10 genuinely useful AI-agent skills every week, generalized from my real AI-agent workflow so anyone can put them to work.
Feed last updated Aug 10, 2026
How to use these
Clone the repo, or copy a single skill folder into ~/.claude/skills/, then ask Claude Code (or your agent of choice) to use it.
A four-label vocabulary (VERIFIED, CODE-SHIPPED-NOT-VERIFIED, BLOCKED, INCONCLUSIVE) that every completion claim from an AI agent must carry, each backed by the exact command or probe that justifies it.
Use it for: Making 'done' impossible to fake in agent status reports, and turning 'are you sure?' arguments into evidence lookups.
View on GitHubA fixed five-section template for summarizing long AI-agent sessions that quotes exact strings (paths, errors, flags) instead of paraphrasing them.
Use it for: Long agentic sessions that get compacted or handed off, so the next session resumes without re-deriving state.
View on GitHubA deep-planning protocol: parallel research agents map current state and blast radius, competing designs are drafted, then a mandatory adversarial pass attacks the plan before any implementation.
Use it for: Non-trivial features, refactors, migrations, and anything handed to an autonomous agent to execute unattended.
View on GitHubStable role-to-model-tier lanes — the strongest model plans and verifies, a cheaper capable model executes, a different-vendor long-context model judges — with live-pricing re-verification before cost-based decisions.
Use it for: Designing multi-model pipelines, monthly cost reviews, and re-mapping lanes when a new model generation ships.
View on GitHubExplicitly pin every spawned subagent to the cheapest adequate model tier, closing the silent-inheritance leak where parallel readers run on the apex-priced main-loop model.
Use it for: Anything that fans out parallel subagents, and investigating a surprising AI bill.
View on GitHubKeep multiple AI coding CLIs behaviorally consistent with one shared constitution file, a capability routing table with polyfill recipes, and a compiled policy block injected identically into each tool's config.
Use it for: The day you add a second AI coding CLI, or after two tools behave differently on the same rule.
View on GitHubThe twelve standing background agents worth running on any machine hosting unattended AI automation, plus the receipt-file and whitelisted-autofix discipline that keeps the roster honest.
Use it for: Growing a scheduled-automation estate without silent multi-week outages.
View on GitHubA git-backed persistent memory layout for AI agents — one fact per file with typed frontmatter, an always-loaded index, and a scheduled commit/rebase/push sync job.
Use it for: Making what an agent learned survive session boundaries, machine switches, and agent-tool switches.
View on GitHubWire an external chat platform (Telegram, Discord, SMS, etc.) into an AI agent so you can send it a prompt from your phone and get the reply back in the same chat, with untrusted-input quarantining and sender allow-listing built in.
Use it for: Sending your agent a prompt from your phone while away from your desk, or building a lightweight remote control for an unattended agent without exposing a web endpoint.
View on GitHubContinuously score live production conversation/interaction transcripts against a quality rubric, auto-draft a minimal prompt-only fix as a pull request, and auto-merge it only once it clears a strict multi-reviewer safety gate.
Use it for: A deployed conversational or text-generating product where you want quality regressions caught from real traffic and small prompt fixes shipped automatically, never auto-merging an unreviewed or oversized change.
View on GitHubA pre-flight routine that frees RAM and drops memory pressure on your machine so it can safely run a larger batch of parallel AI-agent sessions than usual, then restores normal state afterward.
Use it for: Right before spawning a larger-than-normal parallel AI-agent fan-out, or after noticing a big parallel run swapping/thrashing instead of speeding up.
View on GitHubA read-only, dependency-free script that checks a layered codebase against documented architecture rules (layer crossings, legacy reuse, anti-pattern regressions, doc drift) and fails CI when they're violated.
Use it for: Enforcing a layered architecture boundary, tracking a strangler-pattern legacy file's shrink, or catching a previously-fixed anti-pattern before it regresses elsewhere.
View on GitHubCompares a code-level policy module against the natural-language prompt it governs, flags missing anchors and direct ALLOW/FORBID contradictions, and emits a per-policy ALIGNED/WARN/DRIFT verdict without editing anything.
Use it for: Before merging a PR that touches a policy/constants module shared with a prompt file, or after any prompt rewrite.
View on GitHubPre-push check that verifies the active git author email, authenticated gh CLI account, and target remote all match what you declared for this repo, and refuses to push on any mismatch or missing declaration.
Use it for: Working across a personal and a work GitHub account (or several client accounts) on the same machine.
View on GitHubEnumerate every surface tied to a dead project (repos, scheduled jobs, dashboards, notes), get human confirmation, then archive everything reversibly with an undo manifest instead of deleting it.
Use it for: Cleaning up a finished or abandoned project without hunting down every cron job and dashboard entry by hand.
View on GitHubPull a structured, evidence-backed recap (shipped / in-flight / at-risk) of work across multiple local git repos over a time window, cross-referenced with personal notes.
Use it for: Answering 'what did you actually get done this week' across several repos without reconstructing it from memory.
View on GitHubBuild a fully-local, zero-cost observability bundle for an AI coding agent — a status line, a hallucination-blocking turn-end hook, and an optional periodic quality grader.
Use it for: Getting visibility into context usage, cost, and unverified 'done' claims from your AI coding agent without a paid monitoring service.
View on GitHubWraps OpenAI gpt-image-1, Fal's Flux Pro, and Google Imagen text-to-image APIs behind one small script with a bake-off comparison mode, backend health tracking, and a reusable prompt template.
Use it for: Generating hero art or concept backdrops when no image-generation MCP/tool is available, or comparing providers before picking a default.
View on GitHubBefore letting a secondary agent, worker fleet, or automation pipeline act on your behalf, require a proven track record plus deterministic post-hoc verification — never trust the delegate's own self-report of success.
Use it for: Wiring up a second agent or automation pipeline that will act without your eyes on every step.
View on GitHubRecover full-quality output when a safety classifier misreads legitimate, authorized defensive-security work as an attack because of how the request is worded, without ever using this to evade a genuine restriction.
Use it for: An authorized defensive-security task comes back visibly weaker or refused because it was worded in attack-shaped language.
View on GitHubRuns a scheduled and on-demand battery of deterministic, no-LLM checks over a local dev/automation setup and surfaces a ranked, evidence-backed finding list with concrete fixes and an autofix-safe flag.
Use it for: Catching doc-vs-code drift, orphan jobs, security-default drift, and committed secrets before they turn into a real incident.
View on GitHubWhen writing FORBIDDEN/NEVER/✗ example blocks in an LLM prompt, never spell out the literal forbidden phrase — models pattern-match into the bad example and reproduce it; use a short abstract marker instead.
Use it for: Prompts where the model keeps producing an output you explicitly told it never to produce.
View on GitHubWhen a guard/gate/precondition depends on data that might be missing, make the missing-data fallback REFUSE the action, not allow it — and confirm every write-site actually populates the field the guard reads.
Use it for: Writing or reviewing any permission check, gate, or precondition with a fallback default.
View on GitHubGet a Twilio 10DLC A2P SMS campaign approved, or un-stick a rejected one — decodes the common rejection codes and covers the non-obvious resubmission mechanics.
Use it for: A FAILED/rejected Twilio A2P campaign, or building an SMS opt-in flow before submitting one.
View on GitHubContinuously lint your own AI-agent tooling (skills, subagent definitions, workflow scripts, hooks) for the defect classes that quietly rot an agent estate — leaked tokens, retired model IDs, unpinned models, gate docs with no real enforcement.
Use it for: A pre-flight check after writing/editing a skill or agent, or a periodic cleanup pass over your whole skills directory.
View on GitHubMulti-round adversarial questioning that forces you to defend a design BEFORE you build it — generates hard questions across failure modes, edge cases, security, and performance, then iterates on your answers.
Use it for: Stress-testing a plan or architecture decision before you commit to building it.
View on GitHubA pre-flight linter for the highest-signal 'AI-slop' tells in generated frontend/marketing copy — em-dashes, numbered eyebrows, BETA labels, generic filler phrases, decorative status dots.
Use it for: A fast last check after generating any landing page, hero section, or marketing UI.
View on GitHubTurn a set of scores/metrics into an accurate AND illustrative graphic — the data literally drives the shape of the artwork, so it reads as a designed illustration instead of a generic chart.
Use it for: Visualizing scores/metrics/dimensions when a plain bar/line/pie chart would be boring or off-brand.
View on GitHubA curated checklist of ~18 evidence-grounded rules for building reliable AI agents — tool-use gating, verification-before-trust, context engineering, and safety/trust boundaries.
Use it for: Designing a new agent/workflow, or reviewing an existing one for reliability gaps before more autonomous use.
View on GitHubIteratively grade and improve a conversational AI agent's prompt quality without real end-users — synthetic dialogs, hard checks + an LLM judge, targeted human-approved prompt edits until quality plateaus.
Use it for: After any meaningful change to a conversational agent's system prompt, or as a regression gate before a demo/release.
View on GitHubA hardened, gated pipeline for producing short-form AI-generated 'UGC-style' marketing video — hard-won fixes for the media-validation, audio-sync, and judge-hallucination traps that silently ship broken output.
Use it for: Producing or re-cutting a short-form AI-presenter product/marketing video.
View on GitHubSpawn N domain-expert subagents in parallel to audit a codebase, product, or design from different angles, then synthesize convergent vs. divergent recommendations.
Use it for: Breaking out of a surgical-patch loop, getting a second opinion before a risky change, comparing two architecture options.
View on GitHubProcess inline review-comment markers left in shared Markdown docs, apply edits in place, preserve history, and flag contradictions instead of guessing.
Use it for: Async doc review between collaborators without a full comment/suggestion system.
View on GitHubTurn every fixed bug into a permanent regression fixture and replay the whole set before you ship anything.
Use it for: Catching regressions on bugs you already fixed, building a growing regression suite from real incidents.
View on GitHubRender any finished deliverable (report, brief, plan, design, mockup) as a self-contained hosted page and hand over the link instead of a terminal dump.
Use it for: Reports, briefs, audits, roadmaps, and design/mockup previews meant to be read, reviewed, or shared.
View on GitHubRoute self-contained work off your primary AI assistant's metered quota onto a secondary backend you already have, then verify the result yourself.
Use it for: Being low on quota, parallelizing independent work, or getting an independent second opinion.
View on GitHubMap every 'done' claim to the specific proof required before you say it out loud.
Use it for: Preventing confidently reporting unverified tests, builds, deploys, or fixes as complete.
View on GitHubA maximum-depth, multi-agent code review that auto-locates the diff, fans out a cross-functional panel, adversarially verifies findings, and synthesizes a prioritized risk register.
Use it for: Pre-merge high-stakes diffs and production-readiness audits.
View on GitHubA looped, dual-verification stress-test and KPI gate to run before calling any non-trivial change 'done'.
Use it for: Wanting genuine walk-away quality on anything non-trivial you've built, fixed, or deployed.
View on GitHubAn end-to-end, self-verifying build pipeline (plan, swarm, docs, QA, code quality, canary, summary) gated stage-by-stage on verification artifacts.
Use it for: A defined-end-state build you want to walk away from and trust came back correct.
View on GitHubMine your own AI-assistant session transcripts for repeated workflow patterns and draft candidate reusable skills, queued for human approval.
Use it for: Discovering which of your repeated multi-step workflows would make good reusable skills.
View on GitHub