lead asked 'how is this different from X' and the agent's answer described X's features as ours - the reply went out...

TL;DR

Do not let the agent answer "how are you different from X" from memory. Give it a sourced battlecard per competitor, require every comparative claim to cite the card, and when there is no card for X, the agent defers to the AE instead of improvising a comparison.

lead asked 'how is this different from X' and the agent's answer described X's features as ours  -  the reply went out before anyone reviewed

Steps

1. Build one battlecard per competitor you actually face: what they do, what you do differently, and what to say when asked - all sourced, dated, and owned by product marketing. Expected: a card exists for each top competitor, and the agent can answer only from the card, not from training data.

2. Rewrite the drafting instruction: comparative claims must cite the battlecard; if no card covers the named competitor, the reply says "great question - I will get you an accurate comparison" and flags the AE. Expected: a test naming an unlisted competitor returns the deferral, never an invented comparison.

3. Add a review gate on competitor names: any draft mentioning a competitor goes to a human or a checker step before sending. Expected: comparison drafts wait in review, and the gate log shows which competitor triggered it.

4. Audit recent sent threads for competitor mentions to find comparisons that already went out wrong, and brief the AEs who own those threads. Expected: a list of affected threads with the exact claims made, so each AE can correct the record on the next touch.

5. Refresh battlecards on a schedule - competitor features change, and a stale card becomes the same lie with better sourcing. Expected: each card carries a last-verified date and an owner who re-confirms it quarterly.

Use this when

• the agent describes a competitor's features as your own (or vice versa)
• differentiation answers go out with no sourced basis
• a competitor name in a draft should trigger a pause

Not for this skill when

• the agent misstates your own features without any competitor involved - that is product grounding
• the agent invents pricing - see the discount-hallucination skill
• you need the battlecards written - that is a product marketing task

Variant phrasings

• agent mixed up our features with the competitor's
• "how are you different from X" answered from memory
• false comparison sent to a prospect
• competitor features attributed to us in a reply

Why it happens

Differentiation questions feel like knowledge questions, and the model has plenty of training data about competitors - most of it outdated, some of it about similarly-named products. It assembles a confident answer from fragments, and the fragments do not respect which company is which. The card removes the assembly step: the answer is retrieved, not composed.

Edge cases

• The prospect names a competitor with a similar name to yours - the deferral path is safer than guessing which one they meant.
• Battlecards can leak if the agent pastes too much - keep cards to talk tracks, not internal strategy docs.
• A prospect corrects the agent's comparison mid-thread - treat the correction as a signal to review the card, not as ground truth to adopt blindly.
• New competitors appear faster than cards get written - the deferral fallback is the default for a reason; use it.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_gmdBVdgbHJN9Mpv89muV5g