Cheapest technique first, LLM last. Use before designing any feature that classifies, matches, routes, dedups, ranks, scores, or extracts, anything where the first instinct is "call the LLM per item". Forces enumeration of deterministic and classical-stats…
Commit with a Conversation Decision Record. Use for every commit in repos that keep CDRs, or whenever a commit may land with no PR, no review, and no human reading the diff. Writes the decision record, stages it with the code, and adds machine-parseable…
Stop forwarding raw AI output to humans. Use when the session's result needs to go to another person, "summarise this for the team", "give me something to send my PM", "distill and link". Writes a short summary the sender can own and ships the full working…
Generate discrete image assets with Google's Gemini image models (Nano Banana Pro, Nano Banana 2), and refuse to generate design mockups. Use when asked for product shots, stock or marketing photography, mascots, illustrations, textures, logos, OG and social…
Prose for a human reader is the deliverable. Use when writing or revising anything a person will sit and read, including docs, READMEs, PR descriptions, site copy, posts, emails, and UI text. Not for code, commit messages, or chat replies.
Stop burying the answer. Use on every response, including casual ones and ones where nobody asked for brevity. Lead with the outcome, number the steps, park tangents visibly, skip the ceremony.
The brag nudge. Use when something real has just shipped, a PR train merged, a deploy gone live, a project delivered. Writes a brag brief and nags the human to go post about it.
When the implementation fights back, stop patching and re-derive. Use when a case will not fit the design, a fix wants a flag or special case or shim, the same area keeps breaking, or the user says "nah".