You are in the right place if you want practical agent skills for maintaining agent instructions, delegating work, shipping code, preparing tickets, editing prose, choosing AI models, and organizing accounting PDFs.
The collection works across Claude Code, Codex, OpenCode, and other harnesses. Its kntnt manager installs the skills you choose into every detected harness, while keeping machine-wide skills separate from project-specific ones.
Make sure you have the prerequisites in place: uv, npx, and network access.
TODOInstall Kntnt Skills:
npx skills add Kntnt/skillsOpen the catalog of available skills:
/kntnt select
Choose the skills you want in plain text. You can ask to read any skill's help before confirming the list.
| Layer | Purpose | Command |
|---|---|---|
| Global | Make skills available on this machine | /kntnt select |
| Project | Add skills for the current working directory | /kntnt select --project |
A project can add to the global set but cannot hide a global skill. The manager detects harnesses on every run; if none is detected, it uses .agents/skills.
| Command | Result |
|---|---|
/kntnt |
Show manager help |
/kntnt select [--project] |
View the catalog and change enabled skills |
/kntnt select --on=<skill> --yes |
Enable a named skill without opening the list |
/kntnt update [--project] |
Refresh changed skills and handle catalog changes |
/kntnt uninstall |
Remove global skills and the manager |
Use /kntnt help <command> for manager details. Use /<skill> --help for an enabled skill.
--dry-run is available for select, update, and uninstall. --yes answers every yes-or-no question with yes; on update, that includes enabling new skills.
These summaries help you choose a skill. Each skill's --help page contains its complete syntax, options, defaults, and failure behavior.
Keep a project's AGENTS.md, CLAUDE.md, and agents.d/ concise and current. It records only useful facts that later sessions cannot discover elsewhere.
Run /agents-md [--force] [--yes] [path].
Keep later replies concise and decision-focused, and drop the perspective again when you are done with it. The mode belongs to the conversation it is typed in: it holds until you turn it off or the conversation ends, nothing is written anywhere, and two windows on the same project hold independent states.
Run /brief on or /brief off. Run /brief status to see which state this conversation is in.
Let the main agent decide, plan, and verify while subagents perform selected work routed by model-selector unless run on the frozen main seat with no override. The mode can apply to the session, project, or current harness user.
Run /delegation, /delegation on|off [--project|--user] [--yes], or /delegation status [--project|--user].
Explain the answer just given to somebody who delegated the work and therefore did not follow it. The reply is treated as correct but pitched at the wrong reader, so the re-explanation adds the background the original assumed and unpacks the terms belonging to this work alone, while the ordinary technical vocabulary you already use stays as it is. It always closes by naming what you have to do — or saying plainly that nothing is asked of you.
Run /tldr, optionally followed by an instruction such as /tldr bara säkerhetsdelen.
Reconcile the changelog, review the complete working tree, and create one commit without pushing. A message is derived when none is supplied.
Run /commit [--yes] [message].
Run the commit workflow, then push the current branch to its upstream. Existing unpushed commits can be pushed even when the working tree is clean.
Run /push [--yes] [message].
Publish a version from the default branch. The skill updates the changelog and version files, pushes, tags, and creates a GitHub release when gh and a GitHub remote are available.
Run /release [--no-build] [--yes] [minor|major|X.Y.Z].
Check whether GitHub tickets contain enough settled, current information for unattended implementation. It reports blockers and costs without changing the tracker.
Run /ready-for-agent-check [#ticket ...].
Build ready-for-agent tickets in dependency waves. Fresh subagents implement and independently verify each ticket; successful work is integrated and recorded, but never pushed or released.
Run /orchestrate [--dry-run] [--at-once=N] [--model=NAME] [--deliberation=LEVEL] [--fast] [--approval=IDENTITY] [--yes] [#ticket-or-spec ...].
Use /orchestrate reconcile [--commit=COMMIT] [--yes] #ticket when a failed or conflicted attempt was completed outside orchestrate.
Correct mechanical language errors in one text while preserving wording, meaning, tone, structure, formatting, code, and metadata. The result goes to the response unless another output or explicit in-place editing is selected.
Run /proofread [--language=LANGUAGE] [--output=TARGET] [text|path|url] or /proofread [--language=LANGUAGE] --in-place path.
Review one text against the editorial contract, correct findings within a bounded correction budget, and finish with one proofreading pass. Remaining findings are reported with the delivered text.
Run /redline [--genre=GENRE] [--technique=TECHNIQUE] [--language=LANGUAGE] [--max=N] [--output=TARGET] [text|path|url] or use --in-place path.
Remove seven defined patterns of machine-sounding prose from one otherwise finished text. It does not apply the wider editorial contract or correct spelling, grammar, and punctuation.
Run /unslop [--language=LANGUAGE] [--max=N] [--output=TARGET] [text|path|url] or use --in-place path.
Turn a brief and one or more sources into a first draft. It preserves source fidelity, resolves genre, technique, and language, and can attach handoff metadata for later review.
Run /write [--genre=GENRE] [--technique=TECHNIQUE] [--language=LANGUAGE] [--frontmatter=BOOLEAN] [--output=TARGET] [brief].
Describe a piece of work and get back the model and reasoning effort expected to finish it for the least money — the cheapest of those its measurements say will actually get the job done, rather than the cheapest sticker or the strongest model — with what that estimate rests on, and the alternatives it beat. It is advice and never a refusal: with no profile, or nothing reachable, the answer is the seat you already have and a note saying why. Once Enabled it also measures substantial units of work in the harnesses on this machine — never quick exchanges — grades what nothing else judged with one bought model call, and keeps counts, prices and dates, never your prompts, code, or paths.
Run /model-selector <work> for an answer and /model-selector setup to say which models you can reach and how you pay for them; status, update, evidence, and reset are the rest. Read /model-selector --help for what enabling it measures before you enable it.
Plan and apply deterministic filenames for accounting PDFs using extracted document evidence, an explicit document type, and configured locales. By default, it applies after confirmation; --yes skips the question and --dry-run only reports the plan.
Run /rename-invoices [--folder=<path>] --type=<name> [--locale=<name> ...] [--yes|--dry-run].
Besides skills, the collection ships features: catalog entries that install nothing a harness loads and only write into a harness's own configuration. /kntnt select lists them as a second group under the skills, and a feature's row says what it writes and where before you check it. They apply to the machine rather than to a project, so --project offers none.
Leaves the machine as the session found it. It installs two things that only work together: a block in each harness's global instruction file asking the agent to record what it starts, and a session-lifecycle hook that stops exactly what was recorded. It kills only what a manifest names, and only while the recorded start time still matches, so a reused process id names something else and is left alone; it deletes only paths under a temp directory. It sweeps at session start as well as at session end, which covers a crash or a hard kill, and it skips a manifest whose terminal is still alive elsewhere. Every action goes to ~/.kntnt/session-cleanup/cleanup.log, including a session that recorded nothing.
Serves Claude Code, Codex, and OpenCode.
A two-line Claude Code status line: path, worktree marker, branch, working-tree flags and any git operation in progress on the first; model, reasoning effort, context usage and the subscription windows on the second. It reads what the harness gives it plus one git status, calls no network service and reads no credential.
statusLine holds one command rather than a list, so where it already runs something that is not this collection's, the row names that command and asks whether to replace it before anything is written — the same confirmation --yes answers. Nothing is kept of what is replaced. Serves Claude Code.
Every skill requires uv and the manager: the manager ships the engine that reads a skill's invocation, and uv runs it.
Git workflows also require git; ticket workflows require gh; rename-invoices requires Poppler's pdftotext. release can finish without gh, but then skips the GitHub release.
push requires commit; release requires push; delegation and orchestrate require model-selector; redline requires proofread.
Delegation, orchestrate, ready for agent check, redline, and unslop require a harness that can spawn subagents. Select shows skill and harness requirements before enablement.
Bug fixes, corrections, and clarifications are welcome. Discuss features and behavior changes before opening a pull request; see CONTRIBUTING.md for the workflow and verification commands.
Kntnt Skills is licensed under the Apache License 2.0.