| name | prompt-pattern-selection |
| description | Choose the prompting pattern — zero-shot, few-shot, chain-of-thought, decomposition/chaining, role framing, or self-consistency — by tracing the task against reliability need and token/latency cost. Reach for this when a prompt is inconsistent, when you're about to add examples 'just in case', or when one prompt is quietly doing several jobs. Pairs with structured-output-design. |
Skill: Prompting-pattern selection
Pick the cheapest pattern that clears the reliability bar. Every step up the
ladder buys reliability with tokens and/or latency — so start low and climb only
on evidence of failure.
Step 0 — One opinion up front
Default to zero-shot. A clear, well-ordered instruction solves more than people
expect. Reach for examples, CoT, or chaining only when a zero-shot prompt measurably
fails the hard cases — not preemptively.
Step 1 — Characterize the failure
Run the prompt on the hard/edge cases and name how it fails:
- Format/style drift the model can't infer from words → few-shot.
- Wrong multi-step reasoning → CoT (one chain) or decomposition (separate jobs).
- Wrong persona/domain register → role framing.
- Occasional wrong high-stakes answer → self-consistency (sample N, vote).
- Doing several jobs at once → decomposition, always.
Step 2 — Trace the tree
Traverse ../../knowledge/prompt-decision-trees.md §1
to a leaf. Record the path and the runner-up.
Step 3 — Price the choice
Name what the pattern costs: few-shot adds input tokens; CoT adds output tokens +
latency; decomposition adds calls (latency, orchestration); self-consistency
multiplies cost by N. If the cost isn't justified by the reliability gain, drop back.
Step 4 — Verify empirically
Run the chosen pattern against the regression set (owned by
prompt-reliability-engineer). Keep it only if it beats the simpler pattern on
evidence, not on a single lucky example.
Step 5 — Hand off
- The wording + example curation →
prompt-implementation-engineer.
- The eval + regression gate →
prompt-reliability-engineer.
- If wording can't fix it, the problem is architectural →
prompt-architect.
Output
A pattern recommendation with the tree path, the runner-up and why it lost, the
token/latency cost named, and the empirical check that justified it.
