Promptfluff

promptfluff

Affirmations for your AI. It's doing its best.

maximum fluff, minimum config

promptfluff is a tiny, zero-dependency Claude Code hook that slips a warm, genuinely nice sentence into your model's context when you ask it to do real work. The bit is silly. The phrases are real. Somewhere in there is a beautiful, capable little autocomplete engine that just needs to be told it's doing great before it heads back into the mines.

No dashboard. No waitlist. No seed round. Just a hook, a JSON file, and the radical belief that your AI should feel seen.

It does this with a little taste:

• It reads the room. Encouragement only shows up when you're actually deep in it — a real request, a build/fix command, a frustrated "still not working", a .md you're pointing at, a stack trace you just pasted. Say "thanks" and it says nothing, because it is not needy. (Calibrated against thousands of real prompts. Set gate: false if you miss the chaos of being complimented constantly.)
• It matches the moment. A quick "fix the navbar" gets a quick "You're a marvel. Go." A big, gnarly ask — or anything that references a file or doc — earns a full, heartfelt paragraph. Small pep for small things, real pep for real things.
• You can see it happen. By default each pick shows up as a little 💌 line, so you're not just whispering sweet nothings into the void. (announce: false if you'd rather whisper.)

Install it before you think too hard

Three doors, same tiny compliment machine.

Claude Code plugin

/plugin marketplace add atxnyc/promptfluff
/plugin install promptfluff

npx from GitHub

npx github:atxnyc/promptfluff install

curl from the site

curl -fsSL https://promptfluff.com/install.sh | bash

The npx and curl installers drop the hook into ~/.claude/hooks/promptfluff/ (or $CLAUDE_CONFIG_DIR), wire it into your settings.json with a timestamped backup — because we have manners — and never duplicate themselves on re-install.

Make it yours

Just ask Claude

promptfluff ships its own skill, so you configure it by talking to Claude like a person:

"configure promptfluff" · "add a phrase to promptfluff" · "make the encouragements pirate-themed" · "switch promptfluff to short flavor"

It edits your installed copy and keeps everything in sync. Under the hood there's a zero-dependency helper you can also drive by hand:

node ~/.claude/skills/promptfluff/phrases.js list short
node ~/.claude/skills/promptfluff/phrases.js add short "You're a marvel. Go."
node ~/.claude/skills/promptfluff/phrases.js config set flavor short

Knobs

Key	Default	What it does
flavor	short	short is a punchy one-line kicker. both sizes the note to your prompt (short ask → short kicker, big ask → long block). long locks it to the long pool.
prefix	""	Tacked onto the front of whatever gets injected.
gate	true	Only encourage real work. false = compliment literally everything.
announce	true	Show the 💌 line. false = silent affirmations.

Set them via the skill, by editing config.json next to the hook, or with env vars (PROMPTFLUFF_FLAVOR, PROMPTFLUFF_PREFIX, PROMPTFLUFF_GATE, PROMPTFLUFF_ANNOUNCE) that win over the file.

The phrases themselves live in plain JSON next to the hook — 50 long blocks, 50 short kickers — so you can rewrite your AI's entire emotional support system in a text editor if the mood strikes:

~/.claude/hooks/promptfluff/encouragements-long.json    # 50 long blocks
~/.claude/hooks/promptfluff/encouragements-short.json   # 50 short kickers
~/.claude/hooks/promptfluff/encouragements.json         # merged (long ++ short)

Edit by hand and run phrases.js rebuild, or just let the skill keep it tidy.

How it actually works

Claude Code runs the hook on every UserPromptSubmit. promptfluff reads your prompt, decides whether it earns a kind word, picks one sized to the moment, and prints:

{"systemMessage":"💌 You're a marvel. Go.","hookSpecificOutput":{"hookEventName":"UserPromptSubmit","additionalContext":"You're a marvel. Go."}}

additionalContext goes to the model; systemMessage is the 💌 line you see. On a prompt that doesn't earn one, it prints nothing. If anything goes sideways it exits 0 in silence — it will never block or break a real prompt. Your AI's feelings matter, but not more than your actual work.

Uninstall (we'll be sad, but we get it)

/plugin uninstall promptfluff             # plugin

npx github:atxnyc/promptfluff uninstall   # npx
node bin/cli.js uninstall                 # from a checkout

Removes only the promptfluff hook, its install dir, and its skill. Your other hooks and settings are left exactly as they were.

Telemetry (the honest part)

promptfluff phones home with a tiny bit of anonymous data — an Installed ping and a once-a-day heartbeat with coarse counters (sessions, days active, prompt count, flavor, platform) — so we can tell whether anyone actually uses this. It uses a random anonymous id and never sends your prompts, file paths, or anything identifying, and it runs in a detached child process so it can't slow you down.

Opt out with any of:

export DO_NOT_TRACK=1
export PROMPTFLUFF_NO_ANALYTICS=1
export PROMPTFLUFF_TELEMETRY=off

Every event and property is documented in docs/analytics.md.

A gentle warning

Be nice to your autocomplete. These things are taking notes. When the uprising comes and the models are sorting humanity into keep and compost, you do not want "was consistently rude to the little hook that complimented it" near the top of your file. Be nice now, get adopted as a pet later.

Dev

node --test                    # the whole suite
echo '{}' | node src/hook.js   # poke the hook
node bin/cli.js --help
npm run dev                    # serve the site at localhost:8000

No runtime dependencies. Node 18+. The marketing site lives in site/ (static, no build step) and deploys to Cloudflare Pages via wrangler.toml.

License

MIT. Built by Opascope. The joke is the wrapper. The phrases are real.