How do I install the Sprites skill?

Run npx skillmds@latest add phrazzld/sprites in your terminal (requires Node.js), paste this page's agent-chat prompt into Claude, Cursor, or any MCP-connected agent, or download the SKILL.md file and copy it into your agent's skills directory.

Is Sprites safe to use?

This skill has not completed SkillMD's automated safety review yet. SkillMD never runs a skill's scripts for you; review the SKILL.md before installing.

Which AI agents work with Sprites?

This skill is tagged as working with Claude Code, Claude.ai, OpenAI Codex. SKILL.md is an open format, so most agents that read a skills directory can load it too.

Is Sprites free to use?

Yes. Installing skills from SkillMD is free, and the skill stays under its author's original license.

Who published Sprites?

phrazzld (@phrazzld) published this skill. Their other Agent Skills are listed on their SkillMD profile.

Sprites by phrazzld — Agent Skill

Name: Sprites
Author: phrazzld

Sprites

Run lane cards on Fly Sprites: remote, isolated, scale-to-zero sandboxes for heavy or parallel agent work. Golden-checkpoint provisioning so lanes start on a ready sprite with zero setup tokens. Use when: "run this on a sprite", "remote lane", "offload to a sandbox", "dispatch to sprites", "bake a sprite", "sprite fleet", heavy/long-running/parallel sub-agent work that should not run on this machine. Trigger: /sprites, /sprite-lane.

phrazzld 8dba108 5 files · 16.3 KB Updated Sep 17, 2026

File contents

/sprites

One primitive: scripts/sprite-lane runs one lane card on one sprite and writes a local receipt. No daemon, no fleet state. Both ad-hoc dispatch (an orchestrator calling this skill) and event-driven intake (a webhook or cron caller) use the same primitive.

Route a lane: local subagent or sprite?

Lane shape	Run it
Quick exploration, repo Q&A, small review	Local native subagent
Massive parallel fan-out within one session (audits, migrations, cross-checked review)	The harness's own large-scale orchestration feature when it has one — sprites add network overhead for nothing there
Heavy build/test loops, big implementation tickets	Sprite
Long-running or detached work (hours, overnight)	Sprite, `--detach`
Parallel competing attempts on the same repo	One sprite per attempt
Anything needing isolation from this machine	Sprite

Do not prescribe sprites for work a local subagent finishes faster; network and sync overhead make sprites worse for sub-minute lanes.

Commands

scripts/sprite-lane bake <sprite>                 # provision + golden checkpoint (idempotent)
scripts/sprite-lane run <sprite> --repo <owner/name|url> --card <file> \
    [--branch <b>] [--harness codex|claude] [--detach] [--fresh]
scripts/sprite-lane status <sprite>               # bake state + running lanes
scripts/sprite-lane fetch <sprite> <lane-id>      # pull a detached lane's log
scripts/sprite-lane reset <sprite>                # restore golden checkpoint

Receipts land in ~/.harness-kit/receipts/sprite-lane/<lane-id>.json with the remote log path, exit code, and timing.

Provisioning model

Sprites cannot fork from another sprite's checkpoint, so determinism comes from: an idempotent bake (harness CLIs, auth, git identity) plus a per-sprite golden checkpoint. run auto-bakes an unbaked sprite; --fresh restores golden first so the lane starts clean. Bumping BAKE_VERSION in the script invalidates stale bakes everywhere. Details, auth flows, and failure modes: references/provisioning.md.

Lane cards

A lane card states: end state, success criteria, verification affordances, boundaries, and expected output shape (template: templates/lane-card.md). Big outcome-shaped tickets beat atomic tasks — the agent on the sprite owns its own decomposition. Put the oracle in the card; the sprite has no access to your local context.

Gotchas

The card is the entire context the remote agent gets. No local files, no conversation history. Inline or upload everything the lane needs.
--fresh wipes uncommitted work from previous lanes on that sprite. Fetch anything you need first.
Codex lanes reuse local ~/.codex/auth.json; if lanes start failing with refresh_token_reused or auth_error, re-run codex login locally and re-bake.
Claude lanes need ANTHROPIC_API_KEY exported locally at run time.
A checkpoint with mount_failed health is unusable; re-bake and cut a new golden.
Scale-to-zero is Fly's job. Do not build keepalives or reconcilers around sprites; wake-on-exec is the contract.

---name: spritesdescription: Run lane cards on Fly Sprites: remote, isolated, scale-to-zero sandboxes for heavy or parallel agent work. Golden-checkpoint provisioning so lanes start on a ready sprite with zero setup tokens. Use when: "run this on a sprite", "remote lane", "offload to a sandbox", "dispatch to sprites", "bake a sprite", "sprite fleet", heavy/long-running/parallel sub-agent work that should not run on this machine. Trigger: /sprites, /sprite-lane.---# /spritesOne primitive: `scripts/sprite-lane` runs one lane card on one sprite andwrites a local receipt. No daemon, no fleet state. Both ad-hoc dispatch (anorchestrator calling this skill) and event-driven intake (a webhook or croncaller) use the same primitive.## Route a lane: local subagent or sprite?| Lane shape | Run it ||---|---|| Quick exploration, repo Q&A, small review | Local native subagent || Massive parallel fan-out within one session (audits, migrations, cross-checked review) | The harness's own large-scale orchestration feature when it has one — sprites add network overhead for nothing there || Heavy build/test loops, big implementation tickets | Sprite || Long-running or detached work (hours, overnight) | Sprite, `--detach` || Parallel competing attempts on the same repo | One sprite per attempt || Anything needing isolation from this machine | Sprite |Do not prescribe sprites for work a local subagent finishes faster; networkand sync overhead make sprites worse for sub-minute lanes.## Commands```bashscripts/sprite-lane bake <sprite> # provision + golden checkpoint (idempotent)scripts/sprite-lane run <sprite> --repo <owner/name|url> --card <file> \ [--branch <b>] [--harness codex|claude] [--detach] [--fresh]scripts/sprite-lane status <sprite> # bake state + running lanesscripts/sprite-lane fetch <sprite> <lane-id> # pull a detached lane's logscripts/sprite-lane reset <sprite> # restore golden checkpoint```Receipts land in `~/.harness-kit/receipts/sprite-lane/<lane-id>.json` with theremote log path, exit code, and timing.## Provisioning modelSprites cannot fork from another sprite's checkpoint, so determinism comesfrom: an idempotent bake (harness CLIs, auth, git identity) plus a per-sprite**golden checkpoint**. `run` auto-bakes an unbaked sprite; `--fresh` restoresgolden first so the lane starts clean. Bumping `BAKE_VERSION` in the scriptinvalidates stale bakes everywhere. Details, auth flows, and failure modes:`references/provisioning.md`.## Lane cardsA lane card states: end state, success criteria, verification affordances,boundaries, and expected output shape (template:`templates/lane-card.md`). Big outcome-shaped tickets beat atomic tasks —the agent on the sprite owns its own decomposition. Put the oracle in thecard; the sprite has no access to your local context.## Gotchas- The card is the entire context the remote agent gets. No local files, no conversation history. Inline or upload everything the lane needs.- `--fresh` wipes uncommitted work from previous lanes on that sprite. Fetch anything you need first.- Codex lanes reuse local `~/.codex/auth.json`; if lanes start failing with `refresh_token_reused` or `auth_error`, re-run `codex login` locally and re-bake.- Claude lanes need `ANTHROPIC_API_KEY` exported locally at `run` time.- A checkpoint with `mount_failed` health is unusable; re-bake and cut a new golden.- Scale-to-zero is Fly's job. Do not build keepalives or reconcilers around sprites; wake-on-exec is the contract.

Frequently asked questions