My personal collection of agent skills for coding, planning, knowledge capture, transcription, slides, multi-agent work, and day-to-day setup.
This is not a warehouse of every Skill I could install. It is the smaller set I actually reach for: useful enough to earn a place, compact enough to browse. π§°
npx skills add patrick-fu/awesome-skills -gUpdate later:
npx skills update -gnpx skills add patrick-fu/awesome-skills -g -s <skill-name>frontend-harness-slides, parallel-goal-workflows, and llm-wiki-capture
also have standalone repositories, linked from their entries below. Installing
one of them from its standalone repository in the same global scope replaces
the matching Skill; it does not create a second copy.
ποΈ frontend-harness-slides
Build HTML slide decks with a frontend harness: stable scene URLs, a registry, a fixed stage, tests, visual checks, and verified delivery. Repeated edits stay manageable instead of quietly breaking earlier slides.
Use it when a deck needs to look good, feel alive, and ship online, as a PDF or static export, or in all three forms.
π₯οΈ Live demo: explore the dynamic Workbench Demo. The styles show the range; the real win is harness-backed iteration instead of fragile single-file HTML.
π¦ Standalone repository: patrick-fu/frontend-harness-slides
ποΈ llm-wiki-capture
Capture session lessons and external sources into a maintained, Git-backed LLM Wiki. It can also audit an existing Wiki or bootstrap one from zero while preserving provenance, canonical ownership, and explicit repository boundaries.
Use it when knowledge should survive the current session without handing the Wiki's Git history over to opaque automation.
π¦ Standalone repository: patrick-fu/llm-wiki-capture
ποΈ x-twitter-reader
Retrieve X/Twitter posts, reply threads, long-form Articles, metadata, links, and media before summarizing, translating, quoting, or archiving them. Source acquisition stays faithful; downstream transformation happens afterward.
Use it when an X page requires login, exposes only a preview, or simply is not a reliable source of the complete content.
Requires uv and Python 3.10+. Its pinned twitter-cli dependency may reuse a
local X session after approval; the Jina fallback is opt-in.
π say-that-again
A user-invoked reset for explanations that did not land. It adds only the necessary context, keeps established terminology, and explains the point again in clear, concise, unambiguous Chinese.
Use it when you did not follow the previous response and want a clearer Chinese explanation rather than a lasting response style.
π§Ή deslop
Review or rewrite Chinese, English, and mixed-language prose to remove formulaic AI patterns while preserving facts, intent, stance, register, and the author's voice.
Use it on existing text that sounds templated or over-producedβnot as a generic trigger for every writing task.
Give each top-level goal one Goal Owner and a clean, task-local brief. The Owner handles execution, focused helpers when useful, review, repair, verification, and the final report without flooding the main conversation.
Use it for audits, complex research, repair loops, or any task where independent checks matter.
π¦ Standalone repository: patrick-fu/parallel-goal-workflows
π― long-task-control
Run an adversarial control loop over a long task: at each phase boundary a Design Challenger and Drift Watchdog test ROI, scope creep, and drift against the original intent. It preserves verified work, retires activity that no longer serves the goal, and resumes only where new acceptance evidence appears.
Use it when a task is genuinely driftingβnot merely because it is long.
ποΈ codex-session-controller
Turn a top-level Codex App task or session into a controller session: one user entry point backed by a thin control plane. It keeps goals, authorization, owners, dependencies, waits, and reporting cadence in the control plan, then handles routing, deduplication, acceptance, and handoff while workers own research, implementation, and testing. Users normally stay with the controller instead of switching among workers.
π·οΈ codex-session-naming
Maintain lifecycle titles for user-visible top-level Codex App tasks and sessions, keeping state, next gate and owner, checkpoint, final, and handoff visible at a glance. A strict closure contract prevents checkpoints from being mislabeled as final completion. It defines title and closure semantics; it does not dispatch work.
These four coding-agent guides share one workflow: preserve the selected CLI wrapper, discover current model and reasoning controls, monitor semantic output, respect permissions and sandboxes, and verify diffs and tests from the host. Each guide triggers only when its CLI is explicitly selected as the executor.
Run Claude Code CLI as the explicitly selected external coding executor. It preserves the supplied wrapper, discovers current options from the CLI, and monitors long-running work through verbose streaming JSON.
π’ codex-coding-agent
Run Codex CLI as the explicitly selected external coding executor. It uses
codex exec --json with an explicit read-only sandbox for review or a
workspace-write sandbox for approved implementation.
π΅ cursor-coding-agent
Run Cursor CLI as the explicitly selected external coding executor. Because the
generic agent launcher can point to another product, the workflow verifies its
identity before trusting it.
Run Grok Build CLI as the explicitly selected external coding executor. It treats approval policy and sandbox access as separate controls and chooses the appropriate read-only or workspace sandbox explicitly.
π‘ brainstorm
Expand an idea or open problem into a possibility map spanning genuinely different dimensions before deciding where to go deeper.
Use it while the solution space is still open and breadth is more valuable than an early verdict.
πΊοΈ explore-and-plan
Turn a mostly chosen direction into an executable plan: surface blocking decisions, map affected areas, order steps by dependency, and tie verification to each outcome.
Use it when the direction is settled enough that an implementer should not have to rediscover the intent.
Teach through short, active learning loops built around practice, teach-back, and review instead of defaulting to a long explanation. Each loop adapts to the learner's goal, current evidence, and available time.
Use it when the goal is understanding and retention, not merely receiving an answer.
π§ͺ write-unit-test
Write, review, or improve unit tests around behavior, regressions, and domain rules. It starts from a caller-visible contract, keeps decisive inputs visible, and asserts observable effects.
Use it for production code where breaking the contract should turn the test red while harmless internal refactoring stays green.
π¦ log-driven-debugging
Add focused, prefixed logs, have the user rerun the real scenario, and inspect the returned evidence to locate the first meaningful divergence. When static reading is not enough, observability replaces guesswork.
Use it for reproducible bugs whose root cause is still hiding.
Draft a concise English commit message with a sentence-case subject, no conventional prefix, and optional bullets for the important changes.
Use it when the change is ready but the copy-ready commit message is not.
π home-config-sync
Maintain a personal ~/.dotfiles bare-repository workflow: first-time setup,
new-machine deployment, safe pulls and merges, and pushes. Its work tree is the
whole home directory, so the guardrails are part of the feature.
Use it when dotfiles should be versioned without letting a convenience script treat the home directory casually.
ποΈ local-file-organizer
Inventory cluttered directories file by file, classify content by its real use, and build a reversible ledger before moving, renaming, or deduplicating anything.
Use it when a directory needs a safe, explainable reorganization rather than a blind cleanup.
ποΈ audio-transcription
Transcribe local audio and video files with two independent local engines: Qwen3-ASR for speech recognition and MOSS for diarized transcription. The default flow is fully offline, keeps each model's result separate for cross-checking, and never uploads your media.
Requires macOS on Apple Silicon (arm64). First use downloads pinned, versioned models into an app-managed runtime; transcription itself makes no network calls. An optional Doubao (Volcengine) worker transcribes through the cloud, but only after you explicitly ask for online or upload-based transcription, and it uploads a converted copy of the media.
Use it when you explicitly ask an agent to transcribe or extract a transcript
from a local audio or video file. When the agent reports SETUP_REQUIRED,
approve the one-time local runtime setup yourself.
MIT covers this skill's own code. Third-party models and the Doubao API keep
their own licenses and terms; see NOTICE.md.
MIT. See LICENSE. A nested LICENSE or third-party notice takes
precedence for the files it covers.