A fast TUI and CLI for searching every prompt you've sent to your coding agents.
Supports Claude Code, Codex (TUI and GUI), and Pi out of the box. Anything else is a few lines of config away. Built with OpenTUI, runs on Bun.
- Live search across every session, with matches highlighted in the list and the detail pane
- Filter by source and model. Model tabs come from your own history, newest first
- Star the good ones and find them later in Favorites
pk lsfor scripts and agents: the same data, as plain text or JSONL- Make it yours: add your own filters and sources in one
config.ts, built on the same API as the built-ins
| Search, highlighted | Filter by model |
|---|---|
![]() |
![]() |
| Pick from every model you've used | Favorites |
![]() |
![]() |
Only prompts you actually typed. Tool results, slash-command
expansions, skill blocks, injected file/branch/compaction context, system
reminders, and shell (!) commands are all filtered out at parse time.
Two more kinds are kept but hidden unless you ask for them:
| Tag | Meaning | Opt in |
|---|---|---|
agent |
written by an agent or app, not you (subagent transcripts, SDK sessions) | --agents |
unsent |
never got a reply (session killed, quit, or errored) | --unsent |
| Source | Location |
|---|---|
| Claude Code | ~/.claude/projects/*/*.jsonl |
| Codex | ~/.codex/sessions/**/*.jsonl |
| Pi | ~/.pi/agent/sessions/*/*.jsonl |
The model a prompt ran on is read from the assistant reply (Claude/Pi) or the turn context (Codex).
git clone https://github.com/avgvstvs96/prompt-picker
cd prompt-picker
bun start| Key | Action |
|---|---|
| type | search prompt text (live) |
↑ ↓ / Ctrl-P Ctrl-N |
move selection |
PgUp PgDn |
page |
Tab / Shift-Tab |
cycle source: All · Claude · Codex · Pi · ★ Favorites |
← → |
cycle model filter (derived from what's loaded, most recent first) |
Enter / Ctrl-S |
star / unstar the selected prompt |
Enter on Other chip |
open the model picker (type to filter, ↑ ↓ select, Enter apply) |
Ctrl-U |
clear search |
Esc / Ctrl-C |
close the model picker if open, otherwise quit |
Note
The model filter tabs are derived from whatever models show up in the
current view (any source, including All and Favorites) and only appear
once 2+ distinct models are present, ordered by most recent use. Beyond
the first 3 they fold behind an Other chip — Enter on it opens a
searchable picker of every model in the view.
Installs two names for the same command: prompts (explicit) and pk
(short). Use whichever fits; every example below works with either.
pk on a real terminal launches the TUI. Piped or redirected (how LLMs run
it), it lists recent prompts and exits — built for agents to grep/read
without spending context on a UI. Any positional args, or an explicit ls,
also force list mode.
pk # TTY: TUI · piped: same as `pk ls`
pk fix the flaky test # search
pk ls -s 24h # last day, newest first
pk ls -s 3 # bare number = days, same as -s 3d
pk ls --source codex # one source
pk ls -m opus # model filter, see rule below
pk ls --json --since 1w > out.jsonl
pk ls -s 1 -c # compact headersFlags: --since/-s <n>h|d|w|<n> (default 3d; a bare number means days —
-s 3 ≡ -s 3d), --source <id>, --model/-m <q>, --json (JSONL, one
Prompt per line), --raw (skip config filters — custom sources still
load), --compact/-c (session id only, no --json), --agents (include
prompts written by agents and apps — subagent transcripts, SDK-driven
sessions — hidden by default and marked agent in the header), --unsent
(include prompts that never got a reply — killed, quit, or errored
sessions — hidden by default and marked unsent). Results are capped at
100, newest/most-relevant first.
Each prompt is printed under a one-line header:
── claude · Fable 5 · prompt-picker · Jul 21 14:32 · 019f4aec-4067-426b-95da-05acbc07b563
prompt text...
source · model · project · short date/time · session id. The model is
omitted when -m/--model narrowed the results (redundant at that point).
--compact shrinks the header to just the session id, prefixed with the
source unless --source was already given:
── claude · 019f4aec-4067-426b-95da-05acbc07b563 # default compact
── 019f4aec-4067-426b-95da-05acbc07b563 # --source claude -c
--model matches substrings of the model label (-, ., and space are
interchangeable, so opus-4.7 and opus 4.7 are the same query). A query
with no digits keeps only the highest matching version — -m gpt picks the
newest GPT release, not every one you've ever used.
Config filters (see below) apply by default; --raw skips them.
Everything is configured from one file, ~/.config/prompt-picker/config.ts.
It's plain TypeScript, so a filter or a source can be anything you can write
as a function.
Claude, Codex, and Pi aren't special. They're declared with the same
defineFileSource / makePrompt API your config gets, so anything a
built-in can do, your config can do too. The demo above is an example:
its config swaps out the built-ins
for a source of fake prompts.
Export a filters array of predicates. A prompt is shown only if
every filter returns true for it.
// ~/.config/prompt-picker/config.ts
import type { Filter } from "prompt-picker";
export const filters: Filter[] = [
(p) => p.text.length > 8, // drop throwaway prompts
(p) => p.ts > Date.now() - 90 * 864e5, // only the last 90 days
(p) => p.project !== "client-work", // keep one repo out of view
];Each predicate receives a full Prompt (text, source,
model, ts, project, …), so you can filter on anything.
Filters run at load time and never touch the cache, so edits take effect on the next launch with no reindex. A broken config is reported to stderr and otherwise ignored.
Using an agent that isn't built in? Teach prompt-picker to read it. Export a
config factory from config.ts: it receives the source API and returns
{ filters, sources, includeBuiltins }.
defineFileSource scans files on disk and parses them into prompts. Use
makePrompt to fill in stable ids, timestamps, project names, and model
labels for you:
// ~/.config/prompt-picker/config.ts
import type { ConfigApi } from "prompt-picker";
export default ({ defineFileSource, makePrompt }: ConfigApi) => ({
sources: [
defineFileSource({
id: "cursor",
label: "Cursor",
color: "#7aa2f7", // optional tab/badge color
root: "~/.cursor/chats",
glob: "**/*.jsonl",
parse({ file, raw }) {
return raw
.split("\n")
.flatMap((line, lineNumber) => {
if (!line.trim()) return [];
const event = JSON.parse(line);
if (event.role !== "user") return [];
return makePrompt({
source: "cursor",
file,
line: lineNumber,
text: event.content,
ts: event.timestamp,
cwd: event.cwd,
sessionId: event.session_id,
model: event.model, // formatted into a display label
});
});
},
}),
],
});For prompts that don't come from files on disk (generated data, a SQLite
history, an API), use defineSource and return the prompts directly:
import type { ConfigApi } from "prompt-picker";
export default ({ defineSource, makePrompt }: ConfigApi) => ({
sources: [
defineSource({
id: "notes",
label: "Notes",
async load() {
return [
makePrompt({
source: "notes",
file: "manual",
text: "Review this architecture for accidental complexity.",
ts: Date.now(),
}),
];
},
}),
],
});Custom sources are merged after the built-ins and each gets its own tab. The
model filter tabs appear automatically once a view has 2+ distinct models,
no configuration needed. Return includeBuiltins: false to replace the
built-in sources entirely — useful for a demo dataset:
export default ({ defineSource, makePrompt }: ConfigApi) => {
const demo = Bun.argv.includes("--fake-prompts");
return {
includeBuiltins: !demo,
sources: demo ? [/* defineSource(...) */] : [],
};
};File sources are cached by size + mtime per file; defineSource loaders run on
every launch. A broken source is reported to stderr and skipped.
Parsed prompts are cached at ~/.config/prompt-picker/index-cache.json, keyed
by each source file's size + mtime, so only changed sessions are re-parsed on
launch. Search and filtering run in-memory over the full set on every keystroke.
Favorites persist to ~/.config/prompt-picker/favorites.json.
Built by Bassim




