If your AI workflow still makes you repeat the same instructions in every single chat, that's a tax you don't need to pay. I built Ink and Agency to fix that friction point. It’s a practical skills pack that installs as a plugin in Claude Code (and now Codex), giving you reusable prompts for writing, planning, triage, architecture design, and just general day-to-day work.
I wanted better outputs with less effort and more consistency across tasks. So I turned repeatable playbooks into discrete skills, each with clear boundaries and predictable behavior.
Why I Built It
Most agent failures don't happen because the model is bad. They happen because the prompt quality falls apart.
You ask for something broad, get a broad response. Then you spend time correcting it—steering it back on track. That’s fine if you do it once in a while. But when that becomes your daily job, it grinds you down.
Ink and Agency packages task-specific behavior into reusable units. Now, you can route work by intent instead of rewriting the entire process every time.
Get it here
Everything lives at risadams/ink-and-agency. It’s MIT licensed and free to use.
Two older repos of mine, risadams/skills and risadams/claude-subagent, were folded into this one. Both histories came along via git subtree, so nothing was lost. The reason for the merge was simple: the subagent library couldn't ship to Codex, which has no equivalent primitive. Rewriting those specialists as skills made the whole library portable, and collapsed two install steps into one.
What's In The Pack
200 skills, grouped into 15 category folders under skills/, each one a directory with a SKILL.md entry point. The split in the name is the split in the pack:
- Ink covers workflow: writing, sprint and Scrum work, issue management, Obsidian tooling, codebase analysis, debugging, and an end-to-end build loop that runs plan → spec → tickets → implement → tdd → review.
- Agency covers expertise: roughly 138 specialist skills (language and framework experts, infra, data and AI, security, product), plus
clarity-counciland its 46 advisory personas for decisions that need more than one point of view.
There's also a set of executive-function skills built with neurodivergent-friendly defaults, things like task-initiation, hyperfocus-recovery, and energy-budget. Those started as tools for me and stayed in because they earned their place.
Quick start for Claude Code:
/plugin marketplace add risadams/ink-and-agency
/plugin install ink-and-agency
Quick start for Codex:
codex plugin marketplace add risadams/ink-and-agency
codex plugin install ink-and-agency
That's the whole install. On Claude Code the skill names get namespaced at install time, so /ink-and-agency:sprint-plan won't collide with a sprint-plan skill in your own project. Most skills fire automatically when your request matches what they do; a handful are user-invoked only, and you reach those by typing the name.
If you can't remember what's in there, run /ink-and-agency:which-skill. It routes over the whole pack and hands back the exact invocation to use.
The Skills Improve Themselves
This is the part I did not plan for at the start. Every skill now runs a short evaluation at the end of an invocation and writes a journal entry when the run produced real signal: a correction, a retry, something that misfired, something that clicked. Routine clean runs write nothing. On the next invocation, the skill reads its journal first.
The journals are plain markdown on your machine, one file per skill under ~/.ink-and-agency/learnings/. Nothing leaves your disk. You can read them, edit them, or delete them.
When a learning suggests the skill itself should change rather than just your context, it proposes an edit against the source checkout if you have one, or files it as an idea with your consent. A skill never edits its own running copy in the plugin cache, because caches get wiped on every update.
Keeping It Current
Releases follow semver and land as GitHub Releases, with the details in CHANGELOG.md. To pull a new version:
/plugin marketplace update ink-and-agency
/plugin update ink-and-agency
Codex uses the same two commands with a codex plugin prefix. If you're hacking on the pack itself, run Claude with --plugin-dir pointed at your checkout and skip the marketplace entirely; it reads the working tree directly.
The Real Value in Day-to-Day Work
A good skills pack isn't just a bag of clever prompts. It’s a reliability layer. Here is what changes after you start using one:
- Better handoffs: Tasks are framed consistently, so the output quality doesn't feel random.
- Less context thrash: You stop re-explaining your process in every single chat thread.
- Faster execution: Known task types map directly to known skill patterns.
- Easier team adoption: Every skill has a stable surface area and clear purpose—it’s predictable.
If you work in DevOps, delivery, or product engineering, this pattern hits home quickly. You're already used to turning repeated work into scripts, templates, and automation. Skills are just that move for prompt-driven work.
A Few Quick Examples of Use
You don't have to write long instructions; you can just ask for the outcome:
Use break-it-down on this email to explain what it is really saying./ink-and-agency:codebase-explain for this module.Triage PROJ-1234.Run a clarity-council on this design tradeoff.
Skills also chain. Each one declares its workflow neighbors in frontmatter, so a full loop looks like grill → work-plan → plan-to-spec → plan-to-tickets → implement, with tdd and code-review running inside that last step. The named flows are documented in skills/FLOWS.md if you want to see how they fit together.
That's the whole point. You ask for outcomes, not rituals.
