Claude Skills by KyleMit

Triage and address every reviewer comment on a pull request — validate each one against the current code, fix the valid ones and reply with the solution, reply to and resolve the invalid ones with the rationale. Use when asked to address, respond to, or work through the comments/feedback/review received on a PR. To produce a review of a PR (author the critique), use leave-pr-review instead.

Index of all architectural decision records (ADRs). Use before proposing, discussing, or questioning any architectural approach, or when asked why something is built the way it is.

Mine local Codex CLI and Desktop rollout transcripts, including archived sessions, into factual evidence-anchored reports. Use only when the user explicitly asks to analyze, debrief, or report on past Codex sessions; Claude Code transcripts use an independent provider package.

HTTP API reference for the /api/* endpoints — generate-image, report-image, verify-access-code, verify-key, report, csp-report, and the admin bearer-session endpoints, plus the CORS, rate-limiting, and auth model. Use before adding, changing, or calling any /api endpoint.

Splotch tech stack, file-by-file source map of web/src/, route table, and the canonical UI element glossary. Use when navigating unfamiliar parts of the codebase, deciding where new code belongs, or needing the proper name of a UI element.

Comprehensive repository pass that produces a prioritized list of performance, readability, maintainability, and architecture improvements in docs/AUDIT.md. Use when asked to audit the codebase, hunt for improvement opportunities, or do a broad code-quality sweep across the repo.

Review the supported browser and OS floor against current usage and Baseline data, check the API risk register in docs/COMPATIBILITY.md against the code, and classify every runtime feature probe in shipped web/src, then file what drifted as findings in docs/AUDIT.md. Use when asked to audit browser compatibility, revisit or raise the browser/OS floor, check the compatibility risk register, or find fallbacks and feature checks a floor makes unnecessary. Changes no code; a floor raise becomes a...

Inventory and evaluate every third-party dependency — who publishes it, its license, stars, release cadence, maintenance health, abandonment risk, and whether to keep or replace it — written to docs/DEPENDENCIES.md. Use when asked to inventory or catalog dependencies, assess dependency health or provenance, review maintenance/abandonment risk, find dependencies worth replacing, or produce an SBOM-style dependency review. It analyzes dependencies but never upgrades them — to apply version upda...

Scan the codebase for inline code blocks worth extracting into standalone, named, testable functions, and file them as findings in docs/AUDIT.md. Use when asked to find extraction/refactor-into-function candidates or to improve readability by naming inline logic.

Run a Lighthouse page-load audit of the Splotch web app emulated on a slow device + slow internet (phone portrait + tablet landscape), capturing first-visit and repeat-visit runs, and turn the results into audit opportunities. Use when asked to audit page-load/Lighthouse performance, measure load on a throttled device/network, check Core Web Vitals (FCP/LCP/TBT/CLS), or compare first vs repeat visit. For drawing/canvas interaction performance use the `profiling` skill instead.

Retrospective on the current session that files durable, recurring friction (finding, understanding, running, or changing code) as audit findings in docs/AUDIT.md so the repo's Claude tooling gets sharper over time. Use near the end of a working session, before /clear, or when asked to reflect on session friction.

Build the signed release artifacts for the current version (Android .aab, iOS .ipa)

Drive the scripted bulk burndown of docs/AUDIT.md with isolated Codex subprocesses per role and finding (verify → implement → adversarial review → fix). Use when the staged audit backlog is too large to vet and file as individual GitHub issues, or when asked to launch, resume, supervise, pause, report on, or close out an audit burndown from Codex.

Run an unattended code-quality campaign — audit the codebase principle by principle and then area by area with parallel read-only auditors, cluster the verified findings into small units, and ship each unit as its own rival-reviewed, CI-green, merged PR through ship-issue, several disjoint units at a time, until every finding is merged, rejected with a reason, or parked for the user; then fold what the run learned into docs/CODING-STANDARDS.md. Use when asked to burn down code smells, improve...

Take the open Dependabot PRs to a decided end state — verify each bump against upstream source and this repo's actual usage, then merge the safe ones in a conflict-aware order, and close the unsafe ones behind a tracking issue. Use when asked to review, triage, merge, or clean up the Dependabot / dependency-update PRs, or to confirm a batch of dependency bumps is safe to merge.

Drive Splotch's masked Playwright retry count toward zero using the rolling flaky digest, fixes verified under equal worker load, and narrowly tracked quarantine when a confirmed flake cannot be fixed. Use for a periodic E2E flake burndown, not for an ordinary single failing test.

Check package.json dependencies and GitHub Actions pins for updates (incl. majors), then upgrade them one at a time — read the migration guide, fix all usage, verify, and commit each on its own

Pay the size caps back down to their soft targets — mode=files for the per-file max-lines cap, mode=functions for max-lines-per-function. Measures every covered unit against the cap ESLint resolves for it, has a proposer and an adversarial reviewer decide split-or-raise per unit, implements each split in its own worktree behind an independent commit check, retires stale per-file caps, and ships one themed PR with one commit per unit through drive-pr-to-mergeable. Use when asked to burn down, ...

Capture Splotch's drawing, undo, and discrete-action performance suites across macOS web, physical or simulated iPad web/native, and physical or emulated Android web/native targets. Use when producing a deployment-target performance snapshot, comparing renderer architectures, validating a performance change across platforms, or refreshing the committed performance matrix.

Claude Code Cloud session only — start the dev server + reverse tunnel and report the public phone-preview URL

Document a new architectural decision as an ADR in docs/adrs/. Use when a significant decision was just made or confirmed — one that chose an approach over real alternatives, has non-obvious consequences, or encodes a constraint a future contributor would want to understand.

Hand in-flight work to another session — by default as a copyable continuation prompt rendered in one fenced code block, or as a transfer packet committed to docs/handoff/ when the user asks for a file. Use when wrapping up before /clear, running low on context, or when the user says to "hand this off", "write a handoff", "give me a prompt for the next session", or pause work for later. To pick a handoff back up, use resume-handoff instead.

Build a copy-paste prompt directing another agent to review every PR this session produced with the leave-pr-review skill — enumerating each PR (a stacked campaign has several), then adding the spots that deserve special attention and an adversarial pass on the overall strategy as extra focus areas layered on top of a full review sweep. Use when asked to create a PR review prompt, a PR review or feedback handoff, to hand this session's PRs to an independent reviewer, or to prep a second-opini...

Ship dependent changes as a stack of sequential pull requests — each PR based on the branch below it — then link the chain into a GitHub stack, judge CI at the tip, and merge the whole chain as one unit when asked. Opt-in; a multi-issue campaign merges as it goes through ship-campaign instead. Use when the user asks for stacked or sequential PRs, when a multi-part change must land together, or when an existing chain of open PRs needs linking, verifying, or merging.

Review every light/night screenshot in Splotch's responsive page inventory through fresh image-only reviewers, checkpoint each review ID, and finalize a complete hash-bound design critique. Use when asked to critique, assess, finish, resume, or regenerate page-inventory feedback.

Draft and review release notes, then bump versions, commit, tag, push, and create the GitHub Release

Design-system reference — the token vocabulary (color, spacing, radius, type, elevation, motion), the primitives in lib/components/design/, the voice & copy rules, brand/iconography, and the rules for styling UI. Use before writing or changing any component styles, picking a color/size/shadow/easing, adding a UI element, writing user-facing copy, or when visual consistency or the /design styleguide comes up.

Take one open pull request from opened to mergeable — build the reviewer prompt, get an independent rival-agent review, address every thread, repeat that round at most once more, drive CI to green, reconcile conflicts, and return a shippable-or-leftovers verdict. Never merges. Use whenever an open PR needs the repo's standard independent-review and CI loop, standalone or as the current tip of a stack; the skills that open PRs (ship-issue, fix-audits, create-stacked-prs, improve-performance-ma...

Enumerate an epic's children from the GitHub sub-issues API — never from the epic's prose — classify each (done, dropped, in flight, done-but-open, actionable, blocked, needs triage), read every child's comment thread for retractions and scope changes, and recommend a working order with every bucket reconciled against the API counts. Use when asked what remains on an epic or tracking issue, to inventory or plan a multi-issue campaign, or before implementing any issue that belongs to an epic.

Work through every open type:audit GitHub issue autonomously on a dedicated branch — one commit per issue — validating, fixing, and verifying each. Use when asked to fix, clear, or work through the audit backlog.

Review the Claude Code configuration and session history, benchmark against current best practice, and recommend and apply updates for an integrated, token-efficient setup. Use when asked to audit or improve the Claude Code workflow, config, permissions, or token efficiency, or to reduce repetitive permission prompts.

Drive Splotch's deployment-target performance matrix from current evidence to zero unexplained scoreable red cells on the release-gate rows through product improvements and faithful recaptures, keeping harness work subordinate to and immediately useful for a named product experiment. Ships each causal product cluster as its own reviewed PR, merged before the next cluster begins; an unattended run goes through ship-campaign profile=performance. Use for sustained performance improvement; use ca...

Splotch conventions that augment (not replace) the built-in review flow when reviewing a pull request — check out the PR branch locally for offline diffs and to run the code and empirically verify critiques, anchor every finding to a diff file + line as you analyze, then post the surviving findings onto the PR as a single inline-comment review; invoking the skill is itself the authorization to post publicly, and mode overrides instead keep the findings in chat, file them as GitHub issues (eac...

Capacitor native app guide — Android/iOS toolchain setup (macOS + Linux), build/sign/run commands, on-device testing, Chrome remote profiling, and the store release & kids-compliance checklists. Use before touching anything Android, iOS, or Capacitor related.

Coordinate a multi-issue effort through human-relayed Claude or Codex worker sessions — prioritize batches, emit one copyable worker prompt at a time, and verify pasted reports and preservation of evidence before choosing a follow-up or the next batch. Use when asked to orchestrate separate sessions or manage their prompts and reports; ship-campaign runs an unattended queue inside one worker session.

Custom Splotch conventions for visuals in a pull request body — always include screenshots, before/after tables for changes, all states for multi-state components, and gifs/video for animations. Use in addition to the built-in PR flow whenever opening, creating, or updating a pull request that touches anything visible in the UI.

Capture and read an automated performance profile of the drawing app (web, Android, iOS). Use when measuring drawing/canvas performance, investigating jank or a slow interaction, verifying a perf change, or checking for regressions over time. Covers the `npm run perf:*` harness, how to read report.md/summary.json, and the bottleneck decision guide.

Clean up the git workspace in one pass — salvage evidence from and prune stale agent worktrees, delete provably-dead local branches by script, run a judgment pass over the unmerged remainder that makes any finding durable before its branch goes, and triage stale remote branches on origin into a deletion script for the user to run. Use when asked to clean up, prune, or delete old branches or worktrees, when a checkout has hundreds of stale local branches, or after a run of agent sessions has l...

Attach the built .aab/.ipa to the GitHub Release for a version, verifying each artifact's embedded version first. Use after the cut-release and build skills, or whenever a GitHub Release is missing its binaries or has the wrong ones attached.

Review recent decisions from the current conversation and git history, verify existing ADRs still reflect reality, and add or amend ADRs as needed. Use at the end of a session that touched architecture, testing, infrastructure, or build tooling, or when asked to refresh or reconcile the ADRs.

Refresh docs/CODE-MAP.md, the lines-of-code-by-area snapshot of the repository — regenerate its tables with `npm run gen:code-map`, teach the counting rules about new areas and misplaced files, rewrite the prose to match the new numbers, and flag areas whose growth deserves an issue. Use when asked to refresh, update, or reconcile the code map, before an exhaustive whole-repo audit-code pass that shards by map area, or when the map's snapshot commit is weeks behind main.

Bring the living docs, skills, and rules back in line with the code — review the identifiers `npm run check:doc-refs -- --identifiers` cannot resolve, hunt the prose that still describes renamed or deleted files and symbols by name, re-justify every doc-reference exemption, and fix each mismatch. Use when asked to reconcile docs with code, find stale or dead references in docs, check that docs still match the code, or after a large rename, provider swap, or deletion lands. A dead repo path or...

Merge the latest main into a long-running branch and check the incoming commits for semantic conflicts — the ones that merge cleanly and still break the branch — not just git conflicts. Use when refreshing, syncing, catching up, or updating a branch that has been open a while against a main that has moved.

Release the physical iPad and Android capture rig — stop every capture-rig process any checkout of this repo owns (previews, probe hosts, iPad secure-origin fronts, Appium, the hold-awake watcher, inspector proxies, from this session or a previous one), drop the adb forwards they left, and put the phone back to its stock state, so the next start-capture-session begins from a clean rig. Use when asked to release, tear down, clean up, or reset the capture session or capture ports, to free the r...

Pick up in-flight work from a transfer packet in docs/handoff/. Verifies the packet against the current repo, reports what's stale/missing, deletes the consumed handoff, then continues the work. Use when the user says to "resume", "pick up the handoff", "continue where we left off", or names a handoff to resume. To write a handoff, use create-handoff instead.

Pair this Codex session, as the native handler, with a rival agent — a fresh local Claude Code process on the Claude plan login — for an independent review of a PR, a diff, or a free-form question about this checkout. The rival reads a disposable worktree and asks you to run commands through a broker; you run each one under Codex's own sandbox and policy or decline; its findings post to the PR verbatim through a fixed wrapper. Use when the user asks for an outside/independent/second-opinion r...

Launch, run, and screenshot the Splotch web app to confirm a change works in the real running app (not just tests). Use when asked to run/start Splotch, open it in a browser, take a screenshot, draw on the canvas, or visually verify a UI change. Covers the / drawing app, /admin, and /privacy routes. For Android/iOS native builds use the `mobile` skill instead.

Sweep the session just worked for hiccups, surprises, workarounds, and hard-won lessons, judge which are durable, and write each into the home the next tripped-up session will actually see — a harness fix, code comment, rename, script change, doc or skill edit, ADR, GitHub issue comment, PR description, or scratchpad note — applied now rather than staged. Use when asked to self heal, when wrapping up a session that hit friction, or whenever a lesson is about to be lost because it lives only i...

Ship a queue of GitHub issues unattended — an explicit list, an epic's sub-issues, or the newest unclaimed backlog issues — one at a time through ship-issue mode=autonomous, merging successful PRs before the next starts from fresh main. Proves everything that could stall an overnight run while the user is still present, quarantines a stuck issue instead of stalling the queue, and ends with a verified morning report. Use when asked to run a campaign, work through several issues or an epic over...

Take one GitHub issue number or a free-form task all the way to a review-complete pull request — implement it, open the PR, get an independent rival-agent review, address the feedback, and repeat that review round at most twice, driving CI to green and the PR to mergeable. Stops there with the PR open by default; `mode=autonomous` also merges it. Use when asked to ship an issue, take an issue end to end, implement and get a PR reviewed, or drive one task from description to a PR that has surv...