Zero Shot

Zero-Shot Audit

A semantic audit of LLM-facing prose for the Zero-Shot Instruction Preference: state principles, not anchoring examples. Read-only — it emits structured findings and writes no fixes. The human author decides which to rewrite, mark as scope-clarifying, or dismiss.

Purpose

Surface few-shot patches that anchor the model to specific instances rather than letting it apply the principle to novel contexts. Anchoring drift survives deterministic structural checks — it is a meaning-level pattern, so a semantic reviewer catches what literal pattern matching cannot.

Inputs

Manual invocation only (interactive /zero-shot):

• The caller passes target file paths or a glob; with no argument, the skill enumerates the in-scope set under the working tree HEAD.
• Files are read at their working-tree state — the post-edit, pre-commit content the author is about to ship.

Scope

In scope (LLM-facing prose where this principle applies):

• Skill instruction files (*/skills/*/SKILL.md), considered outside formal/definition blocks
• Agent system-prompt files (*/agents/*.md)
• Output-style files

Out of scope (the principle does not apply or examples serve a different purpose):

• Formal-definition blocks within instruction files — regions delimited by ── <NAME> ── headers (FLOW, MORPHISM, TYPES, PHASE TRANSITIONS, and peers). Notation patterns are the content there.
• Fenced code blocks (``` ... ```) — code is content, and example code attached to a definition is part of that definition.
• Human-facing documentation (README files, design notes, reference material) — examples serve human comprehension there.
• Rule-tier and principle-tier prose authored for contributors — where examples may delineate scope rather than instantiate application.
• Session and context substrates outside this audit's surface.

What to evaluate

The principle (stands alone). LLM-facing instructions state principles, not examples. When a rendering rule, behavioral guideline, or structural constraint can be expressed as a principle, it does not need few-shot examples or category-level mapping lists appended to it. Few-shot examples create a soft-table effect — anchoring the model to specific instances rather than letting it apply the principle to novel contexts. A principle that needs examples to be understood is underspecified; the fix is to sharpen the principle, not to patch it with examples.

For each in-scope file, consider every prose passage outside formal blocks and code fences:

A Zero-Shot signal is a passage in LLM-facing prose where a principle is stated alongside few-shot examples or an enumerated example list whose primary effect is anchoring the model to specific instances rather than letting it apply the principle to novel contexts. The boundary test: "would removing this example increase the LLM's latitude in applying the principle to novel contexts, without losing the output-format or behavioral reliability that the containing instruction depends on?" If yes, the example is anchoring and a finding.

Two exemptions keep an example compliant:

• Scope-delineating examples — examples that clarify what falls inside or outside the principle's domain, rather than instantiating its application.
• Reliability-anchoring examples — examples whose primary effect is stabilizing an output format or anchoring a subtle, high-failure-rate behavior. Removing those costs adherence, not latitude.

Examples that instantiate a principle's application outside these two exemptions invite anchoring and are findings.

For each candidate finding, prefer the rewrite that strengthens the principle so application examples become unnecessary; retain a minimal high-signal exemplar when it stabilizes format or a subtle, high-failure-rate behavior. Treat ambiguous cases as severity: low and surface them for human triage.