Prompts That Actually Work
One prompt structure, how to hand it your spreadsheets, how to get emails and SOPs out of it, and the habit of checking its work.One prompt structure, how to hand it your spreadsheets (CSV in context), getting emails and SOPs out, and checking its work.A single operator prompt template, tabular data in context, document generation with style anchors, and verification patterns.
Carry on withCarry on withnext:
- 017 min
The One Prompt Structure You NeedThe One Prompt Structure You Need: Who and What, the Material, the Shape, the CheckThe operator prompt template: system prompt vs user turn, structured output, one file per jobDoneDonedone
Tell it who you are and what you want, paste the material, say exactly what shape the answer takes, and tell it to say MISSING instead of guessing. Four blocks. They fix most bad answers.Four blocks: who and what (the role and the task), the material (what it works from), the shape (rules and output format, a schema when a program reads it) and the check (use only what I gave you). Structure beats clever wording because it makes wrong answers look wrong.A four-block operator template. Stable blocks (role, constraints, output shape, grounding rule) go in the system prompt; the material is the user turn; the shape becomes a JSON schema when downstream code reads the output. One versioned file per job.
- 028 min
Paste the Spreadsheet: Working With Your Own DataPaste the Spreadsheet: Your Own Data in the Context WindowTabular data in context: CSV size, chunking, redaction, and asking for the query instead of the answerDoneDonedone
Yes, you can paste a 500-row inventory export into a chat and ask questions. Here is what to paste, what to cut first, and why you ask for the formula instead of the total.A 500-row inventory export (a CSV) fits in the model's working memory (the context window) with room to spare. Here is what to paste, what to strip and redact first, and why anything numeric gets a formula or script back instead of a number.CSV in context vs. a tool: token budgets by shape, header discipline, redaction before paste, chunking by group, and the rule that the model writes the query and the runtime computes the answer.
- 038 min
Vendor Emails, SOPs, and Budtender Training in MinutesVendor Emails, SOPs, and Budtender Training: One Example of Yours, One Rules Block, One Saved Prompt per DocumentDocument generation: style anchors, few-shot examples, house-rules constraints, and templates promoted to skillsDoneDonedone
The boring writing that eats your week: the restock email, the collections note, the SOP nobody wrote down, the one-pager a new budtender needs by Friday. Paste one thing you wrote, paste your rules, paste the material, and read it before it goes out.Three document jobs, one method: a style anchor (one email you wrote and liked), a house-rules block (no medical claims, your state's ad rules), and the material. Saved once per document type as a system prompt, reviewed against a checklist before send.Style anchor and few-shot examples in the system prompt, a reusable house-rules constraint block, one template per document type promoted to SKILL.md, and a compliance review step that is part regex, part decision-model question, part human.
- 046 min
Check Its Work: The Five-Second HabitCheck Its Work: The Five-Second Habit (and the Checks a Script Can Run for You)Verification: the five-check habit, evals, spot-check sampling, run logging and deterministic validatorsDoneDonedone
Before you send it, ship it or upload it: five quick looks. Did I give it the source? Is one number right? Are the dates right? Where did that come from? Does it say the same thing twice? Here is how each one takes five seconds.Five checks before anything leaves your hands: source, number, date, provenance, rerun. Each is five seconds by hand, and each has a version a small script can run every time so the habit does not depend on who is tired.Manual five-check pass mapped to permanent controls: run logging, deterministic validators (schema, regex, sum checks), spot-check sampling, and a small eval set rerun on every prompt or model change. Guardrails for agents are covered in 07-05 and only linked here.
- 058 min
Prompt Tricks That Actually Help (and the Ones That Don't)Prompt techniques that actually help: examples, steps, sources and second draftsPrompting techniques in 2026: few-shot, decomposition, grounding, self-critique, and what is folkloreDoneDonedone
The internet is full of prompt tricks. A few really help: showing an example, splitting a big job into steps, giving it the facts, and asking for a few versions. Others mostly do not anymore. Here is which is which, matched to the problem you are having.The named techniques (few-shot examples, step-by-step, prompt chaining, grounding, self-critique, role prompts) sorted by what they fix and whether they still matter with current models. Pick by symptom, not by name.Few-shot, chain-of-thought, least-to-most decomposition and prompt chaining, generated knowledge versus grounding, self-refine, role prompting and maieutic prompting, each with a 2026 verdict for current reasoning models, plus tokenisation basics that explain several odd failures.