When to use structured vs. natural language
| Use structured format when… | Use natural language when… |
|---|---|
| The task has multiple conditions or edge cases | The task is simple and unambiguous |
| Output format must be exact (picklist, date, score) | Detection or extraction is self-evident |
| You need precise silence and attribution rules | You’re iterating quickly and testing scope |
| You’re configuring a Summary or Coaching competency | You’re writing a Signal trigger or Smart Tag |
Silence patterns — by feature
| Feature | When nothing is found… |
|---|---|
| Autopilot (CRM field) | Return nothing. Silence = no write. Never fill with placeholders. If Momentum is the only writer to that field, output an explicit value (e.g. N/A) so you can validate the prompt ran. |
| Signal trigger | Output FALSE. Zero explanation. |
| Signal follow-up | Output an explicit “No [X] found.” — confirms the prompt ran and prevents a blank Slack block. |
| Summary section | Output a defined fallback (e.g. None identified). Never skip or leave a section blank. |
| Coaching competency | Note that the call was too short or lacked sufficient context to evaluate. |
Word precision cheat sheet
| Weak phrasing | Stronger alternative | Why it matters |
|---|---|---|
| ”You should…" | "You MUST…” | Suggestions can be overridden by the AI when uncertain. Hard constraints cannot. |
| ”Summarize" | "Extract” | Summarize invites paraphrase and interpretation. Extract pulls what’s explicitly stated. |
| ”Look for" | "Identify only” | Look for is a broad search that surfaces tangential mentions. Identify only narrows scope and reduces false positives. |
| ”If not found, say nothing" | "Complete silence is the correct output. Outputting any text when no qualifying data exists is a critical error.” | Weak silence rules are routinely ignored. Raising the stakes forces compliance. |
The five most common prompting mistakes
| Mistake | Fix |
|---|---|
Trigger prompt extracts content instead of TRUE/FALSE | Rewrite as binary detection. Output TRUE or FALSE only. |
| No silence rule defined | Add: “If [condition] not present, output nothing.” Or define an explicit value if intentional (see Silence Patterns above). |
| Picklist allows freeform output | List exact allowed values. State: return one value only, character-for-character. |
| Summary sections do too much | One extraction goal per section. Split complex ones in two. Keep to 4–7 sections total. |
| Coaching rubric uses vague language | Define observable behaviors — specific actions the rep took or didn’t take, not subjective impressions. |