Spec the 5-Station Repurposing Pipeline
Objective: Given a weekly source asset and a target set of output formats, write a build spec for the lesson's 5-station pipeline, naming the tool for each station and where the forced human review checkpoint sits.
You run content for Allbirds' sustainability team, which publishes a monthly founder Q&A video that currently gets manually clipped into social posts by an intern, taking most of a day. You've been asked to spec an automated pipeline before the team commits budget to it.
Map the founder Q&A video onto the lesson's five stations (trigger, transcribe, extract, draft, schedule), naming a tool per station, and place the required human review checkpoint before scheduling, not after.
Before you start
What you'll need
Free path (everything below is enough to finish)
Free, simple enough for a spec that will be handed to a developer
Free tier is enough for spec writing
Paid upgrades (optional, faster/deeper)
Native LangChain support and 70+ AI nodes cover transcription, extraction, and drafting in one connected workflow
No access? Zapier as the trigger-and-glue layer if n8n's visual builder is too much for a first build
The process
2 steps
Step 01 of 02
The lesson's 5 stations are trigger, transcribe, extract, draft, schedule, each a separate tool connected by an automation platform rather than one tool doing everything.
The founder Q&A video is uploaded monthly to a shared Google Drive folder. Which named tool goes in each of the 5 stations, and what does each one actually receive and hand off?
Procedure
- Station 1 (trigger): name n8n's file-watcher watching the Drive folder for a new upload
- Station 2 (transcribe): name the Whisper API turning the video's audio into a timestamped transcript
- Station 3 (extract): name an LLM step pulling quotes, a summary, and takeaways as structured output, kept separate from drafting
- Station 4 (draft): name three parallel LLM calls, one per target format (LinkedIn post, email blurb, video script)
- Station 5 (schedule): name Buffer receiving only human-approved drafts
PIPELINE SPEC, Allbirds Founder Q&A Repurposing Station 1, Trigger: n8n file-watcher on the Drive folder, fires on new upload Station 2, Transcribe: Whisper API, timestamped transcript Station 3, Extract: LLM call, pulls 6 quotes + 3-sentence summary + 3 takeaways as structured JSON Station 4, Draft: 3 parallel LLM calls, one LinkedIn post, one email blurb, one 45-second script Station 5, Schedule: Buffer, receives only drafts marked approved in the review channel
Healthy
Each station has one named tool and a clear input/output, and extraction is kept separate from drafting.
Unhealthy
One LLM prompt asked to both extract and draft, or a station left as 'a tool that does this.'
What this means
Keeping extraction narrow (structured facts only) makes the drafting stage easier to fact-check against the source.
So what do I do about it?
| Symptom | Action | Effort |
|---|---|---|
| Extraction and drafting are combined into one prompt | Split into two separate LLM calls with a JSON hand-off in between | 5 min |
Step 02 of 02
The lesson requires the review checkpoint be built into the automation itself, routing drafts to a shared Slack channel or pending-review folder before they ever reach the scheduler's queue, not as a manual step someone has to remember.
Where exactly in the Allbirds pipeline does the human review gate sit, and what specifically does it block from happening automatically?
Procedure
- Insert a review stage between station 4 (draft) and station 5 (schedule), not after station 5
- Name the destination: a #content-review Slack channel, tagging the content lead
- Define the gate condition: only an approval reaction moves a draft into Buffer's queue
- Confirm the gate blocks the automation from calling Buffer's API directly from station 4
REVIEW CHECKPOINT SPEC Location: between Station 4 (Draft) and Station 5 (Schedule) Destination: #content-review Slack channel, tags @content-lead Gate condition: n8n only calls Buffer's API after detecting an approval emoji reaction on the draft message Blocked without approval: no draft can reach Buffer's scheduling queue automatically
Healthy
The gate is a required automation step (an emoji-triggered API call), not a note in a doc telling a human to remember to check.
Unhealthy
The pipeline posts directly to Buffer from station 4 and 'review' is just a suggestion in the team's process doc.
What this means
A review step that isn't wired into the automation gets skipped under deadline pressure; a review step that blocks the next API call can't be.
So what do I do about it?
| Symptom | Action | Effort |
|---|---|---|
| Drafts sometimes reach Buffer without anyone reviewing them | Move the Buffer API call so it only fires on a detected approval signal, not automatically after drafting | 30 min |
Final deliverable
A one-page build spec naming the tool for all 5 stations plus a written description of exactly where and how the review gate blocks unapproved drafts.
See a reference example
PIPELINE SPEC, YETI Product Launch Video Repurposing Station 1, Trigger: Zapier watches the marketing Drive folder Station 2, Transcribe: Whisper API Station 3, Extract: LLM call, 5 quotes + summary + 3 takeaways Station 4, Draft: 2 parallel LLM calls, LinkedIn post + email blurb Review checkpoint: #content-review Slack, approval emoji required Station 5, Schedule: Buffer, only after approval
Success criteria
You're done when you can:
- Names a specific tool for all 5 stations with a clear input/output at each
- Extraction and drafting are two separate steps, not one combined prompt
- Review checkpoint is described as a blocking automation step, not a process reminder