Maintain README.md, AGENTS.md/CLAUDE.md, skills, and repo context. Create, scaffold, or initialize AGENTS.md files and project-local agent skills. Use task evidence to delete obsolete skills, merge overlapping skills, and capture recurring workflows.
Paul Berg
Software engineer and open-source developer with 12+ years of experience, now on sabbatical. Co-founded @sablier-labs, a VC-backed crypto payments startup.
Skills
npx skilld add paulrberg/agent-skills --allUse for current official documentation about Codex, Codex CLI, or Claude Code behavior, configuration, prompting, skills, permissions, tools, surfaces, capabilities, troubleshooting, hooks, app-server, or hook trust. Consult the relevant official source before answering.
Use local Codex/Claude Code transcripts to assess recurrence risk for agent behavior. Recommend durable fixes to AGENTS.md or skills from that evidence. Not for routine post-success skill-evolution or agent self-improvement reviews.
Prune agent scratch directories on macOS. Move clearly outdated `.ai/` files into `.ai/archive/YYYY-MM-DD/` and trash clearly outdated `.cache/` entries. Leave `.ai/todos` untouched.
Use for device or simulator UI verification through Apple's Xcode MCP after UI changes or an explicit device check. Inspect screenshots and UI hierarchies, exercise controls, and report functional or visual defects.
Brainstorm with the user to generate, combine, and refine non-obvious ideas into a promising concept. Use when the user wants to brainstorm, ideate, explore possibilities, escape obvious approaches, or find a creative solution.
Use Chrome DevTools through PRB's shared attach-only Chromium browser for browsing, debugging, automation, visual inspection, console or network analysis, performance or memory profiling, screencasts, and Wayback Machine research.
Use for Foundry cast transaction actions: prepare, trace, simulate, sign, or broadcast. Also use to sign messages or encode/decode ABI/calldata. Delegate every standalone RPC read to evm-atlas.
Use for CoinGecko/cg CLI crypto market data: prices, market cap, trending coins, top gainers/losers, coin search, or historical/OHLC data.
Use for GitHub CLI automation: repository reads, workflow runs, search, codespaces, releases, configuration, or gh command syntax. Use yeet for contribution writes to PRs, issues, comments, or discussions.
Use for just/justfile task automation: create justfiles, write recipes, configure settings, add modules/attributes, or set up command-runner workflows.
Polish changed code when the user explicitly asks, or when an active workflow requests post-implementation simplification and risk-profiled review over a fixed file scope.
Shared vocabulary for designing deep modules. Use when the user wants to design or improve a module's interface, find deepening opportunities, decide where a seam goes, make code more testable, AI-navigable, or agent-legible, choose which invariants to enforce mechanically, or when another skill needs the deep-module vocabulary.
Commit staged or intended changes: compose a Conventional Prefix or Natural Language message, then use ai-commit — with --all, --staged, --close, --finding, or --push. Also squash a branch into one commit with --squash.
Copy the active Claude Code or Codex CLI chat transcript path to the macOS clipboard.
Remove unnecessary code, files, and prose, consolidate related logic and guidance, and retire obsolete behavior with evidence. Use when asked to deslop or clean up AI-generated material while preserving required behavior and contracts.
Use for nontrivial Effect 4 work including services/layers, typed errors, Schema/JsonSchema, Config, runtime/concurrency, @effect/vitest, effect/ai and @effect/ai-*, effect/sql and @effect/sql-*, Effect Atom (@effect/atom-*), or @prb/effect-next.
Use for targeted EVM chain, account, transaction, RPC, explorer, bridge, and DEX evidence: chain name/ID, native symbol, RouteMesh, wallet balances and DeFi positions via DeBank/Blockscan in Chromium, cross-chain USD portfolio value or net worth, token/NFT holdings/transfers, tx history, funding origin via Etherscan/Blockscout/Chainscout. Covers Arbitrum canonical withdrawals, Across, Bungee, deBridge, Gas.zip, Hop, Layerswap, LayerZero, LI.FI, Relay, Socket, Symbiosis. Covers Uniswap v1-v4, Universal Router, Permit2, 1inch Classic/Fusion/Fusion+, and CoW Swap, CoWSwap, CoW Protocol, or GPv2 swaps, orders, liquidity, approvals, permits, rewards, migrations, wrapping, cancellations, and refunds.
Audit an entire repository with fresh eyes for correctness errors, bugs, omissions, duplication, inconsistencies, stale, duplicate, or tautological tests, stale comments, and other mistakes supported by evidence. Fix every safe issue and verify the result.
Use when creating or substantially redesigning web interfaces, landing pages, dashboards, components, or other frontend UI where visual direction and implementation quality matter. Produces subject-specific art direction, accessible code, and rendered visual verification.
Interview the user relentlessly about a plan or design until every branch of the decision tree has an explicit recorded decision and no open questions remain. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
Use to build interactive single-file HTML playgrounds, explorers, or tools with visual controls, live preview, and prompt copy-out, or to save the current task's findings as an interactive HTML debrief/report.
Interview the user about a plan, idea, or design through a short sequence of high-leverage questions, then summarize the direction and next step. Use when the user wants a relaxed interview, lightweight clarification, or says "interview me", without exhaustive grilling.
Discover large source-file refactor candidates and propose split plans based on cohesion and risk, using available semantic tooling.
Research and compare current language models for a task or use case. Covers hosted APIs, model gateways, open weights, and local inference, with source evidence, deployment fit, quality uncertainty, and equal-workload cost comparisons. Use only when explicitly invoked by the user or loaded by an authorized parent workflow.
Refactor naming and repository structure exhaustively while preserving behavior and external contracts.
Use for dependency updates: bump npm/pnpm/yarn/bun packages, check outdated, or run taze.
Orchestrate delegated research or implementation with Claude or Codex agents. Default to Claude in Claude Code and Codex in every other harness. Follow an explicit user choice of agent or model. Plan and launch without a separate plan approval step.
Use when PDF files are the primary input or output: read, compare, reconcile, extract text/tables/images, OCR scans, fill forms, split, merge, rotate, rename, compress, or convert between PDF and images. Optimized for private financial, tax, legal, and health documents on macOS.
Cut a release: bump versions, write changelogs, commit, tag, and approve staged npm publishes.
Run one task or skill across many repositories with per-repository agents, coordination claims, validation, commits, and one results table. Use when the same task must run in several repositories.
Audit two or more interdependent repositories for drift, duplication, and practices one repository uses well and another lacks. Verify every finding mechanically, decide with the user, then apply surgical alignment fixes and adapted transfers per repository.
Preview or rename a GitHub repo, or move a local repo to another parent directory. Preserve Codex CLI and Claude Code continuity references.
Run a retrospective on Codex or Claude Code sessions. Propose changes to steering files, checks, review rules, skills, tools, or information access. Check recurrence across transcripts with agents-introspection.
Harmonize a repository's skill portfolio across catalog and user-installed skills.
Post one user-supplied text, with optional images, video, or link, to X, LinkedIn, Farcaster, and Facebook (in Romanian), and to Hacker News when it suits that audience. Lightly tailor Reddit drafts and Hacker News copy. Resize source artwork into OpenGraph and social image presets with ImageMagick. Also refresh stale link previews or find subreddits. Publishing uses shared Chromium.
Use when CSV, TSV, or Excel (.xlsx) is the primary input/output: design or review text-table schemas, inspect, transform, validate, convert, or recalc formulas, or create/fix spreadsheets. Do not trigger when tabular data is incidental.
Use for Tailwind v4 styling: add/fix classes, configure or migrate Tailwind, use tailwind-variants, or tw-animate-css.
Create one decision-complete task handoff in its repository or on the Desktop for cross-repository work, return a command for a fresh interactive Codex session, and recommend Codex subagent delegation for complex tasks.
Archive checked `.ai/TODO.md` tasks into `.ai/todos/YYYY-MM/DD.md`, leaving unchecked tasks.
Use to find, compare, and recommend current tools, packages, libraries, CLIs, VSCode extensions, agent skills, databases, or infrastructure options for a development task, including Swift for iOS/macOS and Kotlin for Android.
Use for Vitest in TypeScript projects (Node, bun, React/Next.js, Effect): write, run, or debug unit/component tests, mocks, testing utilities, and coverage.
Wind down a long-running session fast when the user must leave — freeze new work, preserve active subagents, secure finished work, and hand off the remainder via task-handoff.
Edit user-provided English text or files with surgical corrections to grammar, spelling, punctuation, and clearly awkward phrasing. Use for proofreading or making existing writing sound natural and casual while preserving the author's meaning and voice. Not for drafting from scratch or broad rewrites.
Use for GitHub PR/issue/discussion workflows: create/update PRs, issues, or discussions and post issue or discussion comments; triggers include yeet.
Curated skills from @paulrberg. Checked GitHub just now