothesis, determine:
+
+- **Confirmed**: the output shows the suspected condition is true
+- **Ruled out**: the output shows the suspected condition is false
+- **Inconclusive**: not enough information to decide
+
+### If a hypothesis is confirmed
+
+Summarize the root cause to the user in 2-3 sentences, referencing the specific debug output that confirms it. Proceed to Phase 4.
+
+### If all hypotheses are ruled out
+
+Explain what was learned from ruling them out. Use **AskUserQuestion**:
+
+1. **New hypotheses (Recommended)** — formulate new hypotheses based on what was eliminated, return to Phase 1.2
+2. **Add more instrumentation** — keep existing instrumentation and add more, return to Phase 2
+3. **Stop** — clean up instrumentation and end (go to Phase 6)
+
+### If inconclusive
+
+Explain what was learned and what remains unclear. Use **AskUserQuestion**:
+
+1. **Refine instrumentation (Recommended)** — adjust logging to get clearer signal, return to Phase 2
+2. **New hypotheses** — start fresh with new hypotheses, return to Phase 1.2
+
+## Phase 4 — Fix
+
+### 4.1. Implement the fix
+
+Write the minimal change that addresses the confirmed root cause:
+
+- Change only what is necessary — no drive-by refactors.
+- Match surrounding code style.
+- If the fix grows beyond the original scope, stop and suggest `/plan` → `/build`.
+
+### 4.2. Remove debug instrumentation
+
+Remove ALL debug instrumentation added in Phase 2. Search for `WINGSPAN-DEBUG` across the codebase and remove every tagged line or block. Verify none remain:
+
+```bash
+git grep -n "WINGSPAN-DEBUG"
+```
+
+If any remain, remove them. The codebase must contain only the fix — no debug instrumentation.
+
+### 4.3. Validate
+
+Follow the [validation and fix procedure](references/validate-and-fix.md).
+
+## Phase 5 — Verify
+
+Use **AskUserQuestion**: "I've implemented the fix and removed all debug instrumentation. Please reproduce the original bug to confirm it's resolved."
+
+1. **Fixed** — the bug is resolved
+2. **Still broken** — describe what happened
+3. **New issue introduced** — describe the new problem
+
+**If "Fixed"**: proceed to Phase 7.
+
+**If "Still broken"**: read the user's description. Use **AskUserQuestion**:
+
+1. **Adjust the fix** — root cause was right but fix was incomplete, revise it and return to Phase 4.1
+2. **Re-investigate** — root cause hypothesis may be wrong, return to Phase 1 with new information
+3. **Stop** — revert all changes and end
+
+**If "New issue introduced"**: fix the regression, re-validate, and ask again.
+
+## Phase 6 — Cleanup Only
+
+Reached when the user stops debugging without a fix.
+
+Remove ALL debug instrumentation. Search for `WINGSPAN-DEBUG` across the codebase:
+
+```bash
+git grep -n "WINGSPAN-DEBUG"
+```
+
+Remove every match. Confirm to the user that all debug instrumentation has been removed and the codebase is clean.
+
+## Phase 7 — Complete
+
+### Final validation
+
+Run the project's formatter, linter, and test runner one last time. Fix any failures.
+
+### Handoff
+
+Use **AskUserQuestion**: "Bug fixed and verified! What would you like to do next?"
+
+1. **Commit the fix (Recommended)** — create a commit with the fix
+2. **Add a regression test first** — write a test that reproduces the original bug, then commit
+3. **Done** — end the session
+
+**If "Commit the fix"**: create a single commit:
+
+```text
+fix: <concise description of what was fixed>
+
+Root cause: <1-2 sentence explanation>
+```
+
+**If "Add a regression test first"**: write a test that fails without the fix and passes with it. Run validation, then create the commit.
+
+## Gotchas
+
+- The `WINGSPAN-DEBUG` tag is the single source of truth for cleanup. Every debug addition must include it — no exceptions.
+- If the bug is non-deterministic (race condition, flaky test), instrumentation may need multiple reproduction attempts. Ask the user to reproduce several times if needed.
+- If more than 3 hypothesis-test cycles pass without progress, suggest a different approach: `/brainstorm` for deeper analysis or pairing with a teammate.
+- Debug instrumentation must never be committed. Phase 4.2 and Phase 6 exist specifically to prevent this.
+- If the user's reproduction environment differs from the development environment (production, staging, specific device), note that instrumentation must be compatible with that environment.
+
+## Evaluation queries
+
+### Should trigger
+
+1. "Debug this crash — the app dies when I tap the submit button."
+2. "What's causing the test failure in the auth module?"
+3. "Help me find the bug — users report stale data after refresh."
+4. "Troubleshoot why the API returns 500 on this specific payload."
+5. "Why is this broken? It worked yesterday before the deploy."
+
+### Should NOT trigger
+
+1. "Add a dark mode toggle to the settings screen."
+2. "Review this PR before I merge it."
+3. "Write unit tests for the checkout flow."
+4. "Refactor the data layer to use the repository pattern."
+5. "Create a plan for the new onboarding feature."
+
+### Edge cases
+
+1. "This test is flaky — passes locally, fails in CI." (debugging-adjacent; should trigger)
+2. "The build is broken." (may be a config issue, not a code bug; should trigger)
+3. "Performance is slow on the dashboard page." (performance investigation, not a bug; should not trigger — suggest `/brainstorm` instead)
+
+## Important
+
+- This skill is interactive. It requires the user to reproduce bugs and report results between phases.
+- Keep debug instrumentation minimal and tagged. Every addition gets `WINGSPAN-DEBUG`.
+- The goal is to narrow down the root cause systematically, not to guess-and-check.
+- Remove all debug instrumentation before finishing — the only lasting change should be the fix itself.
diff --git a/skills/debug/references/validate-and-fix.md b/skills/debug/references/validate-and-fix.md
new file mode 120000
index 0000000..173ee62
--- /dev/null
+++ b/skills/debug/references/validate-and-fix.md
@@ -0,0 +1 @@
+../../shared/references/validate-and-fix.md
\ No newline at end of file