essence

Only what you mean. Not a word more.

essence strips an AI coding agent's writing down to what the reader needs: chat replies, code comments, commit messages, PR/issue text, docs, published copy. The exception is private scratch work: internal reasoning and throwaway notes were never written for a reader.

An agent's writing usually picks up cruft from the conversation that produced it: hedges, meta-commentary about what was asked, preamble, restated context. A reader who never saw that conversation should understand the result fully, with nothing in it that refers back to something they weren't part of.

essence vs caveman

Caveman compresses words: telegraphic grammar, dropped articles, same content in fewer tokens. essence filters content: full grammar, full sentences, but only the sentences that actually carry meaning. Different axis: caveman changes how something is said, essence changes what gets said at all. They compose fine; essence doesn't touch phrasing style.

Before / after

Chat reply (the same input, shown two ways, to make the axis concrete):

Before:  Great question! So, not to reopen what we already decided on the retry
         logic, but I noticed something worth flagging: right now if the API call
         fails, we just log the error and swallow it, so the caller has no way of
         knowing the write didn't actually happen. I think it'd probably be a good
         idea to return an error here, though I could be wrong about the approach.

essence: If the API call fails, the error is logged and swallowed: the caller has
         no way to know the write didn't happen. Return the error so the caller
         can decide what to do.

caveman: API fail: error logged, swallowed. Caller unaware write skip. Return
         error, caller decide.

essence cuts "great question", "not to reopen what we decided", and the hedging, but keeps full sentences and normal grammar.

Code comment

Before:  // Fetches the user's profile from the database and returns it. Handles
         // the case where the profile doesn't exist yet for brand-new users who
         // haven't finished onboarding.

essence: // profile may be missing until onboarding completes; caller must handle nil

PR description (no fixed template; this is prose the reviewer needs, not a form to fill in):

Before:  Per our discussion, I split this into two PRs instead of one big one like
         you suggested. This one only covers the backend endpoint; I didn't touch
         the frontend since we agreed that could land as a follow-up once the API
         is stable.

essence: No endpoint exists for revoking an API key once issued. Adds
         `DELETE /api-keys/:id`. Frontend follow-up is a separate PR, pending API
         stabilization.

Meta-commentary, caught in the wild (a real mistake, made while writing this very README):

Before:  essence is the default for everything you write, not a mode that switches
         on for certain artifact types.

essence: essence is the default for everything you write.

"Not a mode that switches on for certain artifact types" only makes sense to someone who'd just watched an earlier, conditional design get replaced with this one, which reads as arguing with a draft the reader never saw. Same rule, applied to itself: state the fact, drop the rebuttal.

What you get

Command	What it does
/essence [on|off]	Toggle the always-on filter. On by default every session.
/essence-pr [target]	PR/MR description: bug + fix, nothing else.
/essence-doc <file>	Rewrite a doc/comment/note file to essence; preserves code/URLs/paths exactly, backs up the original.
/essence-commit	Commit message: Conventional Commits, why over what, no AI-attribution trailer.

Install

Claude Code (no clone needed, the plugin lives in this repo):

claude plugin marketplace add jurassix/essence && claude plugin install essence@essence

Codex, Gemini, Cursor, Windsurf, or Cline (these write into a local rules file, so clone and run the installer):

git clone https://github.com/jurassix/essence.git
cd essence
node bin/install.js            # detect installed agents, install for each
node bin/install.js --list     # see the agent matrix + what was detected
node bin/install.js --dry-run  # preview without writing anything

See INSTALL.md for the per-agent matrix and manual steps.

Repo layout

• skills/essence/SKILL.md: the ruleset, single source of truth, read at runtime by the activation hook and by the installer
• skills/essence-pr/, skills/essence-doc/, skills/essence-commit/: deliverable-specific skills
• commands/: Claude Code slash-command definitions (.md)
• src/hooks/: SessionStart / UserPromptSubmit hooks powering Claude Code's always-on mode
• bin/install.js: multi-agent installer; install.sh / install.ps1 are thin shims over it