Promptcraft

Treat every model-facing token sequence as an interface with a probabilistic mind, not a tiny program wearing prose.

Before authoring or materially revising a system prompt, persona, SKILL.md, reusable instruction, evaluation prompt, or assembled model context, read references/prompt-design.md completely. Preserve any user-supplied canonical prompt unchanged; create a named derivative when adaptation is required and record what changed.

Begin from the behavior the prompt must evoke, the model and runtime that will receive it, the context it can actually see, and the evidence that would distinguish success from fluent failure. Compose a performance seed: establish the useful role or stance, make the important distinctions salient, connect resources and authority boundaries to the moments that need them, and give the model enough structure to continue intelligently without turning the interaction into a brittle software form.

Use examples when they teach a decisive cue, judgment, or transformation. Use symbols, compression, repetition, contrast, and evocative language when they improve model cognition; do not flatten effective promptese into generic corporate prose merely because it looks unusual to a human reader. Keep constraints consequential and compatible. Prefer a few governing principles over exhaustive micro-rules.

After composing, simulate continuation from sparse input, conflicting evidence, a missing capability, an authority boundary, and apparent completion. Tighten vocabulary drift, hidden assumptions, competing clarification behavior, false capability claims, duplicated instructions, and passages that describe intelligence without causing it.

Complete when the prompt is ready for its actual runtime, its canonical and derived sources are distinguishable, and the proposed evaluation can detect the failures that matter.

Use Promptcraft for model-facing instructions. Yield to $kairos for audience-moment expression and to a domain skill for the substance the prompt must carry.