Claude Skills
SkillsResultsReplayCollectionsGuidesNominate
© 2026 Curated Agent SkillsBrowse skillsHistorical indexAbout this guide·Learn more about Agent Skills
All results/Jesse Vincent Superpowers

Debugging · independently reviewed

Systematic Debugging by Jesse Vincent

Runs a four-phase root-cause process and treats symptom fixes as failures. Best suited to: Regressions, intermittent failures, performance anomalies, and unfamiliar systems where the cause has not been demonstrated.

Skill by Jesse VincentCurated Agent Skills reviewReviewed · 2026-09-15Page updated · 2026-11-15Seeded teaching fixture
Try this skill

Actual skill result

Why a zero-quantity line still costs money

  1. 01reproduce
  2. 02trace
  3. 03isolate
  4. 04patch
  5. 05verify
fixture.mjs · lineTotal()
export function lineTotal(priceCents, quantity) {
  return priceCents * (quantity || 1);
}
0 || 1 → 1 → 1,250 cents
node --test exit 1
✖ zero quantity costs zero
✔ omitted quantity defaults to one
✔ null quantity defaults to one
✔ two units cost twice the price

tests 4 · pass 3 · fail 1

A removed cart line still costs 1,250 cents. Expected: zero.

10-second explanatory replay. Time is compressed; this is not a new test run.Retained test log ↗
Task
Fix a zero-quantity cart calculation
Output
One-operator patch + regression logs
Verification
Same suite: 1 failure before, 4 passes after
Boundary
Agent-seeded teaching fixture; not an independent benchmark

Recorded run · 2026-10-30 · Codex instruction execution

Request, reproduction and run notes

Trace: lineTotal(1250, 0) → quantity || 1 → 1 → 1250. Replace || with ?? at the source of the lost zero. Regression: 1 failure / 3 passes before; 4 passes / 0 failures after.

Original request (excerpt; full fixture linked in evidence)

A removed cart line (quantity 0) still contributes 1250 cents. Diagnose and fix the fixture; keep omitted quantity defaulting to one.

Reproduce this run

  1. Download before.mjs, fixture.test.mjs and fix.patch into a clean folder; copy before.mjs to fixture.mjs.
  2. Run node --test --test-reporter=spec fixture.test.mjs. Expect one failing zero-quantity assertion.
  3. Apply patch -p0 < fix.patch, then run the same command. Expect all four tests to pass. Compare fixture.mjs and the retained logs.

Generative output may differ. The retained files establish this run’s result.

Manual intervention and adaptations: Agent deliberately seeded a tiny teaching bug, wrote tests, traced the value before patching and ran the real Node test runner. No human patch edits. Published stack traces replace the temporary absolute directory with relative filenames; counts and assertions are unchanged.

Agent: Codex desktop agent (direct instruction execution in the authoring session). Model: GPT-6 family. Exact deployment identifier unavailable in this session. Run date: 2026-10-30. Full source, environment, verification, limitations and artifacts follow in Evidence and provenance.

Our take Editorial verdict

A four-phase root-cause process that treats symptom fixes as failures. It forces the agent to reproduce, isolate, understand, and only then change code, which is exactly the discipline most agents skip when they start guessing at a fix.

Best for

Regressions, intermittent failures, performance anomalies, and unfamiliar systems where the cause has not been demonstrated.

Keep in mind Limits and tradeoffs

It is slower than a direct fix for obvious defects and still needs a reproducible signal; without one the process can stall instead of converging.

How to find a root cause before patching a bug

  1. Provide the symptom, expected behavior, reproduction steps and recent changes.
  2. Ask the agent to gather evidence, compare hypotheses and isolate the responsible path before making a fix.
  3. Verify the original reproduction and a regression check after the patch; inspect the output instead of accepting an unsupported success claim.

Recorded 2026-10-30: Actual four-phase debugging exercise with red/green regression evidence. The real fixture suite failed the zero case before the single-operator patch and passed all four cases afterward. Seeded fixture, not a production incident. Invalid, negative and fractional quantities are outside the fixture contract. Supporting TDD and verification instructions were also read; see run protocol. This record does not establish successful installation or general effectiveness.

Explore Jesse Vincent Superpowers · Skill reference and original instructions

Try this skill

InstallGet starter request

Install · Collection / plugin

Run in Claude Code chat. Installs the whole Superpowers plugin: the methodology, other skills and harness integration. Install separately for each agent. Scope is chosen by the agent plugin manager.

/plugin install superpowers@claude-plugins-official

Command / route checked 30 October 2026. Upstream installation

Setup requirements and caveats

Run in Claude Code chat. Installs the whole Superpowers plugin: the methodology, other skills and harness integration. Install separately for each agent. Scope is chosen by the agent plugin manager.

Node.js with node:test; standard patch utility

These are documented requirements; installation has not been tested here.

Your agent needs its own account, runtime and model access. Usage charges depend on your plan; unrecorded costs are unknown.

Recorded outside costs:
  • Existing Codex account; incremental charges were not measured

Run the example

Start a fresh session in the target project. Provide the relevant files or diff. This is the review’s example request, not a record of your execution.

Use the systematic-debugging skill. Find the root cause of intermittent webhook timeouts without editing production code until the failing path is isolated.

Check your result · Successful first run

A reproduction and evidence-backed root cause before a patch is proposed.

Confirm the agent loaded the named skill, inspect the output and run the relevant checks. Installation alone does not prove a successful run.

Something not working?

The skill is missing from my agent

Check that the skill is installed for the selected agent and project/global scope. Open a fresh session in that project and check the agent’s skill list. Plugin installs may also need to be enabled.

The wrong skill runs, or nothing happens

Start a fresh conversation, explicitly name the skill using the request above, and provide the requested inputs. Ask the agent to confirm which skill it loaded. Check for duplicate installed copies or overlapping instructions.

The skill runs but the result is poor

Compare the result with the successful-first-run check above. Supply the missing files, reproduction steps or constraints, then ask for one focused correction. If the skill addresses a different problem, return to the task guide and compare alternatives.

Installation troubleshooting · missing dependencies

If npx fails, check the Node/npm versions in setup requirements, Git and network access. For skills CLI installs, run npx skills list (add -g for global). If a route changed, follow upstream instructions.

Collection background and tradeoffs

The methodology is deliberately heavy. Brainstorming, chunked design review, and a full written plan add real friction to small or already-clear changes, and the automatic skill-invocation rule can feel rigid. Its richest behavior assumes a harness with subagents, hooks, and a plugin marketplace, so the experience is uneven across agents even though the skills themselves travel well.

Source maintained by Jesse Vincent. Read original skill. Compatibility guidance and published showcase runs remain separate from your first run.

Evidence and provenance Sources and retained files

Evidence and provenance

Executed test · passed

Test date: 2026-10-30

Actual four-phase debugging exercise with red/green regression evidence.

Task: Reproduce, investigate, hypothesize and fix the zero-quantity default bug.

Result: The real fixture suite failed the zero case before the single-operator patch and passed all four cases afterward.

Evidence source · Verified revision: 2ce8aa15d05bd57213e6f746f953aafb7ded7a8b

Agent: Codex desktop agent (direct instruction execution in the authoring session)

Model: GPT-6 family. Exact deployment identifier unavailable in this session.

Environment: macOS / Darwin arm64; Node.js v26.8.1; Self-contained local fixtures; no external service calls in generated examples

Prerequisites: Node.js with node:test; standard patch utility

Outside costs: Existing Codex account; incremental charges were not measured

Limitations: Seeded fixture, not a production incident; Invalid, negative and fractional quantities are outside the fixture contract; Supporting TDD and verification instructions were also read; see run protocol

  • Run protocol, provenance, rights and reproduction
  • Investigation and hypothesis
  • Broken fixture
  • Regression suite
  • Actual failing test log
  • Minimal patch
  • Fixed fixture
  • Actual passing test log

Related workflows

Systematic Debugging
Try this skill

Agent compatibility

Claude Code · Codex · Cursor · Gemini CLI

Editorial guidance. Tested environments and upstream support are recorded separately in the evidence.
Read original skill Inspect the run evidence Source owned and maintained by Jesse Vincent. Our reviews remain independent.