skills/retry-fix/SKILL.md
Investigate recurring bugs and regression cycles before attempting another fix. Use when: (1) User says 'still broken', 'X is blue again', 'this is the second/third/fourth time', 'we tried this before', 'I'm doing this many times', 'the regression is back', 'didn't stick', (2) A bug has been 'fixed' before but came back, (3) User mentions doing the same fix again, (4) Code changed for a feature but the previous fix got overwritten. ALWAYS use this skill before writing any code when a regression is detected — surface the prior-attempts table and get explicit confirmation before proceeding.
npx skillsauth add takazudo/claude-resources retry-fixInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
A recurring bug has been patched before and regressed. This skill excavates all prior attempts, diagnoses the structural pattern that lets it regress, and gates any new code change behind explicit user confirmation.
Argument: $ARGUMENTS — a problem slug (e.g. active-frame-border, zoom-transform, modal-focus).
Run these searches in parallel. Use the slug and obvious keyword variants (kebab → spaces, camelCase, partial words):
# Git commits
git log --all --oneline -i --grep="<slug-keyword-1>"
git log --all --oneline -i --grep="<slug-keyword-2>" # variant if needed
# Closed GitHub issues
gh issue list --state closed --search "<slug keywords>" --limit 30
For each result: record SHA (or issue number), date, and title. Deduplicate overlapping hits.
For each unique commit SHA found:
git show <sha> --stat # what files changed
git show <sha> # full diff (skim for the structural change)
For each closed issue number: read the issue body (use /gh-fetch-issue if it has attached images). Capture:
Look for a project-scoped lessons skill matching the slug:
ls .claude/skills/ | grep "l-lessons"
If .claude/skills/l-lessons-<slug>/SKILL.md exists:
last-updated: frontmatter fieldProduce the following sections in this exact order. Do not skip sections. Do not merge sections.
| # | Date | Commit SHA | Issue | What it changed | What it missed |
|---|------|-----------|-------|-----------------|----------------|
| 1 | YYYY-MM-DD | abc1234 | #N | One-line description | What was left untouched |
| 2 | … | … | … | … | … |
One row per distinct fix attempt. If a commit and an issue describe the same attempt, merge into one row.
One paragraph per attempt (numbered to match the table). Name the specific assumption that was wrong — not the symptom. Example: "Attempt 1 changed the JS default but not the CSS fallback, assuming CSS consumed the JS value at runtime. In fact the CSS fallback fires before JS runs, so the old palette index was visible until settings loaded."
Answer this question: What is the architectural pattern that lets this bug regress?
Concrete examples of structural diagnoses:
One or two paragraphs. Specific over general.
Bulleted requirements the next attempt MUST include to prevent regression:
Add or remove bullets as appropriate. Tailor to the actual structural diagnosis.
Run
/retro-notes <slug>Capture in
l-lessons-<slug>: [one-sentence summary of what to record — the structural diagnosis above, the guard test requirement, and any traps from the checklist].
After emitting the report, do not propose or write any code. Instead, ask:
"I've surfaced all prior attempts above. Before I propose the next fix, please confirm: 'Yes, I've seen this and want a structurally different next attempt.'"
Wait for explicit confirmation before proceeding. If the user confirms, default to invoking /big-plan (or describe the next attempt as a structured plan) — do not write code directly unless the user explicitly says so.
When the user is ready to close out this session, or after the fix lands:
/retro-notes <slug> to append a dated section to .claude/skills/l-lessons-<slug>/SKILL.md.last-updated: YYYY-MM-DD frontmatter field in the lessons file (using today's date)./retro-notes will scaffold one — confirm the last-updated field is present after it runs./big-plan or a follow-up task.active-frame-border, active frame border, activeFrameBorder, frame border) to avoid missing commits that used a different naming style.Reopened events or a follow-up issue referencing it, include both in the table as separate rows.$HOME not ~ in file paths — ~ is not expanded by non-login shells or Node.js.development
Link Claude Code skill names mentioned in a CodeGrid article (data/{series}/{n}.md) to the author's public claude-resources repo, pinned to the latest commit hash so links don't rot. Use when: (1) user says 'linkify cc resources', 'link the skills', 'link skill names', or invokes /dev-linkify-cc-resources; (2) editing a CodeGrid article that mentions `/commits`, `/pr-complete`, `/skill-creator` or other Claude Code skills and they should point to claude-resources. Only links skills that actually exist in the public repo; skips hypothetical examples and code blocks.
development
Second opinion from Claude Opus on a plan or approach. Use when: (1) Planning phase of /big-plan needs a higher-quality review than /codex-2nd / /gco-2nd / /gcoc-2nd, (2) User says 'opus 2nd' or 'opus opinion', (3) Wanting Anthropic's larger model to critique a plan. Spawns a general-purpose Agent with model: opus that reads the plan file and returns structured feedback. Anthropic quota — not free.
tools
AI-based testing via subagent + a per-task test-flow skill. Use when the user wants to verify something that mechanical assertions can't fully capture — image recognition, visual size/position comparison, animation smoothness, multi-step manual flows that need AI judgment. Triggers: 'AI-based test', 'AI test', 'visual verify', 'image recognition test', 'manual operation test', 'human-eye check', 'verify visually', 'compare screenshots', 'looks the same', 'looks correct'. The skill's job is to (1) author a focused test-flow skill that captures the exact procedure + verdict criteria, then (2) dispatch a verification subagent via the Agent tool that loads BOTH the test-flow skill AND a browser-driving skill (/verify-ui primary, /headless-browser fallback) so the subagent has clear context and consistent verdicts. NEVER uses `claude -p` — subagent dispatch goes through the Agent tool exclusively.
development
End-of-workflow audit of touched GitHub issues, PRs, and branches via a Sonnet subagent. Use when: (1) /big-plan, /x-as-pr, or /x-wt-teams finishes its main work and needs to verify every touched resource is in the right state (closed when done, kept when ongoing, deleted when dead), (2) User says 'cleanup resources', 'audit cleanup', or 'check what should be closed', (3) A long workflow ends and the manager wants a structured paper trail of what it closed/kept/deleted. Auto-execute by default — the Sonnet agent proposes, the manager (you) executes safe actions and prints a final report.