AI Marketing Operator Logs · AI Marketing Engineering · Context Engineering

Engines are the named AI marketing systems Hendry Soong built and ran, and this section is the raw record of each one. Seven appear. Create-Articles and Create-Images are stable, at v8.3.0 and v4.8.0. Create-Compiler is retired. Listen-Competitors and Create-Social are dormant. Two replication engines carry the same architecture to other brands, one in production and one validated. Entry counts run from 1 to 68. Each engine holds its own context window, validation system and version history, which is why the entries record what changed, what broke and what was learned rather than a finished method. The three content engines are not independent products either. They hand work to each other in a fixed order, and the record is written from inside that pipeline.

These logs capture the real work of operating the AI Marketing Framework. Not polished thought leadership. Timestamped entries showing iterations, failures, and extracted principles from running production systems. The Foundation entries (v0.1 to v0.3) show context engineering in action: designing the brand voice, ICP, and positioning that shape every AI output.

Scope: These seven engines hold 123 of the 721 entries. A further 591 are in the track logs, filed by engineering concern rather than by engine: Data, GTM, Signal, Orchestrate, Governance and System Architecture. The remaining 7 are the deep dives.

The Pipeline: Articles flow through three engines: Create-Articles generates structured JSON with visual insertion points (0 SVGs). Create-Images generates SVG diagrams and hero images through a 9-check exit gate. Create-Compiler validates fields with 22 checks plus a 7-question review agent, classifies issues through a 4-tier router, and sends reverse manifests back to upstream engines. Output publishes to a headless CMS (Neon + Payload + Vercel) via publishing SDK. Closed-loop feedback. Each engine has its own context window, validation system, and version history.

Deep dives are the long-form counterpart to the log entries. Each one is a full retrospective on a learning that turned out to matter. A log entry records what changed and what broke. A deep dive returns to that change with the complete methodology behind it. It includes the code, and it closes with the principles pulled out afterward. The difference is depth rather than subject. Everything here has already appeared in a log somewhere, compressed into a line. The retrospective exists because that line was carrying more than it could hold. 7 published so far. Read a log when you want to know what happened. Read a deep dive when you want to know why it worked and which part of it transfers.

System Architecture is the operator log for the substrate: cross-engine architecture, contracts, and the website build. It holds 81 records, dated February 2026 to July 2026. The mix leans toward design work, with 29 architecture records and 20 fixes, then 14 learnings, 9 decisions, 6 features, 2 releases and 1 failure. Each record names what changed, what broke when something broke, and the principle taken from it. Every one traces to a dated source in the repository it came from, so the claims here can be checked instead of trusted. This is one section of the AI Marketing Operator Logs, which cover production AI marketing engines across hendry.ai, growthsetting and enterprise work.

Read the full System architecture log (81 records).

Governance is the operator log for authorisation: how each change was approved, what it granted, and what it refused. It holds 56 records, all dated August 2026. Exactly half are decisions, 28 of them, followed by 8 fixes, 6 releases, 5 learnings, 4 architecture records, 3 failures and 2 features. That shape is the point. A governance log is mostly rulings, not repairs. Each record names what changed, what broke when something broke, and the principle taken from it, and each traces to a dated source in the repository it came from. This is one section of the AI Marketing Operator Logs, covering production engines across hendry.ai, growthsetting and enterprise work.

Read the full Governance log (56 records).

Orchestrate is the operator log for the control plane, the layer that runs the other engines and records what it did. It holds 83 records, dated July 2026 to August 2026. The mix is close to even: 20 learnings, 18 decisions, 17 fixes, 14 architecture records, 6 failures, 5 releases, 3 features. A layer that touches everything leaves a wide trail. Each record names what changed, what broke when something broke, and the principle taken from it, and every one traces to a dated source in the repository it came from. This is one section of the AI Marketing Operator Logs, which cover production AI marketing engines across hendry.ai, growthsetting and enterprise work.

Read the full Orchestrate log (83 records).

Data is the operator log for making inputs usable and proving where each one came from. It holds 21 records, dated July 2026 to August 2026. Failures and architecture tie at 6 each, ahead of 5 learnings and 4 decisions. Each record names what changed, what broke when something broke, and the principle taken from it. Every one traces to a dated source in the repository it came from, so any claim here can be checked rather than taken on trust. Records are ordered newest first. This is one section of the AI Marketing Operator Logs, which cover production AI marketing engines across hendry.ai, growthsetting and enterprise work.

This engine runs across 3 pages: Data verification (70 records), Data measurement (58 records) and Data pipelines and sources (71 records).

Signal is the operator log for holding and scoring what matters, with extraction kept structurally separate from judgement. It holds 48 records, dated July 2026 to August 2026. Learnings dominate at 20, then 10 decisions, 9 failures, 4 architecture records, 4 fixes and 1 feature. A section that is mostly learnings is a section still being figured out. Each record names what changed, what broke when something broke, and the principle taken from it, and every one traces to a dated source in the repository it came from. This is one section of the AI Marketing Operator Logs, which cover production engines across hendry.ai, growthsetting and enterprise work.

Read the full Signal log (48 records).

GTM is the operator log for sequencing the engines that go to market and the contracts between them. It holds 17 records, dated July 2026 to August 2026. Architecture leads with 5, then 4 failures, 3 learnings, 3 features, 1 decision and 1 fix. Close to a quarter of the entries are failures, which is what sequencing work looks like while the contracts between engines are still being settled. Each record names what changed, what broke when something broke, and the principle taken from it. Every one traces to a dated source in the repository it came from. This is one section of the AI Marketing Operator Logs, covering production engines across hendry.ai, growthsetting and enterprise work.

This engine runs across 2 pages: GTM measurement (55 records) and GTM verification and sources (69 records).

Create-Articles is the operator log for the content generation engine and its three tiers of validation. It holds 68 records, dated December 2025 to August 2026. By type: 14 architecture, 13 features, 13 releases, 9 fixes, 9 failures, 6 decisions, 4 learnings. Features and releases together outnumber every other category, which is what a shipping engine's log looks like. Each record names what changed, what broke when something broke, and the principle taken from it. Every one traces to a dated source in the repository it came from, so any claim here can be checked rather than taken on trust. This is one section of the AI Marketing Operator Logs, covering production engines across hendry.ai, growthsetting and enterprise work.

Read the full Create-Articles log (68 records).

Create-Images is the operator log for SVG diagram and hero generation, its perception rules and its exit gates. It holds 26 records, running through August 2026. Architecture accounts for 10 of them, then 5 fixes, 4 releases, 3 decisions, 3 features and 1 learning. Almost no learnings and a heavy architecture share point to an engine solved by design rather than by iteration. Each record names what changed, what broke when something broke, and the principle taken from it. Every one traces to a dated source in the repository it came from, so any claim here can be checked rather than taken on trust. This is one section of the AI Marketing Operator Logs, covering production engines across hendry.ai, growthsetting and enterprise work.

Read the full Create-Images log (26 records).

Create-Compiler sits at the end of the content pipeline as its assembler. It merges finished articles with finished images into one publishable unit. That position makes it the last thing to run and the first to notice when something upstream went wrong. Before anything ships, it runs a 14-check quality gate. Failures do not simply stop the build. The compiler classifies what failed and routes that classification back to the upstream engines. The defect then gets fixed at its source instead of patched at assembly. That return path is what separates it from a packaging step. The count now stands at 7+ versions. The logs below trace how the checks accumulated and how the feedback loop closed.

Create-Social is the engine that generates LinkedIn carousels. Its log is the quietest in this collection, and the quiet is the finding. The version number stopped at v1.0.2 in January 2026 and has not moved since. That is unusual in a system where most components keep getting rebuilt. Stability here is not neglect. A carousel has a narrow job and a fixed output shape, so once the generation logic was right there was little left to tune. Adding versions would have meant adding changes nobody needed. The entries below cover how it reached that state rather than what has happened after, because the honest answer is nothing. A log that stops is a result, not a gap.

Listen-Competitors began as monitoring and ended as competitive intelligence. The early versions did what most competitor tracking does today. They watched what rivals published and reported the changes back. That produces a feed. A feed tells you something happened without telling you whether it matters. The later versions moved to synthesis, which is a different job entirely. The engine now reads across what it collects and returns a position rather than a list. Collecting became concluding. 7 versions cover the distance between those endpoints, and the arc matters more here than any single release. The entries below show where the shift happened and what each version added on the way there.

Replicate exists to answer one question: do these engines work for anyone other than the person who built them? A system tuned to its own author is a workflow. A system that survives transplant into someone else's brand is an architecture. The difference only resolves by attempting it. The engines were rebuilt inside other brands and run there under real conditions. 48+ versions across 3 brand implementations came out of that work. The answer held. The architecture is portable rather than Hendry-specific, which is the finding this whole track was built to test. The logs below record each implementation in order. They include what broke on arrival and what had to be rebuilt before the engine ran clean.

The Replication Journey: Started with the Hendry Content System (25 versions), then tested on three external brands. Each test compressed: Brand A took 2 weeks and 15+ versions. Brand B took 3 days and 5 versions. Brand C took 2 days and 5 versions. The pattern became formulaic.

81 principles have been extracted so far, drawn from 721 curated iterations across every engine. None were written in advance. A principle earns its place by recurring, usually after something broke in a way that trying harder would not fix. Each is tagged with the engine version that produced it, so a rule can be read next to the failure that generated it. The set keeps growing, which is the expected behaviour of a system that logs its own mistakes rather than its own wins. Read this way it is less a philosophy than a maintenance record: the same problems kept arriving, and these are the sentences that stopped them arriving again.

Top 10:

1. The agent is disposable. The orchestration layer is permanent.

2. The LLM generates numbers. The human sees shapes.

3. Evidence-based validation defeats hallucination.

4. Agents improvise unless explicitly forbidden.

5. Context engineering is not prompt engineering. It's designing the information layer.

6. One example beats 89 lines of instructions.

7. Context window is a finite resource. Separate what from how.

8. The problems come first, then the principles, then the product.

9. Cross-session memory turns isolated agents into a learning system.

10. Memory is infrastructure. Without it, the system cannot learn from its own history.