Insights
Field notes from running LLM-augmented marketing in production. Each post is a standalone read — anchored to a specific job-to-be-done, named sources, no vendor pitches.
Sequence follows the Playbook's reading order — start at the top if you're new to LLM prompting.
-
Foundations · #1
The four marketing jobs LLMs change first
The work that shifts before strategy or measurement does — briefs, drafts, distribution copy, and lifecycle variants.
-
Operational · #2
The model-choice matrix you do not have
A new flagship model lands every six weeks; vendors push you to "the smartest model". Choose by job shape instead, and re-evaluate on a cadence rather than at every release.
-
Operational · #3
Why your prompt library does not survive a second team
The single most common reason internal prompt libraries fail when handed to a new team — undeclared variables, missing worked examples, no failure-mode notes.
-
Operational · #4
From chat histories to a single source of truth
Your team's best prompts live in seventeen places and no one's library survives the new starter. The fix is what every other artefact has already been through — store it, version it, change it on purpose.
-
Operational · #5
Prompts as code — why your library needs a commit history
A prompt without a version history is folklore that happens to be in writing. The fix is not a Notion page; it is a repository, a diff, and a changelog per template.
-
Operational · #6
Why the same prompt and the same model gives different output
The mechanism behind run-to-run drift — sampling, system-message stability, hidden state in the chat session, and silent model updates. The fix is to treat the prompt as a tuple, not a body.
-
Governance · #7
How to cite a marketing statistic without becoming a vendor pitch
Three rules for sourcing claims your CMO can defend in a vendor review — primary sources, named methodology, and the date check.
-
Governance · #8
Auditing hallucinated citations in your back catalogue
The model invents a Forrester benchmark and you ship the post before catching it. The audit checklist for a back catalogue you suspect contains fabricated cites.
-
Governance · #9
Output you cannot defend — the audit trail every AI-generated draft needs
When the CMO asks where the bolded statistic came from, the answer cannot be "the model wrote it". The fix is the audit trail your team builds before a draft ships, not the apology you draft after.
-
Governance · #10
The measurement spine your finance team will accept
The team is shipping more and your finance team wants the figure. "We feel faster" is not a number. Three metrics, named calculations, and the rate trick that makes the percentages defensible.
No posts under this theme yet. Pick another, or subscribe to /updates for new ones as they ship.