Skip to content

generation-loop.md

Latest commit

 

History

History
117 lines (86 loc) · 4.75 KB

generation-loop.md

File metadata and controls

117 lines (86 loc) · 4.75 KB

Generation Loop

Ghost sits as pipeline infrastructure for AI-driven UI generation. The expression.md is the grounding input; the review recipe is the post-generation gate; the verify recipe drives the loop over a prompt suite to expose where the expression leaks.

Only the grounding step is a deterministic CLI verb (ghost-drift emit context-bundle). Generate, review, and verify are skill recipes the host agent follows — installed with ghost-drift emit skill.

Pipeline shape

expression.md  ──►  [ghost-drift emit context-bundle]  ──►  SKILL.md / tokens.css / prompt.md
                                              │
                                              ▼
                                       any generator
                                  (host agent, Cursor, v0,
                                   in-house tool)
                                              │
                                              ▼ HTML / JSX
                                       [review recipe]  ──►  drift disposition
                                                             (block / annotate
                                                              / ack / track)

Pieces

ghost-drift emit context-bundle [flags] — the one CLI verb

Emit a grounding bundle any generator can consume. Default output writes SKILL.md + expression.md + tokens.css into ./ghost-context/.

Flags:

  • --out <dir> — output directory (default: ./ghost-context)
  • --prompt-only — single prompt.md only; skips SKILL.md / expression.md / tokens.css
  • --no-tokens — skip tokens.css
  • --readme — include README.md
  • --name <name> — override the skill name (default: expression id)

Point a Claude Code or MCP client at the output directory and the agent reads SKILL.md.

The generate recipe

Driven by the host agent. Loads the expression, builds a system prompt from Character/Signature/Decisions + tokens, asks the underlying model, extracts the artifact (HTML/JSX/etc.), and hands it to the review recipe for self-check. Retries with drift feedback until it passes or the agent gives up.

Not a replacement for Cursor / v0 / in-house tools. It exists so the loop is provable end-to-end, and so the verify recipe has something to drive.

Source: packages/ghost-drift/src/skill-bundle/references/generate.md.

The review recipe

The agent diffs generated output against the expression. Flags hardcoded colors outside the palette, spacing off the scale, and type choices that violate decisions. For pre-baked, per-project review commands use ghost-drift emit review-command (which writes a slash command at .claude/commands/design-review.md).

Source: packages/ghost-drift/src/skill-bundle/references/review.md.

The verify recipe

Runs the generate→review loop over a versioned prompt suite. Aggregates drift per dimension and classifies:

  • tight (mean < 1): expression reproduces faithfully
  • leaky (1–3): generator drifts here often — tighten Decisions
  • uncaptured (≥ 3): expression likely under-specifies this dimension

The killer demo: run verify on a mature expression, intentionally drop a section (e.g. motion), re-run, watch drift rise in dimensions that lost grounding.

Source: packages/ghost-drift/src/skill-bundle/references/verify.md.

The standard prompt suite

A versioned set of UI-construction tasks, each tagged with the expression dimensions it stresses. Tagging prompts with dimensions lets the agent distinguish targeted drift (a pricing-page prompt leaking spacing) from incidental drift (the same prompt leaking color, which it wasn't supposed to stress).

How the three-layer expression format earns its keep

Each layer has a concrete job somewhere in the loop:

Layer Role in the loop
Character Prompt context — shapes feel
Signature Drift-sensitive moves the reviewer weights heavily
Decisions Lookup table the generator consults for specific choices

If a layer doesn't pull weight somewhere, that's a signal the format is over-specified. The verify recipe is the schema-discipline mechanism.

Integration patterns

CI: a per-project design-review slash command emitted from ghost-drift emit review-command, invoked by the host agent as a required check on PRs that touch UI files.

In a generation pipeline: ghost-drift emit context-bundle writes the skill bundle into the generator's context; the generator produces; the review recipe gates the output. Drift disposition belongs to the pipeline owner (block, annotate, require ghost-drift ack).

Expression maintenance: run verify periodically. When a dimension shows up consistently leaky, the expression needs more Decisions for that dimension.