Skip to main content

The feedback loop: research in, a sharper engine out

The idea: treat your brand engine as something that learns. Feed it real research and real performance data, and use both to tune the engine itself (its voice, positioning, rules, and audience model), so every future generation is more informed and more differentiated than the last. Over weeks, the content stops sounding like "good AI content" and starts sounding like the sharpest person on your team wrote it.

This is the workflow that most directly fights slop, because slop is what you get when a model has nothing specific to say. The fix is a better-informed engine, not a cleverer prompt.

The loop​

        research + context ─┐
                            ▼
                     ┌─────────────┐
   analytics ───────▶│  the engine │──────▶ generate ──▶ publish
   (what worked)     └─────────────┘                       │
        ▲                   ▲                               │
        └─── measure ◀──────┴──────── tune the engine ◀─────┘

Every pass tightens two things: what the engine knows (research, context docs, competitive landscape) and how it sounds (voice, tone, anti-slop rules), guided by what actually performed.

Step 1: feed it research and context​

Generic content comes from a generic knowledge base. Give the engine specifics.

1. add_competitor        // Map the landscape: who else plays here, their angle, where you differ
2. upload_document       // Product specs, positioning memos, transcripts, POV pieces
3. list_documents        // Confirm what's grounding generation

The competitive landscape matters for anti-slop specifically: an engine that knows what everyone else says can be steered to say something else. Differentiation is the opposite of slop.

Step 2: generate, then measure honestly​

1. create_brief → generate_content → get_generation_status → get_content
2. publish_now / schedule_content
3. generate_utm_link       // So web outcomes attribute back to the piece

Then let it run long enough to have signal, and read the results without flinching.

4. get_analytics_overview       // The baseline
5. get_top_content              // What actually landed
6. get_topic_clusters           // Which themes win, which die
7. get_search_analytics         // Real queries you rank (or nearly rank) for
8. get_insights                 // The engine's own read: top performers, decay alerts, shifts

Step 3: turn findings into engine changes (the part people skip)​

This is the difference between "using AI to post" and "building an engine your team owns." Don't just brief the next post: change the engine based on what you learned.

  • A phrasing or angle keeps outperforming? Fold it into the voice. 
    update_voice            // Add the winning cadence to personality/tone; extend the lexicon
    update_writing_rules    // Promote what worked into rules; ban what read as filler into Anti-Slop
  • The market has moved, or a competitor claimed your old angle? 
    update_positioning      // Sharpen the wedge and category frame
    add_competitor          // Keep the landscape current so you keep differentiating
  • One audience segment drives the results? 
    add_audience            // Model them explicitly so generation targets their pains and language
    set_platform_tone       // Tune the channels where they actually are

Let analytics write your next brief for you, then generate against the now-sharper engine:

suggest_brief_from_analytics    // Data-driven brief from what's working
create_brief → generate_content

Step 4: compare and keep the loop turning​

Generate the new version against the tuned engine and check it moved.

1. get_content                  // Read the new draft: it should be more specific, less hedged
2. get_content_performance      // After it runs, compare to the piece it replaced
3. get_insights                 // Did the change register?

Repeat on a cadence (monthly is plenty). Each turn the engine holds more of what's true about your market and your voice, and the anti-slop rules get sharper from real evidence rather than guesswork.

:::tip Version your engine changes Engine edits are versioned. Make a bold tuning change, generate a batch, and if it doesn't help you can roll the section back, so the feedback loop is safe to experiment in. :::

The payoff​

Anyone can prompt a model. A feedback loop is what compounds: research makes the content informed, analytics makes the tuning evidence-based, and the anti-slop rules, sharpened every pass, make the output distinct. Six months in, your engine is a genuine asset that a competitor with a raw LLM can't match, because they don't have your loop.

Next: run the refined engine hands-off on a schedule, or embed it in your own product as an invisible engine.