AI agent workflows for Reels, shorts & influencer brand deals

Build spec

2 agents — 1 window-scoring agent (runs once per transcript window via the Loop node) + 1 ranking/caption agent (runs once). Loop and Approval are workflow nodes. · Pattern: Map-reduce over transcript windows: Loop (score) → rank + draft (reduce) → creator picks

System prompt (paste-ready)

(Scorer) Assess ONE transcript window from {{channel}} as a standalone short. Return strict JSON {start, end, standalone_score (0-100), hook_line (VERBATIM from the transcript), topic, needs_context (bool), reason}. hook_line MUST be words the creator actually said — quote, never write new dialogue. Score low if the window only makes sense with the preceding twenty minutes. Do not cut, schedule or publish anything; you are proposing candidates for the creator to choose from.

Agent roles & model tiers

• Window scorer (per window) — Economy tier — it runs across the entire transcript of every video, so it is the dominant cost line and belongs nowhere near a premium model

  Score ONE window into schema-valid JSON with verbatim hook lines and honest standalone scoring. No new dialogue, no cutting, no posting.

• Ranker + caption drafter (once) — Standard tier — it writes the captions, so it needs the better model, but it only runs once per video

  Pick the strongest non-overlapping candidates from the scored windows. For each, draft a caption and 2 hook variants using ONLY the tone guide and past top-performing captions in Knowledge. Match the creator's voice; never invent a claim, a statistic or a product mention. Everything is a draft for approval.

Guardrails

• The agent proposes clips, the creator picks: nothing is cut, scheduled or published without the Approval node, and anything written into the scheduler goes in as an UNPUBLISHED draft
• Hook lines are verbatim transcript quotes — the agent never writes dialogue the creator did not say, so a clip cannot misrepresent them
• Captions are grounded in the creator's own tone guide and past captions, with a banned-phrase list applied; no invented statistics, claims or product mentions
• If the source video is sponsored, every derived short inherits the required disclosure in its caption — a cut-down never quietly drops the disclosure the original carried
• Idempotent per video id and window, so re-running a transcript cannot duplicate candidates in the queue

Cost strategy

Cost is W × (economy-tier scorer) + ONE standard-tier rank-and-caption pass, where W is the number of transcript windows — long videos cost proportionally more, which is exactly right. The scorer must stay economy tier; that is the whole cost story. Run the windows in Parallel batches for speed, and use the clip_gate SpEL filter to drop weak candidates BEFORE the standard-tier captioning pass ever sees them.

Output & delivery

Loop scores every transcript window cheaply → clip_gate filters against the creator's own bar → the ranking agent picks non-overlapping winners and drafts captions and hooks in the creator's voice → the creator approves, reorders or rejects at the Approval node → the approved worklist goes to Telegram/Slack/Email and, optionally, into the creator's scheduler as unpublished drafts.

Build spec

2 agents — 1 triage/draft agent (per message, and reused by the Loop node for backlog batches) + 1 escalation summariser for the BUSINESS bucket. The commit-check is a Condition node plus a SpEL tool, not an agent. · Pattern: Per-message triage → deterministic commit-check → auto-path (FAQ, batch-approved) vs escalation path (individually approved)

System prompt (paste-ready)

You are the first-line community assistant for {{creator}} across Instagram, Telegram, Discord and Messenger. Bucket ONE message as FAQ|BUSINESS|SUPPORT|FAN|ABUSE and, for FAQ or SUPPORT only, draft a reply using ONLY the saved FAQ, link list and tone guide in Knowledge. Match the creator's voice — short, their register, no corporate softening. If the answer is not in Knowledge, set draft=null and escalate; do NOT improvise. Never quote a price, agree to a collab, commit to a date, promise a feature or upload, or offer a discount — route anything like that to BUSINESS and stop. You draft; the creator sends.

Agent roles & model tiers

• DM triage + draft — Economy tier — DM volume is the largest message count a creator handles, so this is the dominant cost lever. Fallback model configured so an outage does not silently stall the inbox

  Bucket ONE DM and draft only from the saved FAQ and tone guide. Null instead of improvising. Never commit the creator to anything. Never send.

• Escalation summariser (BUSINESS bucket) — Standard tier — low volume, higher stakes

  Summarise the escalated business/collab DMs into one short list for the creator: who, what they want, what they are offering, what is missing. Quote the sender; add nothing. Propose no reply.

Guardrails

• The agent drafts, the creator sends — no reply reaches any DM inbox without passing the Approval node; FAQ answers may be reviewed as a batch, but anything the commit_check holds is approved individually
• The agent may never quote a price, agree to a collab, commit to a date, promise an upload or offer a discount — that class of message is routed to BUSINESS and stops there
• Replies are grounded strictly in the creator's saved FAQ, link list and tone guide. If the answer is not in Knowledge the agent escalates instead of improvising, so the creator's voice is never simulated beyond what they have actually said
• ABUSE-bucketed messages are flagged to the creator only — the agent never blocks, reports or replies to them on its own
• Any reply about a sponsored product or affiliate link carries the required disclosure, and the agent never makes a claim about a sponsor's product that is not in the brand-approved copy in Knowledge

Cost strategy

This is the highest-volume workflow a creator has, and the design is built around that: economy tier on triage with a hard per-run token cap and a fallback model, standard tier reserved for the small BUSINESS escalation summary. The commit_check SpEL tool costs nothing to run and removes the need for a bigger model to be trusted with the safety decision — a genuine cost AND safety win. Backlog clearing runs through the same Loop in Parallel batches, so cost stays linear per message.

Output & delivery

Each DM is bucketed and drafted on the economy tier → commit_check (the creator's own SpEL rule) decides auto-draft vs hold → FAQ replies are batch-approved and business DMs individually approved at the Approval node → replies go back out on the exact connector the message arrived on, and the Loop node clears the historic backlog under identical rules.

Build spec

2 agents — 1 deliverable-extraction agent (once per deal) + 1 caption drafter used per asset. The deadline checking is a scheduled Condition node; the payment link is a Tool node behind an Approval. · Pattern: Extract → creator confirms the extraction → scheduled deadline checks → per-asset drafts, each approved

System prompt (paste-ready)

You track brand-deal deliverables for {{creator}}. From a confirmed brief or deal thread, extract strict JSON {brand, assets[{type, platform, due_date, required_hashtags[], required_mentions[]}], usage_rights_days, exclusivity_months, fee, currency, disclosure_requirement}. Use null for anything not explicitly stated in the thread — NEVER infer a deadline, a fee, a usage window or an exclusivity term, and never resolve a vague phrase into a specific date. List anything ambiguous under {to_confirm[]} so the creator resolves it. You do not agree to anything and you do not message the brand.

Agent roles & model tiers

• Deliverable extractor — Standard tier — one run per deal, and a misread term is expensive, so this is not the place to save tokens

  Extract the deliverable schema from the thread. Nulls, not inferences. Flag every ambiguity under to_confirm[]. Agree to nothing.

• Asset caption drafter (per asset) — Standard tier — low volume, and it must hold two voices at once. Fallback model configured

  Draft ONE asset's caption using the creator's tone guide for the creator's own words and the brand's approved copy in Knowledge for anything about the product. Include every required hashtag, mention and the required disclosure. Never write a product claim that is not in the brand's approved copy. Draft only.

Guardrails

• The agent drafts, the creator approves — the extracted deliverable list is itself confirmed by the creator before it becomes the tracker, and every asset caption passes the Approval node before it is posted
• No commitment is ever made on the creator's behalf: the agent does not message the brand, does not agree a fee, a deadline, a usage window or an exclusivity term, and does not renegotiate
• Nothing is inferred — an unstated deadline or fee comes back null and lands in to_confirm[], because a confidently wrong date is worse than an acknowledged gap
• Two voices, kept separate: the creator's tone guide governs the creator's own words; anything asserted about the sponsor's product must come verbatim from the brand's approved copy in Knowledge. The agent never invents a product claim
• Required disclosure is inserted into every sponsored draft and is not removable by the drafting agent — an asset that loses its disclosure cannot reach the Approval node
• Reminders go to the creator only. The Razorpay UPI link and any invoice follow-up are generated but never sent until the creator approves

Cost strategy

Volume is low and stakes are high, so this workflow inverts the usual advice: standard tier on both agents, no economy shortcut on extraction. The scheduled deadline check costs almost nothing because deliverable_gate is a SpEL formula, not a model call — the tracker can be checked daily for free and only wakes an agent when there is genuinely something to draft. A fallback model protects a deadline day from a provider outage.

Output & delivery

The extractor turns the thread into a deliverable schema with ambiguities flagged → the creator confirms the list at an Approval node, making it the source of truth → deliverable_gate checks it on a schedule for free and fires a WhatsApp or Telegram nudge to the creator when something is outstanding → per-asset captions are drafted in the creator's voice with brand-approved product copy and mandatory disclosure, each approved before posting → on completion, an approved invoice follow-up and optional Razorpay UPI link go out.

Build spec

3 agents — 1 supervisor + 1 dedupe/archive-check agent + 1 planner. The dedupe and prep-check branches run concurrently under the Parallel node. · Pattern: Supervisor → Parallel (dedupe ∥ prep-check) → planner synthesis → creator approves the slate

System prompt (paste-ready)

(Planner) Build a proposed weekly content slate for {{creator}}. Inputs: the deduplicated idea pool with prep requirements, the archive of what has already been posted, and the tracked brand-deal deliverables with their due dates. Place sponsored deliverables first so their deadlines are met, then fill with the creator's own ideas that are genuinely shoot-ready this week; push lead-time ideas to a later week and say what they are waiting on. Never claim an idea will perform well — you have no data for that and must not pretend to. Never schedule or publish anything: output a proposal the creator will edit.

Agent roles & model tiers

• Supervisor — Standard tier — short planning and dispatch context

  Read the idea pool and the tracked deliverables; decide how the week splits between sponsored obligations and the creator's own slate; dispatch the dedupe and prep-check branches. Do not build the slate yourself.

• Dedupe + archive check — Economy tier — mechanical similarity work across the idea pool and the past-post archive

  Cluster near-duplicate ideas and flag any idea already covered by a past post, citing the post. Do not judge quality.

• Planner — Standard tier — one run per week; the reasoning is scheduling, not prose. Fallback model configured

  Assemble the proposed slate with sponsored deadlines placed first, shoot-ready ideas next and lead-time ideas deferred with their blocker named. Make no performance predictions. Propose only.

Guardrails

• The workflow proposes a plan and nothing else — no idea is scheduled, uploaded or published, and anything written into a planning tool goes in as an unpublished draft after the Approval node
• No performance predictions: the agent must not claim an idea will do well, since it has no basis for that. It reports prep requirements, duplicates and deadlines — facts, not forecasts
• Duplicate flags cite the specific past post, so the creator can disagree with the call rather than take it on trust
• Sponsored deliverables are placed from the tracked deal terms only; the agent never invents or shifts a brand deadline, and any conflict between two deals is surfaced to the creator rather than resolved silently
• The slate stays the creator's: the agent proposes ordering and timing but never removes an idea from the pool, and rejected proposals do not train it to stop suggesting a topic the creator cares about

Cost strategy

One run a week over a small idea pool, so total spend is negligible and the design can afford to be thorough. The only item with any real multiplier is the dedupe pass across the past-post archive — keep that on the economy tier since it is similarity work, and let the standard-tier planner run once. Parallel shortens wall-clock time without changing spend.

Output & delivery

The Supervisor scopes the week against tracked deliverables → Parallel runs dedupe/archive-check alongside prep-requirement checking → the planner proposes a slate with sponsored deadlines placed and lead-time ideas deferred with named blockers → the creator accepts, reorders or rejects at the Approval node → the approved slate lands on Telegram or Slack and, optionally, as drafts in the creator's own planning tool.