Prompts

A prompt entry is a reusable, parameterized piece of text for an AI coding agent. Fill its placeholders in the launch menu, then fire it at the agent of your choice:

skit add review.prompt.md            # managed placeholders become input fields
skit run review                      # pick the agent, fill in the inputs, go
skit run review --runner codex --set target=src/app.py --no-input

Adding one

A .prompt.md (or .prompt) file is recognized as a prompt outright. A plain .md works too — interactively skit asks whether to add it as a prompt (defaulting to yes); non-interactively it never guesses, so pass --prompt. You can also draft one from scratch with skit add --prompt.

Placeholders

During interactive add review you choose which detected {{placeholders}} become input fields. Up to 30 are selected by default; when detection exceeds 30, none are selected by default — so code samples pasted into a prompt are not mistaken for variables. Managed fields get the full preset / last-values / --set experience.

There are no escape sequences to learn: anything that isn't a placeholder you manage — unmanaged {{holes}} included — travels to the agent byte-for-byte as written. A per-prompt switch (--no-interpolate, or one checkbox in Entry settings) turns insertion off entirely, for prompts that were never written with it in mind.

Runners

A runner is the agent CLI that receives the rendered prompt. Five come preconfigured — seeded on first use, then yours to edit:

The -- / = binding forms are deliberate: a prompt that happens to begin with --help or --model must not be parsed as a flag. All of these open the agent's interactive session with the prompt as the opening message — except amp, which has no such mode and uses its official amp -x one-shot to execute the prompt and return.

Register anything else with:

skit runner add mycli -- mycli run {{prompt}}

Template rules: each word is one argv token (no shell anywhere); exactly one token must be {{prompt}}; it can't be the first token (that's the binary); no other {{holes}} are allowed; single-brace text is literal.

How the runner is chosen:

• Interactive runs: picked in the launch menu, prefilled with your last pick. A pin you leave untouched doesn't count as a pick.
• Non-interactive runs: --runner beats the entry's pin; with neither, the run fails with exit 126. There is deliberately no global default and no ranking — the runner changes the result. The last pick only ever prefills pickers; it never resolves a --no-input run.
• Per-run agent flags ride after --: skit run review -- --model opus. The tail is remembered like any entry's — see Parameters & values.

Delivery

Running a prompt renders the text and invokes the selected runner's configured argv command with it. The rendered prompt travels as a single process argument — no shell in between, so nothing in the text is ever shell-interpreted.

Oversized prompts are refused before spawn with a clean error (exit 125) naming the measured size, instead of a raw OS failure: the assembled command line caps at 60,000 UTF-16 units on Windows and 100,000 bytes on POSIX — CJK text counts in bytes, not characters. A NUL byte in the rendered text is refused too.

Not a secrets channel

Whatever the rendered text contains ends up in the receiving agent's own session logs. Don't put secrets in prompts — no delivery mechanism here pretends otherwise.