Paul Berg Agent Skills - Setup Guide¶

Source: paulrberg/agent-skills via skills.sh - 32.7K combined installs across 97 indexed listings; first seen 2026-10-31 (morning sweep)

GitHub: paulrberg/agent-skills - 104 stars, MIT, branch main, last pushed 2026-10-30; 43 skill directories under skills/ with one SKILL.md each and 169 markdown files between them, plus a toolkit/ of Rust CLIs and an mcp/ server

Category: Smart contract work for agents: Solidity style, Foundry testing, deployment and code polish

Quality Tier: 🟢 Green - all five sampled skills Pass all three auditors and the repository is MIT markdown, though it is one engineer's personal catalog with 104 stars rather than a maintained platform

Paul Berg's repository is a working engineer's setup published as a skills pack: 43 skill directories, plus the command-line tools his skills call (ai-commit, ai-coord, ai-handoff, ai-notify, ai-skillet), and an MCP server directory. The README opens with a warning worth repeating in full substance: this catalog reflects Paul's preferred tools, defaults, safety boundaries and writing voice, it is not a neutral template, and anyone installing it should review every workflow and customize it for their own stack and agents.

The label on this page reads Solidity and Foundry, and that is the thinnest part of the offer. The top of the install table is TypeScript and styling: effect-ts at 3,122, tailwind-css at 3,010, cli-gh at 2,835, code-polish at 2,781. The EVM work is real but small: cli-cast covers Foundry's cast CLI for preparing, tracing, simulating, signing and broadcasting transactions, evm-atlas covers chain, account, bridge and DEX evidence lookups, etherscan-api sits at 26 installs and evm-chains at 176. What you install is a disciplined agent workflow kit with a crypto corner in it.

Installation¶

The README's skill route is one line:

bunx skills add PaulRBerg/agent-skills

The CLIs are separate, and several skills assume they are on the path. They build from source with Rust nightly:

cargo +nightly install --locked --git https://github.com/PaulRBerg/agent-skills ai-commit ai-coord ai-handoff ai-skillet

macOS users add desktop notifications:

cargo +nightly install --locked --git https://github.com/PaulRBerg/agent-skills ai-notify

From a clone, just toolkit::install-cli puts the supported CLIs into ~/.local/bin, and the toolkit README covers prerequisites and Linux clipboard setup. That last point matters on a Linux host: reading and writing the clipboard is how several of these workflows hand text between agents.

Hermes installs skills from a directory, so clone and copy:

git clone --depth 1 https://github.com/paulrberg/agent-skills.git /tmp/prb-agent-skills
cp -r /tmp/prb-agent-skills/skills/. ~/.hermes/profiles/corpusiq/skills/

That copies all 43 skill directories. A practical subset for a non-crypto profile is code-polish, commit, release-bumper, node-deps-bumper, cli-gh, yeet, codebase-design, large-file-refactor, naming-refactor, fresh-eyes-sweep and agents-brain. The CLIs stay optional until a skill actually asks for one: commit calls ai-commit, and orchestration wants native subagents in Claude Code or Codex, or an authenticated CLI for shell execution.

What It Provides¶

Skill	Installs	Use For
effect-ts	3,122	Effect 4 work: services and layers, typed errors, Schema and JsonSchema, Config, runtime and concurrency, `@effect/vitest`, `effect/ai`, `effect/sql` and Effect Atom.
tailwind-css	3,010	Tailwind v4 styling: adding or fixing classes, configuring or migrating Tailwind, `tailwind-variants` and `tw-animate-css`.
cli-gh	2,835	GitHub CLI automation: repository reads, workflow runs, search, codespaces, releases and configuration, with contribution writes left to `yeet`.
code-polish	2,781	A post-implementation pass over an explicit file scope: simplification and risk-profiled review with autofix, driven by `--simplify` and `--review`.
yeet	2,584	GitHub contribution writes: create and update PRs, issues and discussions, post comments, with each mode named in the argument hint.
bump-release	2,362	No directory in the clone and no row in the README's 43-skill table; the nearest live entry is `release-bumper`, which bumps versions, writes changelogs, commits and tags.
bump-deps	2,342	No directory in the clone; dependency work now belongs to `node-deps-bumper`, which builds one taze plan and applies compatible ranged updates.
code-simplify	2,310	No directory in the clone; `code-polish` carries the simplification half today through its `--simplify` flag over a fixed scope.
code-review	2,014	No directory in the clone; `code-polish` carries the review half through `--review`, risk profiles and autofix.
md-docs	1,702	No directory in the clone and nothing in the live table covers markdown documentation authoring, so this install figure is historical.
biome-js	1,067	No directory in the clone; the repository ships no formatter or linter skill, so treat the count as a listing that predates the current tree.
cli-just	789	`just` and justfile automation: creating justfiles, writing recipes, settings, modules and attributes for a command runner.
commit	344	Committing staged or intended changes: a Conventional or natural-language message through `ai-commit`, with `--all`, `--staged`, `--push` and `--squash`.
cli-cast	305	Foundry's cast CLI: prepare, trace, simulate, sign and broadcast transactions, sign messages, and encode or decode ABI and calldata.
spreadsheets	259	CSV, TSV and XLSX as the primary input or output: designing text-table schemas, transforming, validating, converting and recalculating formulas.
git-squash	235	No directory in the clone; `commit` takes `--squash` to fold a branch into one commit today.
work	230	No directory in the clone and no live equivalent in the README table, so the figure survives from an earlier index.
autoresearch	219	No directory in the clone; `tool-finder` covers finding and comparing developer tools, which is the adjacent live skill.
oracle-codex	190	No directory in the clone; `orchestration` covers research, planning and implementation across Claude or Codex agents.
evm-chains	176	No directory in the clone; `evm-atlas` covers chain names and IDs, native symbols and the evidence lookups this name implies.

The remaining 77 indexed listings range from 2 to 175 installs.

Thirteen of the 20 rows above point at names the repository no longer ships. Across the whole index the gap is larger: 54 of the 97 listings have no directory in the clone, which is the single most important thing to know before trusting this table as a shopping list. The 43 directories are all in the install counts, so the pack is not stale, but the index has a long memory and the repository does not.

Why This Matters for Hermes Agents¶

The skills that carry this pack solve the unglamorous problems that decide whether an agent session is useful. code-polish takes an explicit file scope and produces either a simplification pass or a risk-profiled review with autofix, which is a much narrower contract than "review my code" and far easier to trust. commit composes a message and hands the actual snapshot to ai-commit, which exists to make immutable commits in shared working trees. release-bumper and node-deps-bumper each do one mechanical job with a dry run. Those are the shapes that survive daily use.

The agent-tooling half is unusually introspective. agents-brain maintains repo context and generates AGENTS.md files or repo-local skills, agents-introspection assesses recurrence risk from Codex and Claude Code transcript history, agents-docs fetches current official Codex and Claude Code documentation including hooks and trust settings, and skill-harmonization reconciles a repository portfolio against user-installed skills. For a profile that runs many suites at once, skill-harmonization is the answer to skill sprawl rather than another source of it.

The repository also documents its own maintenance loop. skill-harmonization reconciles repository and user-installed skill portfolios, repo-harmonization aligns interdependent repositories and transfers practices between them, agents-system redesigns repo systems for agent understanding and control, and .agents/internal-skills/ holds the author's own procedures for skill authoring, sync, lifecycle and publication. That is a level of self-hosting most packs skip, and it means the conventions an agent learns here transfer to writing new skills rather than only to using existing ones.

The crypto corner is small but specific. cli-cast delegates every standalone RPC read to evm-atlas rather than reimplementing chain queries, and evm-atlas covers bridges (Across, Hop, LayerZero, LI.FI and more) and DEX routing (Uniswap v1 through v4, 1inch, CoW Swap) as evidence lookups. For Solidity authoring there is no compiler-level skill here: cli-cast operates Foundry's cast tool, and testing and deployment discipline has to come from your own project conventions.

Usage¶

You say	What happens
"Polish the code I just changed before I review it"	`code-polish` takes the file scope, runs the simplification or risk-profiled review pass, and applies autofix where the profile allows it.
"Commit this with a conventional message"	`commit` composes the message and hands the snapshot to `ai-commit`, honouring `--all`, `--staged` or `--squash` and closing issues named in the diff.
"Cut a release for the two workspace packages"	`release-bumper` bumps versions, writes the changelog entry, commits and tags, with `--dry-run` available before it writes anything.
"Which dependencies are stale, and what will break if I bump them"	`node-deps-bumper` builds a single taze plan, applies compatible ranged updates and isolates the major-version decisions for review.
"Simulate this transaction before we broadcast it"	`cli-cast` prepares and traces the call, simulates it, and only then signs and broadcasts, with ABI encoding handled in the same skill.
"What is this address, and where did its funds come from"	`evm-atlas` resolves the chain and account, pulls balances and history, and follows funding origin through Etherscan, Blockscout and Chainscout.
"Our AGENTS.md is out of date and the repo has no skills"	`agents-brain` regenerates repo context and writes AGENTS.md or repo-local skills from what the repository actually contains.

Verification¶

The raw path for the highest-install skill returned HTTP 200 on branch main when this guide was written on 2026-10-31:

curl -sL -o /dev/null -w "%{http_code} effect-ts\n" \
  https://raw.githubusercontent.com/paulrberg/agent-skills/main/skills/effect-ts/SKILL.md

Expect 200. After the copy, check that the named directories landed and read a header, since these skills use frontmatter fields beyond the Agent Skills minimum:

ls -d ~/.hermes/profiles/corpusiq/skills/{code-polish,commit,effect-ts,cli-gh,yeet,agents-brain}/ | wc -l
head -12 ~/.hermes/profiles/corpusiq/skills/code-polish/SKILL.md

code-polish should show an argument-hint line with [--simplify] and [--review], a description: explaining the fixed file scope, and any skill-dependencies: block. If you also installed the CLIs, confirm the binary resolves before a skill calls it, for example command -v ai-commit for the commit workflow.

Security¶

skills.sh verdicts for sampled skills (verified 2026-10-31) - re-check the security pages on skills.sh before production use:

Skill	Gen Agent Trust Hub	Socket	Snyk
effect-ts	Pass	Pass	Pass
todo-archive	Pass	Pass	Pass
task-handoff	Pass	Pass	Pass
ai-prune	Pass	Pass	Pass
agent-context-polish	Pass	Pass	Pass

All five sampled skills Pass all three auditors. No Fail and no Warn appears in the sample.

The auditors scan skill text, and the sharper risk here is what the text tells an agent to delete or publish. ai-prune archives stale .ai files and moves stale .cache entries to the trash on macOS, todo-archive moves checked tasks out of TODO.md, ai-commit writes immutable commits into shared working trees, yeet creates and updates public PRs, issues and discussions, and social-media cross-posts. Each of those is a write that cannot be trivially undone, so scope the agent's repository and credentials before enabling them, and prefer the --dry-run flag where a skill offers one.

The second surface is the CLIs. They are Rust binaries built from this repository with a nightly toolchain and installed from git rather than from a published registry, so there is no checksum or release provenance to verify, only the commit you cloned. Treat cargo +nightly install --locked --git ... as a build from source with the trust that implies, and keep the installed set small. A Hermes profile can use the 43 skill directories without any of the binaries; the only loss is that commit and the handoff skills will ask for tooling that is not there.

Integration with CorpusIQ¶

CorpusIQ is the connector layer that lets AI assistants read the systems a business already runs on - accounting, payments, CRM, email, files and analytics - and answer with the source named. It is read-only by design: it retrieves data and never writes to, or acts in, a connected system, which is what makes it safe to point at live accounts.

This page is part of the CorpusIQ directory, the reference we maintain for the MCP server and agent-skill ecosystem. Run alongside CorpusIQ, a third-party server like this one can be combined with a company's own business data in a single answer.

See what CorpusIQ connects

Limitations¶

Snapshot data, verified 2026-10-31: 32,742 combined installs across 97 indexed listings; 104 GitHub stars; MIT; last pushed 2026-10-30. Counts drift over time.
Index and tree diverge badly. The clone holds 43 skills and skills.sh lists 97, and 10 of the 20 rows in the table above name skills the repository does not contain, so the table is a record of what people installed, not a description of what you get.
The catalog is one person's defaults. The README states plainly that it is not a neutral template and reflects Paul's tools, safety boundaries and writing voice, so an install carries opinions about commit message format, release tooling and agent delegation that your team may not share.
Install counts are not a ranking. The most-installed skill is effect-ts, a TypeScript Effects library guide, while the Foundry and EVM skills this page is named for sit at 305 installs and below; popularity here tracks the author's own stack, not the category.
Several skills assume macOS and third-party accounts. pdf, spreadsheets and ai-prune say macOS in their own descriptions, evm-atlas wants the routemesh CLI and DeBank or Blockscan access, and the clipboard workflows need Linux setup from the toolkit README.
Skill text assumes the CLIs. commit calls ai-commit, and orchestration needs native subagents in Claude Code or Codex or an authenticated CLI, so a skills-only install will hit instructions that reference binaries you chose not to build.
Smart contract advice stops at tooling. There is no Solidity style guide, no compiler-version triage and no Foundry test-pattern skill in the tree, so this is not a substitute for a dedicated Solidity suite.

Skills Catalog | Skills Marketplace

Pro Tips¶

Install by directory name, not by index row. Ten of the twenty names in the table above do not exist in the clone, so check ls skills/ before adding a skill to a config, and skip the historical rows entirely.
Take code-polish instead of the retired code-review and code-simplify pair. It is the live skill with both halves, and its fixed-scope contract makes the output reviewable rather than open-ended.
Leave the CLIs out until a workflow demands one. The skills work as markdown; ai-commit and friends cost a nightly Rust toolchain and a build from git, which is a large ask for the commit-message step alone.
Copy agents-brain even if you skip the crypto skills. It writes AGENTS.md files and repo-local skills from the repository's actual shape, which is the part that makes a Hermes profile usable on an unfamiliar codebase.
Prefer evm-atlas over ad-hoc RPC calls for read-only chain questions. It routes evidence gathering through named explorers and bridges, and cli-cast defers every standalone RPC read to it rather than duplicating the lookup.

← Skills Catalog | Skills Marketplace → Powered by CorpusIQ