← All case studies
◢ Chapter 05B · Case Study

The agent flow behind the briefing.

A 7-stage pipeline from raw signal sources to LinkedIn publish — with the LinkedIn Post Generator engineered as Fan-Out + Reflection to solve the hook problem. This is what a PM-led agent design looks like end to end.

7
Stages
3
Patterns used
2
Human touchpoints

The 7-stage pipeline

Click any stage to expand its input/output, design notes, failure modes, and the checkpoints generated there. The pain point is auto-opened.

◢ Checkpoint legend — what gets generated where
Fan-Out

Where N parallel generators run with distinct strategies — diversity, not redundancy.

Stages: 5
Reflection

Where a critic agent scores output against a rubric and triggers regeneration if it falls short.

Stages: 5 · 7
Contract Spec

Where a typed input/output schema (and rubric) gates what passes between stages.

Stages: 5 · 6 · 8
01

Newsletters, RSS feeds, Twitter/X lists, Slack channels — wherever signal-worthy AI PM news surfaces.

02

Reads raw inputs, identifies items that qualify as AI PM signals, and scores relevance.

03

Removes signals published in prior editions, formats survivors in the 📌 numbered style.

04

Generates 3 variants with different hook angles, critic scores each against a rubric, surfaces the best to human.

Fan-out: 3 parallel generators
Hook A: Surprising claimHook B: Trend readHook C: Contrarian takeCritic scores all 3
Input
Final signal list + voice guide + prior posts + brand rules
Output
3 scored variants, ready for human selection
Design note
Fan-Out gives creative diversity. Reflection enforces quality before human sees anything. This is the stage that costs the most time today.
Failure modes
  • ·All 3 variants using the same hook angle
  • ·Generic hook ("AI is changing things")
  • ·Critic missing real tone problems while nitpicking style
05

Generates 3 variants with different hook angles, critic scores each against a rubric, surfaces the best to human.

◢ Checkpoints generated at this stage
Contract spec generated here

This is the stage that gets the full Post Generator Contract — input/output schema, weighted critic rubric, failure modes. See the Contract tab.

Fan-Out happens here

3 parallel generators run with predefined hook strategies (surprising claim · trend read · contrarian take). One inference call per variant.

Reflection happens here

Critic agent scores each variant against the rubric (Hook 30% · Voice 25% · Signal 20% · Format 15% · CTA 10%). Below 7.5 weighted → regenerate.

Fan-out: 3 parallel generators
Hook A: Surprising claimHook B: Trend readHook C: Contrarian takeCritic scores all 3
Input
Final signal list + voice guide + prior posts + brand rules
Output
3 scored variants, ready for human selection
Design note
Fan-Out gives creative diversity. Reflection enforces quality before human sees anything. This is the stage that costs the most time today.
Failure modes
  • ·All 3 variants using the same hook angle
  • ·Generic hook ("AI is changing things")
  • ·Critic missing real tone problems while nitpicking style
06

Rahul picks the preferred variant — or approves the auto-selected top-scorer if score ≥ 8.0.

07

Applies any selection edits, finalises spacing, emoji, and CTA. One lightweight reflect pass.

08

One-click final approval. Posts to LinkedIn via API or clipboard. Logs to edition archive.