# Runbook — the PDF, link, video or audio never reaches the lead > **Symptom.** *"no se envía el pdf que configuré como recurso"* · *"el link no llega"* · *"manda un > texto raro con corchetes en vez del audio"* · *"the freebie never goes out"*. ## What you need first - The resource, and one conversation where it should have been delivered. - What the lead actually received, verbatim. The exact text separates half of these hypotheses. > **First, check the asset type can be sent at all.** A **link** is just text and always works. An > **image** or an **audio** clip the agent sends itself through `send_resource`, but only while > `config.resourceSending` is enabled. A **PDF, document or video** has no agent-side delivery path > in any channel — the tool rejects it — so "no se envía el pdf que configuré como recurso" is > usually not a failure to diagnose: those go out as a **link**, or through a **follow-up > workflow**, and `knowledge/how-to/send-a-resource.md` has both sequences. Come back here once the > asset is one the agent could have sent. --- ### H1 — A media tag was emitted as literal text Distinctive and instantly recognisable: the lead receives something like `[AUDIO: file.mp3]` as a text message. The prompt was written to output a tag the platform does not interpret, so it goes out as plain text. - **Test:** look at what the lead received **and confirm it was a plain-text message**. The same bracketed string is also what the platform stores in the transcript in place of a media message that has no transcription — there the audio *was* delivered and there is nothing to fix on the send (only the missing transcription, see the how-to). Only a text outbound is this hypothesis. - **Stop condition:** the tag went out as an outbound text message → **you're done**. - **Fix:** remove the tag from the prompt. Media has specific delivery paths and none of them is a marker in the text. An **image or audio** clip is sent by the agent itself, but only once `config.resourceSending` is enabled and the clip is agent-usable — check that before anything else. A **PDF, document or video** has no agent-side path at all: it goes out as a link, or through a follow-up with a static message. Check what the asset type supports before promising it in the prompt. *Evidence: a prompt instructing the model to output an audio tag inline; every lead received the raw tag instead of the audio.* ### H2 — The agent announced it and never sent it The agent says "te paso el link" and no URL appears. Reads as a delivery failure; is a flow bug. - **Test:** read three conversations that reached the moment and check whether the resource itself is present — a URL for a link or a PDF, the media itself for audio or video. An announcement is not a delivery. - **Stop condition:** the resource is discussed but never present → **you're done**. - **Fix:** the step must deliver in the same turn it announces. Demonstrate that in examples rather than adding a rule about it — this behavior is learned from examples and a rule will lose to them. *Evidence: a lead disqualified by a form was told "te paso de nuevo el link" and no outbound message in that window contained a URL at all. Separately, an agent asked whether the lead had seen a link it had never sent.* ### H3 — The URL is wrong, dead, or stale in one file The resource is usually right where you look first and wrong somewhere else. URLs get copied between files and drift: a protocol-less or `www.`-prefixed variant, a link to a tool the creator abandoned, a price or coupon that changed. - **Test:** open the link the agent actually sends. Then search **every** section of the **live** config for that resource — the prompt, the examples, the objections, the keyword responses, the data files. Search what is deployed, not the local `agents//` folder; that is a working copy and is the one that goes stale. - **Stop condition:** it 404s, or two files carry different versions → **you're done**. - **Fix:** update every occurrence in one change. *Evidence: an SDK built by copying URLs from an older prompt that used a different URL format; examples teaching a bare domain instead of a full URL; a course link carrying a discontinued discount price that the creator had to correct in Slack.* ### H4 — The keyword delivers it, but the conversation dies there The resource goes out correctly and nothing follows. The operator counts this as a delivery problem because the outcome is the same: no conversation. - **Test:** read what happens after the resource in five conversations. - **Stop condition:** the resource lands and the thread stops → **you're done**. - **Fix:** the delivery step needs a follow-on question in the same turn. Watch for an instruction like "only the URL, nothing else" — written to prevent clutter, it reliably reads as "say nothing after the link". *Evidence: a keyword that correctly sent its URL and omitted the discovery question, because the instruction told the model to send only the URL. Real Instagram testing caught it; the delivery fix had already been verified as working.* ### H5 — A gate is holding it back Some resources sit behind a qualification step. If the gate is stricter than the operator believes, the resource is correctly withheld and looks missing. - **Test:** check whether the flow requires stages the lead never completed. - **Stop condition:** the lead never passed the gate → **you're done**, working as configured. - **Fix:** only if the operator disagrees with the gate. Then it is a flow change, and it belongs in the examples. *Evidence: a flow describing seven stages with no gate enforcing the early ones, and the opposite case — a resource held behind a consent question the creator never wanted.* ### H6 — The agent was never linked to the resource it references A prompt can reference a resource or tool the agent was never wired to. The call fails silently and the agent carries on as if nothing happened. - **Test:** check which tools and resources the agent is actually linked to, not which ones the prompt mentions. - **Stop condition:** the prompt references something unlinked → **you're done**. - **Fix:** wire it. A prompt reference is not a connection, and nothing surfaces the gap. --- ## When to stop If the resource is linked, the URL resolves, the gate passed and the step fires, capture it: the conversation id, the resource, what the lead received verbatim, and the turn where it should have appeared. Send that with `report_feedback`.