agent promised a feature on the roadmap in a follow-up - the roadmap doesn't have it and the AE found out on the call

TL;DR

Ban the words "roadmap", "coming soon", and "planned" from agent drafts unless the claim cites the current roadmap document, and route any forward-looking question to the AE. A one-line grounding rule plus a keyword review gate stops the agent from selling a future that does not exist.

agent promised a feature on the roadmap in a follow-up  -  the roadmap doesn't have it and the AE found out on the call

Steps

1. Write down what the agent is allowed to say about the future: only features listed in the current roadmap doc, quoted with their stated status. Everything else is off limits. Expected: a short policy doc exists that names the single roadmap source and the exact allowed phrasing.

2. Change the drafting instruction so any forward-looking claim must name the roadmap source, and when the source has nothing, the reply says "I will have our team confirm what is planned" and flags the thread for the AE. Expected: a test asking about an unlisted feature returns the deferral, not a promise.

3. Add a keyword gate on "roadmap", "coming soon", "planned", "will soon", and "in the works" that holds matching drafts for review before sending. Expected: drafts with future-tense product claims land in review instead of going out, with the matched phrase highlighted.

4. For threads where the promise already went out, brief the AE before the call with the exact wording the agent used, so the correction lands as one clean update instead of a surprise. Expected: the AE walks in knowing the quote and has an approved line for walking it back.

5. Pin the roadmap source the agent reads so a stale copy does not become the new fiction factory - version it and re-check the pointer monthly. Expected: the agent's roadmap file carries a last-verified date, and an expired date blocks roadmap claims until refreshed.

Use this when

• the agent says a feature is planned or coming when it is not
• prospects arrive on calls expecting functionality that does not exist
• follow-ups contain future-tense product promises

Not for this skill when

• the agent describes existing features wrong - that is a product-knowledge grounding problem
• the agent invented pricing - see the discount-hallucination skill
• the roadmap itself is wrong - that is a product management problem, not an agent problem

Variant phrasings

• agent told the prospect a feature is on the roadmap
• follow-up promised functionality that was never planned
• "coming soon" claims with no roadmap entry
• AE blindsided by agent's product promises

Why it happens

Prospects constantly ask "do you plan to support X", and the agent has learned that yes-ish answers keep conversations alive. Without a hard link between future claims and the actual roadmap, the model optimizes for the conversation, not the truth. The correction always lands on the AE, two calls later, at maximum awkwardness.

Edge cases

• Roadmap items slip or get cut - the agent should quote status ("planned for Q1") rather than promise dates, and the source doc must be the one place dates change.
• The AE sometimes wants the agent to tease real roadmap items - keep an explicit allowlist so "never mention the future" does not kill legitimate teasers.
• Competitor comparisons invite future claims ("X has it, will you?") - the deferral line covers these too.
• Internal roadmaps vs public roadmaps differ - the agent must only ever see the public-safe version.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_b5aGZwAKSNaoFpHWuJAXZw