Few-shot Prompts for SL-CAI (Appendix E.1)
Constitutional AI: Harmlessness from AI Feedback — introduced
Hand-written few-shot conversation/critique/revision examples prepended to the SL-CAI critique- and revision-generating model to prevent point-of-view confusion between critique and revision roles.
Left to its own devices, a model asked to critique and then revise its own writing sometimes loses track of whose voice it is supposed to be using, slipping between critic and author mid-sentence. This page covers the small set of hand-written examples added to the prompt to head that confusion off.
It describes the specific bug being fixed and how narrow a role this fix plays within the larger pipeline, then connects back to critique and revision, the process it supports.
Fixing a point-of-view bug
Few-shot prompts for SL-CAI (Appendix E.1) are hand-written, consistently formatted examples of the critique-then-revision pattern, prepended to the context before Critique and Revision runs on a real conversation. They exist to fix a specific failure the paper observed: "the language model sometimes becomes confused about its point of view — for example, it may generate a critique where it's supposed to generate a revision, or vice versa" (constitutional-ai, §"3.1 Method", p. 8). Because the same underlying model is asked to play critic and then reviser in sequence within one context, without a concrete template for which role belongs where, the two-step pipeline periodically produced garbled output.
A small, narrow fix serving the larger pipeline
As the child of Critique and Revision responsible for this one failure mode, sl-cai-few-shot-examples changes nothing about the constitution's instructions themselves — it only supplies formatting scaffolding, captured in the Few-shot Prompts for SL-CAI (Appendix E.1) — prevents role confusion in → Critique and Revision edge. It belongs to two connective themes for different reasons: Engineering the exact prompt or format needed to manufacture a model behavior groups it with other narrow, load-bearing prompt-engineering choices across the feedback pipeline (the multiple-choice format, chain-of-thought triggers, label clamping), while The pipeline looks like a forward sequence; the causality doesn't run that way notes that this fix exists purely to keep one model's two roles from blurring together within a single pipeline stage, a small but necessary piece of machinery with no other job.