# The Recipe Card Deck ## The Deck, and How to Read It Every card runs four parts, terse on purpose, the Chapter 13 register: trigger, steps, guardrails, verify, then the judgment handed back to you. The cards sort the work. The judgment stays yours. Part 1 reprints the five cards Chapter 13 taught for work that lives behind clicks. Part 2 adds five the book banked but never carded. Part 3 is the guardrail card. Click paths rot fast, so current versions of every card live on the companion site (`abeburnett.com/vibe-coding`). ## Part 1: The Full R-13 Set The five cards from Chapter 13, unchanged. ### R-13.1: Multi-System Data Entry - **Trigger:** the same record goes into two or more systems, by hand, more than once a week. - **Steps:** map the fields once, on paper or in a sheet: source value to each system's field name → build the source file, the master list of records → run one record at a time, field by field, system by system → finish each record before starting the next → run the verify pass before touching the next record. - **Guardrails:** - credentials never go in the prompt, and the agent never types them - you log in, or your tool holds the session - the source file stays out of the chat entirely, descriptions only - the full Chapter 3 checklist applies to every screen the agent sees - nothing submits until the verify pass passes - **Verify:** - row counts match across every system, every time - spot-check three fields per record against the source - the session log shows one line per entry, with timestamp and system - if a count is off by one, stop the run, not the record Your morning, your call: the day the systems start talking, delete this card and go back to a script, which will do the job faster. ### R-13.2: Portal Watch Loop - **Trigger:** a portal whose deadline or status changes matter to you, and you currently check it by habit. - **Steps:** schedule a daily check at a fixed time (the scheduling mechanics live in Chapter 9) → capture the status fields into a file → diff today against yesterday (compare the two files) → alert only on change, with the old value and the new value in the alert. - **Guardrails:** - read-only, always - the watcher never clicks submit, never accepts anything, never replies - alerts go to you, not to a group, until you decide otherwise - if the portal shows customer data, the watch runs only in a setup that passes the Chapter 3 checklist for that data - **Verify:** for the first week, check the portal yourself once a day and confirm every alert matches a real change, and no real change went unalerted. Whether a portal deserves a watcher is your judgment. The week it stops mattering, turn the loop off the same week; a watcher that nags about nothing teaches you to ignore it. ### R-13.3: No-Export-Button Extraction - **Trigger:** a system shows the data you need on screen, and offers no way to export it. - **Steps:** write down exactly which rows and columns you need, before you start → page through the screens, capturing what's there → write it to a structured file, one row per record → count the rows → compare the count against the total the screen itself displays. - **Guardrails:** - Chapter 3 applies to screens: PII (information that identifies a person) visible on screen is PII - if the screen shows customer names, this card runs only after the Chapter 3 checklist passes for that data and every service receiving the screen - capture what you need, not everything the screen offers - the file you build is data too, so it follows the same rules as any other file - **Verify:** - the row count matches the on-screen total - spot-check five rows against the screen - the file opens clean in the tool you'll actually use next If your screen shows names you don't need, that's a data-rules decision no card can make for you. Leave them on the screen. ### R-13.4: Scheduled Dashboard Pulls - **Trigger:** the same numbers come from the same dashboard on the same day every week. - **Steps:** open the dashboard on schedule → read the tiles in a fixed order → write the numbers to the tracker with the date → note anomalies in the same row, in your words → leave the dashboard. - **Guardrails:** - read-only - the pull never sends, posts, or files anything - dashboard credentials live in your tool's own login and never appear in a prompt - the tracker stays on your machine or a team share you can already write to - **Verify:** - for the first two weeks, pull the numbers yourself and compare - after that, the file should read like a history, with no silent gaps - any gap is an alert, not a shrug If Monday's dashboard stops being the truth, retire the card before it faithfully reports the wrong numbers for a month. ### R-13.5: CRM Entries, Timesheets, Receipts - **Trigger:** call notes that should become CRM records, time that should land in a timesheet, receipts that should land in expense categories. - **Steps:** one note, one record: read the note, write the record, mark the note done → timesheets: pull the week's calendar and sent email, draft the entries, review them against the day's shape → receipts: name, date, amount, category, in that order, and flag anything that brushes a policy edge. - **Guardrails:** - nothing sends without your review, ever - expense-policy questions go to a human, not a guess - don't categorize what you can't read, blurry receipt included - time entries you didn't actually spend are the one thing this card will not do - **Verify:** - record count matches note count - a colleague could follow any entry back to its source - the flagged items are a list, not a mystery When a receipt fights the policy, the policy wins, and a human decides. That human is you. ## Part 2: Expansion Cards Same shape, new nouns. ### R-G1: The Weekly Digest Run - **Trigger:** the incoming pile lands the same way every week, and you read it by hand. - **Steps:** schedule a fixed morning each week → the agent answers four questions per item: what it is, what it asks, by when, decision or FYI → rank by what needs your judgment soonest, not by sender → anything that asks nothing and owes nothing goes to the bottom → the digest lands as a draft. - **Guardrails:** - draft-status only: nothing auto-ships, ever - nothing confidential in the prompt; describe, don't paste - no stated ask gets "unknown," never a guess - **Verify:** - the ranked list matches the pile, item for item - decisions match your own read of the week - re-read the first paragraph before you act The digest sorts. You decide. ### R-G2: The Meeting Minutes Machine - **Trigger:** a recurring meeting produces decisions and commitments that live only in your memory. - **Steps:** capture the record before the meeting ends, transcript or running notes → summarize: decisions, one line each, with who made the call; action items with owner, task, deadline if stated, and where tracked; open questions; and any "we should probably" that never got an owner → track: one line per item → follow up Monday: last week's items against the tracker, flag past-due items, draft one-line status requests you review and send. - **Guardrails:** - the record stays in safe tools, per Chapter 3 - nothing sends without your review, ever; the Monday nudges are drafts - "we should probably" becomes a decision needed, never a silent action item - **Verify:** - every decision and commitment appears in the summary - every action item carries owner, task, date, home - the Monday list matches the tracker Read the Monday list like information about your team. ### R-G3: The Onboarding Kit - **Trigger:** a new hire starts, and the team's knowledge lives in files nobody reads and heads nobody asks. - **Steps:** point the agent at the team's AGENTS.md files and the docs they reference → draft a first-week kit: who owns what, where files live, which systems open first, what done looks like → the kit lands as a draft for the manager's edit, then ships. - **Guardrails:** - no personal data: nothing beyond role and file ownership, nothing from HR files - references by name and path, not pasted content, per Chapter 3 - the kit is a draft until a human reads it - **Verify:** - a new hire could complete the first task from the kit alone - a colleague could correct any line by opening its source - the manager's edit pass takes under an hour You're handing the new hire a map, not the territory. ### R-G4: The Compliance Binder Refresh - **Trigger:** the first week of the quarter, and the Chapter 16 binder hasn't been touched since last quarter. - **Steps:** run the refresh on a fixed date each quarter → pull the quarter's log entries: tool used, date, tier → check each against prompts, drafts, sources, sign-off → flag entries missing a sign-off or a named tier → close with your sign-off: name, date, what you reviewed → file the refresh note on top. - **Guardrails:** - the binder holds method, not data: regulated tasks record the tool and tier, never the data - prompts that failed the Chapter 3 checklist get redone, and the refresh flags them - nothing gets deleted in a refresh; stale entries get an as-of date - the sign-off line is yours, every quarter - **Verify:** - an auditor could reconstruct any logged task unaided - every past-quarter entry carries a current sign-off - you could produce the binder without wincing Auditors trust folders. Keep yours true. ### R-G5: The Pre-Meeting Dossier - **Trigger:** a meeting is on the calendar, and you'd walk in knowing less than you should about the people at the table. - **Steps:** pull the attendee list from the invite → for each name: last touch, recent shipped work, open threads between you → one page per meeting, assembled the night before → write the one question worth asking each person; that column is yours. - **Guardrails:** - descriptions only, per Chapter 3: nothing confidential goes in, nothing leaves safe tools - dossiers are for your eyes only - an empty thread stays empty; no invented history - **Verify:** - spot-check two dossiers against your memory - every claim traces to a source you could open - the meeting starts with zero blank-stare moments The dossier ends blank-stare meetings; the question column stays yours. ## Part 3: The Guardrail Card The one card that rules them all. No trigger of its own; every trigger is its trigger. Print it. ### R-0: The Guardrail Card - **Trigger:** any card in this deck is about to run. - **Steps:** name what the run touches: money, permissions, sends, deletes, credentials → confirm the log habit: timestamp, system, action, result → run → read the log first. - **Guardrails:** - the five never-gates hold each time, no exceptions, not once per session, not once per week: - money never moves without you at the gate - permissions never get granted without you at the gate - sends never go out without you at the gate - deletes never happen without you at the gate - credentials never enter a prompt, and the agent never types them - a gate you set once is a door; "approve everything" isn't a gate - Chapter 3 applies to screens: PII on a screen is PII - when colleagues depend on the output, the log becomes an audit obligation - **Verify:** - a stranger could read the log and tell you what happened, in order - nothing moved, sent, or deleted without your gate - every entry traces to timestamp, system, action, result The typing is the boring part, and the judgment is the part you keep. This card is where the two shake hands. ## The Judgment Stays Yours Cards rot slower than screenshots, but they rot. Fields get renamed, screens redesigned, systems replaced, and a card that faithfully describes last year's portal will faithfully waste this year's Tuesday. Retire or revise it when that happens; the site holds current versions. The judgment stays yours. It always was.