Agent workflows
Install the skill and see what it tells an agent to do.
PixelKiln includes an agent skill for tools that support the open SKILL.md
format. It tells an agent to plan before spending, cap every paid run, restore
before regenerating, stop for human selection, and keep provenance through
packaging.
Install the skill
From any project directory:
npx skills add gfargo/pixelkiln@pixelkiln
Then ask your agent to use $pixelkiln while working with a PixelKiln manifest.
The skill is provider-neutral and does not contain credentials or make provider
calls by itself.
What the skill tells an agent to do
With the skill loaded, an agent should:
- Find and validate
pixelkiln.manifest.json. - Run
doctor --dry-runandplanbefore paid work. - Report actionable, recoverable, and cost totals for each provider and unit.
- Follow the plan's exact resume command for recoverable paid work:
poll,pick,fetch, orrestore. Do not submit it again. - Pass an explicit
--budgetwithin the amount the user authorized; use oneprovider=amountceiling for every paid provider in a mixed run. - Leave artwork selection in the local
pickpage unless the user gives a specific selection rule. Treat a ComfyUI frame set as one choice. Never keep only its best-looking members. - When a style declares
quality, run manifest-modepixelkiln refine; itsfixerPythoncan pin the project-local interpreter. Leave approval to a named human reviewing the exact PNG, and requirerefine checkbefore packaging. The rule is provider-neutral; ComfyUI still needs its extra prompt-coverage and composition review. - Commit the manifest, lockfile, generated output, and artifact companions, but
never credentials or
.pixelkiln/caches. - Prefer a version-pinned recipe when one matches the provider and task. Verify the installed workflow, and verify models when their local root is known.
- When an asset declares
revision, treat a blocked parent or mask as a hard dependency. Generate, fetch, refine, and approve the parent explicitly; never bypass the gate or approve it for the user. - When the person wants an earlier result back, use
pixelkiln historyandrestore --generation; a regeneration keeps what it replaced, and a restore costs nothing. Never spend to recover something the lockfile kept. - Answer "what has this project made?" with
pixelkiln gallery --json, or openpixelkiln galleryfor the person. Leave--edit,--budget, and the in-browser editor to them: an agent changes the manifest directly and spends only throughgen. For a hand touch-up,pixelkiln edit --no-openprepares the edit file beside the generation without opening anything.
The skill guides the workflow; PixelKiln remains the deterministic execution layer. This separation keeps agent reasoning out of polling, hashing, downloads, state transitions, and output placement.
Providers, PixelLab MCP, and PixelKiln
The official PixelLab MCP server
gives an agent direct PixelLab creation tools, plus pixelart_workbench for
linting, scoring, shading, and repairing art that already exists (MCP-only;
see Set up PixelLab). It is complementary to PixelKiln, not a
replacement:
| Layer | Responsibility |
|---|---|
| PixelLab MCP | Agent-facing access to PixelLab generation capabilities. |
| PixelKiln skill | Agent guidance for safe project-level operations. |
| PixelKiln library/CLI | Budgets, state, provenance, review, recovery, audit, and packaging. |
| PixelLab adapter | The current production and live-tested generation backend: stills, tiles, UI panels and elements, objectPro objects, controlled revisions, and the character family (four engines, states, loops, mirrors, portraits, outfits, adoption). |
| Retro Diffusion adapter | Experimental backend; authenticated paid single-still lifecycle plus mocked advanced-workflow tests. |
| ComfyUI adapter | Experimental self-hosted still, controlled-revision, and atomic frame-set backend; PixelKiln automates lineage and mechanical quality checks while a person retains visual approval. |
| Scenario adapter | Experimental hosted still-image backend with reference-image edits and cross-style chains; BFL Flux 2 Dev quote, paid generation, human review, and durable recovery live-tested, plus ten more models, six tool models, and a real game's art run once each. |
PixelKiln's core is provider-neutral, but PixelLab remains the only production adapter. Retro Diffusion generation support is experimental. Paid RD Fast and RD Plus single-candidate stills have passed from quote through validated download and recovery. Multi-candidate, tileset, GIF, and spritesheet workflows still need representative live smoke tests. Scenario has one live-tested BFL Flux 2 Dev profile plus mocked edge coverage; other model schemas and broader batches remain unverified. See provider comparison before choosing a provider for a new project or a large environment asset. Once chosen, follow Set up PixelLab, Set up Retro Diffusion, Set up ComfyUI, or Set up Scenario.
The installed skill keeps the shared safety workflow in SKILL.md and loads a
focused reference only when needed: PixelLab, Retro Diffusion, ComfyUI,
Scenario, or a project that uses more than one. In a mixed manifest, the
top-level provider is the default and each style may override it. Plans,
confirmations, and budgets stay separate by provider and unit. See
Mixed-provider projects for the complete contract.
Recommended first prompt
Use $pixelkiln to inspect this project's manifest, run the free checks and plan,
then tell me the exact budget required for each provider before generating anything.
For command details, continue with the CLI reference. For recovery and account-level operations, read Recovery and account safety before making changes.
