Open source/5 agent skills/Host-neutral

Reusable playbooks for coding agents that need to do real work.

These are open-source engineering playbooks for coding agents and teams — focused procedures for the work where vague advice is not enough.

Each one captures a repeatable method: preserve behavior during a Rust migration, make a smaller code change, keep version-control recovery open, or verify technical claims before publishing them.

Install

npx skills@latest add RealAdarsh/skills
Claude Code
/skill-name
Codex
$skill-name
One skill only
--skill <name>
Source on GitHub ↗

The agent playbooks

Use what fits the work

01

The front door

ask-adarsh

Routing · Any agent

Describe what you are trying to do, and it tells you which of these skills to reach for — one primary skill, a second only if the task genuinely has two halves, plus the exact command to invoke it and the first concrete step.

Why it exists — A collection of skills is only useful if you can find the right one without reading all of them. This is the routing layer. It is also allowed to say “none of these fit” and name the missing capability, instead of stretching a skill to make the set look complete.

  • Routes on the outcome you want, not on keywords in your sentence.
  • Prefers one focused skill over stacking three together.
  • Admits when nothing fits rather than inventing a match.
02

The heavy one

cpp-to-rust-port

Rust · C / C++ · FFI · Migration

Turns “port this C++ to Rust” into an actual migration: agree what must behave identically, freeze that behavior in tests first, port one subsystem end to end, and only delete the C++ once the Rust passes the same inputs with the same outputs.

Why it exists — Ask an agent to port C++ and you get a line-by-line translation that compiles but is not Rust — class hierarchies turned into Arc<Mutex<T>>, errors still stringly-typed, the borrow checker silenced with .clone() everywhere. Migration is a redesign problem, so the skill teaches the process rather than the syntax.

  • Parity harness before the rewrite: characterization tests, golden files for binary formats, and differential tests that run both implementations on the same input.
  • Real rules for the dangerous edges — where unsafe is allowed to live, how to stop a Rust panic crossing an FFI boundary, how to make CMake and Cargo agree on build profiles.
  • A named list of the anti-patterns it will not let you ship, and completion gates covering behavior, safety, performance, and observability.
  • Proven on asdcplib — a production digital-cinema C++ library with deep ties to cryptography, KLV parsing, and SMPTE standards — rather than on a textbook example.
03

Version control

jj

Jujutsu · Git · Workflow

How to drive Jujutsu safely, including repositories that are colocated with Git: the daily flow, history edits, bookmarks and pushing, conflict resolution, and recovery — with a cheat sheet for translating the Git command you were about to type.

Why it exists — Agents bring Git muscle memory to a tool that does not work like Git. They open commands that wait for an editor and hang, and when history goes wrong they reach for a destructive reset. This skill keeps every command non-interactive and points recovery at the operation log, where nothing is actually lost.

  • Every command form is safe to run unattended — messages passed as flags, splits done by file path.
  • Recovery is jj undo and jj op restore, not a hard reset.
  • In colocated repos it keeps Git read-only, so the two tools do not fight over the working copy.
04

Everyday editing

reuse-first

Code review · Refactoring · Any language

Make the smallest coherent change that fits the repository you are already in. Look for the existing utility, test helper, error type, and convention before writing a new one — then reuse, extend, extract, or add, in that order.

Why it exists — Left alone, an agent writes new code. You end up with a wrapper that renames one call, a manager that forwards to a service, and comments narrating what the next line obviously does. None of it is wrong, and all of it makes the codebase harder to read. The skill exists to make the smaller change the default.

  • Names the shallow abstractions it refuses to add — wrappers, helpers, managers, and forwarding methods that pass a call through.
  • Equally clear about when not to reuse: do not distort a working API just to claim you reused something.
  • Has an opinion on comments — explain the rationale and the constraint, not the control flow.
05

Writing

technical-blog-writing

Writing · Documentation · Verification

Plan, draft, and fact-check a developer-facing article. Pick the shape that matches what you are writing — tutorial, deep dive, migration story, benchmark, postmortem — then verify the claims in it before anyone else has to.

Why it exists — Agents write fluent, confident technical prose, and some of it is invented: benchmark numbers with no run behind them, citations to pages that do not say what was claimed, code samples nobody executed. The centre of this skill is an evidence ledger — every substantial claim gets a source and a verification status before drafting continues.

  • Code is treated as a tested artifact: run it, show the real output, or label it unverified and say what is left to check.
  • Benchmarks have to carry their methodology — environment, baseline, sample counts, limitations.
  • Editing happens in named passes, so accuracy is checked separately from clarity.

The common thread

Leverage still needs verification.

Every one of these does the same thing in a different domain: replace a plausible answer with a checked one. Agents are extremely good at plausible. Freeze the behavior before you rewrite it, look before you add, keep the undo path open, and verify the claim before it gets published under your name.

Browse the repo ↗