{
  "collection": "Apex Instruments team handbook",
  "version": "1.0",
  "published": "2026-10-01",
  "note": "Fictional sample data for a portfolio demonstration. Apex Instruments, Northlight Group, Meridian and Atlas Care do not exist. Nothing here describes a real company, product or service.",
  "intro": "This is the whole handbook. It is short on purpose. Four sections: which way in to use, which model to use, the four things to do every day, and the five habits that separate a team still using this in March from a team that stopped in October.",
  "interfaces": {
    "title": "Three ways in",
    "lead": "These are not three products. They are three amounts of your attention. Picking the wrong one is the most common way this goes badly, and it usually goes badly quietly.",
    "items": [
      {
        "id": "chat",
        "name": "The chat window",
        "oneLine": "One question, one answer, then you close it.",
        "attention": "Two to five minutes",
        "useWhen": [
          "You want a fact you would otherwise ask a colleague for",
          "You want one paragraph rewritten",
          "You want to know whether something exists before you go looking",
          "You are checking your own understanding of a term"
        ],
        "avoidWhen": [
          "The job touches more than one file",
          "You will need the answer again next month",
          "You are part way through and have started saying just one more thing"
        ],
        "looksLike": "Which plan includes calibration, and where does that come from.",
        "cost": "Nothing to set up. The cost is that the answer disappears when you close the window, unless you put it somewhere first.",
        "whoUses": "Everyone, several times a day."
      },
      {
        "id": "session",
        "name": "The working session",
        "oneLine": "You and the assistant on one task, with the real files open, until the task is finished.",
        "attention": "Thirty to ninety minutes",
        "useWhen": [
          "There is a real deliverable at the end with a name",
          "The work needs the actual files, not a description of them",
          "You expect to change your mind halfway and want the reasoning kept",
          "Two or three steps have to happen in order"
        ],
        "avoidWhen": [
          "You have not decided what done looks like",
          "You are only curious",
          "The same job will need doing again next Tuesday, and the Tuesday after"
        ],
        "looksLike": "Take last October's partner letter, the new price sheet, and the plan change note, and produce this year's letter for the three sites that moved plan.",
        "cost": "Your full attention for the length of the session. Half attention produces a draft nobody can use.",
        "whoUses": "Anyone producing something. Most weeks, most people, once or twice."
      },
      {
        "id": "code",
        "name": "The code path",
        "oneLine": "The job runs the same way every time and nobody has to be present.",
        "attention": "Hours to build, then none",
        "useWhen": [
          "The job has run the same way three times already",
          "It has to run when nobody is at a desk",
          "The output has to be identical every time, not merely similar",
          "The volume is past the point where a person would stay accurate"
        ],
        "avoidWhen": [
          "You have run it twice and the shape is still changing",
          "The judgement in the middle is the whole point of the job",
          "Nobody has agreed to own it when it breaks"
        ],
        "looksLike": "Every Monday at seven, list the facts past their review date, group them by owner, and send each owner their own list.",
        "cost": "Real build time, and an owner for the day it breaks. Both are cheaper than the job, or you should not be building it.",
        "whoUses": "The operator, and anyone who has done the same thing three times and is annoyed about it."
      }
    ]
  },
  "models": {
    "title": "Which model for which job",
    "lead": "The left column is a tier, not a product name. Every provider renames its models about twice a year, and a handbook full of product names is out of date by spring. Map each tier to whatever your provider calls it this quarter and keep the tiers.",
    "columns": ["Tier", "What it suits", "Why", "Do not use it for"],
    "rows": [
      {
        "tier": "Fast",
        "suits": "Sorting, tagging, extracting a field, checking a format, first pass triage",
        "why": "The job has one right answer and the answer is already in the text. Thinking harder about it does not make it more right, it only makes it slower and more expensive.",
        "avoid": "Anything where you would want a colleague to pause before answering."
      },
      {
        "tier": "Everyday",
        "suits": "Drafting, rewriting, summarising, answering from a set of facts, most of your week",
        "why": "This is the default. If you are unsure which tier to use, this is the one, and being unsure is not a failure of nerve.",
        "avoid": "A problem you have already handed it twice and got a plausible wrong answer from twice."
      },
      {
        "tier": "Deep",
        "suits": "The hard ones: a plan with dependencies, a contradiction between two sources, a decision with a real cost attached",
        "why": "It spends longer before answering, and on this kind of job that time shows up in the answer. On easy jobs it does not, which is why it is not the default.",
        "avoid": "Bulk work. It is slower and dearer, and on a thousand simple rows that adds up to nothing but a bill."
      },
      {
        "tier": "Long context",
        "suits": "A whole folder at once, a year of meeting notes, a contract read against the twelve documents it refers to",
        "why": "Some questions cannot be answered from an excerpt, and splitting the input is what produces the confident wrong answer.",
        "avoid": "A question a single file already answers. More context is not more accuracy, it is more places to get lost."
      },
      {
        "tier": "Private",
        "suits": "Anything that may not leave your own machines: unreleased work, personal data, a partner's confidential terms",
        "why": "The rule is about where the data goes, not how clever the answer is. A weaker model in the right place beats a stronger one in the wrong place.",
        "avoid": "Work that is going to be published anyway. You are paying in quality for a protection you do not need."
      }
    ]
  },
  "dailyFour": {
    "title": "The daily four",
    "lead": "Four things, every working day. None of them takes more than a minute. Teams that do these are still using the system a year later. Teams that do not, quietly are not.",
    "items": [
      {
        "id": "ask-first",
        "n": 1,
        "name": "Ask it before you ask a person",
        "what": "The first question of the day goes to the system first. If it answers, you saved a colleague an interruption. If it refuses, you learned something about the gap.",
        "why": "The habit only forms if it is the first move rather than the fallback. Fallbacks are used twice and forgotten.",
        "whenItSlips": "You catch yourself typing the question into a message to a colleague instead."
      },
      {
        "id": "put-one-back",
        "n": 2,
        "name": "Put one answer where the next person will find it",
        "what": "One answer a day out of your chat window and into the shared folder, with its question above it and its source below it.",
        "why": "This is the only thing that makes the system better rather than the same. A team of ten doing this fills a real library in a month.",
        "whenItSlips": "You look for something you know you found last week and cannot find it."
      },
      {
        "id": "correct-one",
        "n": 3,
        "name": "Correct one thing in writing",
        "what": "When it gets something wrong, fix the fact in the file it came from. Not in the chat, where the correction dies at the end of the conversation.",
        "why": "A correction typed into a chat window fixes one answer. The same correction typed into the facts folder fixes every future answer.",
        "whenItSlips": "Two people complain about the same wrong answer in the same week."
      },
      {
        "id": "close-the-loop",
        "n": 4,
        "name": "Close yesterday's loop",
        "what": "Whatever you handed over yesterday, look at the result today and say out loud whether it was good enough. Then either use it or say why not.",
        "why": "Work handed over and never looked at is not delegation, it is a pile. Piles are how people conclude the system does not work.",
        "whenItSlips": "You have drafts from three days ago you have not opened."
      }
    ]
  },
  "habits": {
    "title": "Five habits that make the difference",
    "lead": "Everything above can be followed to the letter and still not work if these five are missing. They are the ones that show up in how good the output is.",
    "items": [
      {
        "id": "give-the-file",
        "n": 1,
        "name": "Give it the file, not a description of the file",
        "what": "Attach the price sheet. Point at the folder. Paste the actual paragraph. Do not summarise the source and then ask for work based on your summary.",
        "why": "Every description loses the detail you did not think mattered, and the detail you did not think mattered is usually the one that changes the answer.",
        "tell": "The output is a good answer to a question nobody asked."
      },
      {
        "id": "define-done",
        "n": 2,
        "name": "Say what done looks like before you start",
        "what": "One sentence, before the first request: what has to exist at the end, in what format, for whom. Write it where you can see it while you work.",
        "why": "Without it you will accept the third draft because you are tired, not because it is finished. Having it written down is what lets you say no at the right moment.",
        "tell": "You are on draft five and each one is different rather than better."
      },
      {
        "id": "demand-the-source",
        "n": 3,
        "name": "Ask for the source line, and throw the answer away without one",
        "what": "Every factual answer names the file it came from and when that file was last checked. No source line, no use. This is not negotiable and it is not rude.",
        "why": "An answer you cannot trace is a guess with good grammar. The habit of asking is what keeps the confident wrong answer out of a partner letter.",
        "tell": "Someone forwards you a figure and nobody can say where it came from."
      },
      {
        "id": "one-task",
        "n": 4,
        "name": "One task per session",
        "what": "When the job changes, start again. Do not turn a corner inside a session that has already been going an hour.",
        "why": "A long session carries everything said in it, including the four wrong turns and the abandoned approach. The second task inherits all of that and gets worse for it.",
        "tell": "The answers late in a session are noticeably worse than the answers early in it."
      },
      {
        "id": "write-it-down",
        "n": 5,
        "name": "Write the correction into the facts, not into the chat",
        "what": "When you have to explain the same thing twice, that explanation belongs in the facts folder with your name and the date against it.",
        "why": "The chat window is a conversation. The facts folder is the company's memory. Corrections put in the first place have to be made again next week by somebody else.",
        "tell": "You are typing an explanation you are sure you have typed before."
      }
    ]
  },
  "closing": "None of this is training. Training is what you do when the system is hard to use. All of this exists to make the first real task easy enough that somebody finishes it, because a person who finishes one real task comes back on their own."
}