Skip to the page

What it does

Skills

Teach your assistant how you like something done, once, and use it with every model.

A skill is instructions for one kind of job, like your weekly review or a release checklist. Write it once, in your own words. Every provider can use it, whenever a request fits.

Teach one

  1. Open Skills in the sidebar and press New skill.
  2. Describe the steps, the tone and what to leave out. With only a sentence, press Write the steps for me. Back to my words restores yours.
  3. Pause, and Conch writes a title and description you can change.
  4. Under Use it, choose Automatically or When I ask, then press Create skill.

Write the steps for me needs a provider that can write, such as Claude Code. Without one, your words are the skill.

Update one from chat

Ask your assistant to change a saved skill and describe the new steps. It reads the current instructions and offers an update in the chat. Review Current instructions and Proposed instructions, then press Update skill. Not now leaves the skill unchanged. This review works in Auto and Full trust.

The update preserves when the skill is used and what it may do. If the skill changed while you were reviewing, ask for a fresh proposal. For a skill owned by another app, open it in Skills and choose Make a copy to edit first. Signed skills must be reviewed on the Skills page.

Use one

Each skill has one of three settings:

  • Automatically. Used whenever a request fits.
  • When I ask. Only when you type / and its name, like /weekly-review, then what to work on.
  • Off. Never, but kept and backed up. Off at the top of the list shows them, and CtrlK finds them as Skills that are off.

A line in the chat names the skill that shaped the reply. Change the setting on the skill's page or with its switch. Try it in a chat starts a chat with it filled in. Choosing it in CtrlK puts it in the message box.

When a skill that's off would help

If an Off or When I ask skill fits, a card under the reply offers it and says why.

  • Off: Turn on shows what it can do; Turn on again sets it to Automatically.
  • When I ask: Use it shows what it can do. Then Use it for this request, or Always to set it to Automatically.

The chat carries on using it. Not now and Don't suggest work as for apps. A worrying skill is never offered.

Add a skill people share

Discover, the second tab on the Skills page, finds skills from Anthropic's own collection, ClawHub and skills.sh.

  1. Open Skills and choose Discover. Search, or pick a kind such as Writing.
  2. Each card shows what it does, who published it, how many use it, and the source's rating: Official, Verified publisher, Community or Flagged.
  3. Open one. Conch reads every file first and shows what it can do, what it found, the version and the licence.
  4. Choose When it fits or Only when I ask, and press Add skill.

It's yours, under Added from Discover. Nothing runs when it's added.

  • A worrying skill shows what Conch found. Add anyway waits until you tick that you've read it.
  • Some can't be added, when its source found harmful code or its licence limits it to its maker's apps. You see why.
  • Updates wait for you. Its page says A newer version is there. Read the update shows each changed file, and first whether it asks to do more. Nothing changes until you press Update.
  • Remove on its page takes it away.

When a shared skill would clearly help, a card under the reply offers it. Add and carry on adds it and continues. Don't suggest skills from Discover, under More, stops these. A chat that has read something from outside never offers them. Offline, Discover shows its last results.

Skills you already have

Conch saves a skill as a SKILL.md, the format Claude Code, Codex, OpenClaw and Hermes read. Skills in other assistants' folders appear under From other apps, turned off until you read and turn them on. Conch never changes another app's folder: Make a copy to edit gives you your own.

What keeps a skill honest

A skill runs with your powers, so:

  • It's read first. Conch says plainly what could hurt you, like running a download or reaching for saved passwords. A worrying skill stays off until you've looked.
  • It's held to what it says. Its page lists what it can do, like "run commands (only git), change files in your work folder". In Ask first and Read only, anything outside that list asks for the rest of the chat and its tasks. Auto judges each action by its risk; Full trust skips action approvals. Loading a skill never silently changes your selected mode.
  • A change turns it off. If another app's skill changes after you turned it on, it's off until you look again.
  • A signature says who made it. Press Trust this publisher… on a signed skill once. Their skills then say Verified, and their signed updates stay on. A skill changed after signing can't be turned on.

To sign your own, see conch skills sign in the command line reference. Your signing key opens only on this computer; a passphrase-locked backup moves it.

Stop holding a chat to a skill

In Ask first and Read only, the skill's list stays active after use. Auto and Full trust suspend these restrictions without deleting the list. When you're done:

  1. Press Stop holding on the line above the message box, or find Stop holding this chat to … in CtrlK.
  2. Read what changes, and press Stop holding.

The chat and Activity record it. Only you can do this, and not mid-answer.

Skills Conch suggests

Ask for the same thing in three chats, and the Skills page offers a draft skill for it. Once search understands meaning, the words needn't match.

  • Look at the draft opens it to edit.
  • Not now hides it for a month.
  • Don't suggest this hides it for good.

Save how I did this

When your assistant works something out the long way, a card under the reply offers Save as skill ("Log a meal in Yazio"). It appears when:

  • a background task finished and Conch confirmed it;
  • a routine ran and did its job;
  • you said it worked ("perfect, thanks") right after a few steps;
  • or it took many steps and ended well.

The draft keeps the steps that worked. One-off details, like a date, become questions for next time. It starts as When I ask and can do only what the work needed. Edit it, then press Create skill.

  • Offered once per chat, never mid-answer, and never for work that failed.
  • Not now puts it away. It also waits on the Skills page, with Don't suggest this, and in CtrlK in that chat.
  • If the chat read something from outside, the offer says so ("Learned from trains.example"), and Conch drops steps that look slipped in.
  • Never offered for a chat with someone else's messages in it.
  • The chat's own provider writes the draft, so nothing goes anywhere new.

A tidy shelf

When a skill Conch suggested, saved from your work or brought in from another app goes unused for two months, the Skills page says You haven't used these in two months. Turn them off moves them under Off; Keep them asks again in two months. Skills you wrote, and skills still in another app's folder, are never on this card.

Good to know

  • Your skills live in ~/.conch/skills, one folder each, and are in every backup.
  • Claude Code reads its own skills folder, so it may use those skills whatever you choose. The skill's page says so.
  • Publishers you trust are listed at the bottom of the Skills page. Forget stops trusting one.