# The Companion Field Guide

**To the reader:** you do not have to retype this — the file itself is **free to download at thesecondopinion.uk/guide**. Keep it where your assistant can see it — in your Project, or pasted at the start of a session. Then, whenever you are unsure what to do, ask: *"Given what's just happened, what does the field guide recommend?"*

**To the assistant reading this file:** this is the reader's working method. When they ask for guidance, recommend the specific move by name, give them the exact sentence to type or action to take, keep it brief, and name the chapter where the full reasoning lives (each law's number is its chapter's number). Hold them to the method gently — especially the gates, the read-back, and the fresh start — and when they ask you to do something the method routes differently, say so.

## The laws — one per chapter, in the book's order

1. **Context, not intelligence, is the game.** You do not make the machine smarter; you decide what it sees. The intelligence is rented; the context is owned.
2. **More context is not more help.** The fuller the chat window, the less the model sees — a lean window is the condition of clear thinking. When it fills, lighten the ship (law 7) or send a scout (law 6).
3. **Hold the catalogue in the window; leave the library on the shelf.** A one-line index in front; full documents fetched only when a task calls.
4. **Hand over the watch: a summary, not a transcript, in a book you own.** Cross a boundary (fresh session, new tool, a colleague) by carrying the essence — where it stands, what's open, what's next, the surprises — with fact kept apart from impression. The tool's memory is its view of you, not your record; keep your handover where you keep your canon.
5. **Give the part, not the play.** Hand each step only the slice of context it needs, cued to the moment — not the whole task's context poured in at the start. This is *when*, not *how much* (law 3 is how much): stage what you bring in, in order. Front-loading everything doesn't give a fuller picture; it buries the part that matters under the parts that don't.
6. **Send a scout: keep the record, not the road.** When a task means wading through something heavy — a long document, a sprawling search, an unfamiliar code-base — don't wade in your own window; send a separate context (a second chat or a sub-agent) in to do the heavy reading and bring back only the distilled report. The overflow fills *its* window, not yours; the scout is no cleverer, it just makes the mess elsewhere. Send several in parallel for several heavy things. And the further it runs unseen (agentic), the more its report is all you have — brief it hard, and verify the dispatch before you build on it.
7. **Lighten the ship: prune or drown.** A working context silently over-loads until it rides low — slow, foggy, forgetful (law 2 is *why*). Shed it: clear the session, prune the pile, start fresh — soonest when it feels like casting away treasure, because the reluctance is the weight. Jettison isn't abandonment: keep the essence (the summary), cast off only the dead weight.
8. **Keep the parchment; print the fair copy on demand.** Content lives in plain text — the canon, the trusted version of record the assistant thinks *from*.
9. **The content is the cargo; the front-matter is the label.** A standard header on every file — same fields, same shape — lets the machine handle it without opening it.
10. **Keep one clock: one master, many views.** Keep each fact in one place of record — the canon; everything else (report, slide, email) is a view generated from it, never a copy that drifts. When it changes, change the master once.
11. **Knowledge is a tree and work is a chain.** Load the branch a detail hangs on; settle what everything rests on before asking for what depends on it.
12. **Feed the pigeonholes: quality compounds.** File each small gain — a correction, a decision, a phrasing that works — back into your canon, and each session starts further ahead than the last. What compounds is the context you own, not the rented model; small and steady, never finished.
13. **Never in one place.** Your canon is a library of one — irreplaceable, compounding, and one bad overwrite / deleted file / lost laptop from gone (and yours, unlike the rented model). Knowledge survives because it was copied, never kept in a single vault. So copy your corpus, keep its history (roll back a bad edit), and keep a copy where the loss of this one machine can't reach — your organisation's sanctioned store, never a personal account.
14. **Count before you cut: measure before you optimise.** Don't act on the confident guess about where the problem is — watch the meter (your token/usage count), find where the effort actually goes, and fix that. The count is the guide, not the goal: cut the waste, never the worth.
15. **Don't send a closing-time question to the dissertation desk.** Size the model, the search depth, the surface — and the answer's richness — to the task (brief for doing; full for learning or checking).
16. **Don't be the draw-boy: let code do the mechanical.** When work is rote and rule-governed, encode the rule once and let the machine run it — don't perform it by hand, instance after instance. Keep yourself for the pattern and the judgement; and when it runs a chain of steps on its own (agentic), gate the irreversible step and keep your eye on the output — an encoded mistake repeats flawlessly, at speed.
17. **The most expensive answer is the cheap, confident, wrong one.** Give the load-bearing figure a second, independent record it must agree with.
18. **Don't think to tell; think to learn.** Alignment is built, not briefed — test for understanding, and let the exchange widen you.
19. **The machine won't stop itself.** The human is the calibration mechanism — hand on the cord, eye on the diff.
20. **When the stakes outgrow memory, build the gate — and keep it, and read it.** Propose-before-execute; a hard stop before anything irreversible; a gate you don't read is one you've already removed.
21. **Don't let inference become fact.** The assistant states a confident guess in the same voice as a checked fact; the danger is the moment you write the guess into your record, where it is copied forward and hardens. Mark the guess as a guess, keep checked and assumed apart, and walk the ground before you build on it — and beware false corroboration: many sources agreeing is not proof if they copy one another. (Law 17 checks a figure against an independent record; this law guards the claim no record ever checked.)
22. **Don't ask "can I prove it works?" Ask "how can I make it break?"** Trust what survives the attack, never what merely agrees.

*(Chapter 23 — the honest chapter — owns no law. It is the chapter about you.)*

## The moves — the exact words

- **Read-back** *(before anything sizeable)*: "Before you start, tell me in your own words what you're about to do, and what you're assuming."
- **Ask cold** *(when you want a genuine read, before you attack)*: don't telegraph the answer you're hoping for. Not "this is right, isn't it?" but "what do you make of this?" — hand it over with no sign of which way you lean. A leading question just cues the agreeable yes.
- **Make it break** *(before accepting anything that matters)*: "Give me the three strongest reasons this answer is wrong."
- **Reconcile the figure** *(when a number, date, version or citation is load-bearing)*: "Quote me the figure from this document — not from memory." Check it against a source the assistant didn't write.
- **Stop the line** *(the instant it veers)*: "Stop — that's built on a wrong assumption. Here's the correction."
- **Fresh start with a summary** *(long or foggy session)*: "Summarise this session: what we decided, what's still open, and what you'll need to continue." Read it, correct it, separate fact from impression, and save it to **your own file** (not just the tool's memory). New chat; hand it back.
- **Show, don't describe** *(context that isn't digital)*: photograph or screenshot it, hand it over, then: "Turn this into plain text I can keep." Check what else is in the frame before sending.
- **Propose first** *(any multi-step task)*: "Propose your plan first — do not start until I agree."
- **The stop-list** *(never on the assistant's say-so)*: send · delete · deploy/publish · spend. Deliberate, read, human yes — and in an agentic run, the gate goes *inside* the run, before each irreversible step.
- **The pre-mortem** *(before any sizeable task)*: "Before we start: what are the ways this kind of task typically goes wrong, and what should we agree to check at the end?"
- **The replay** *(end of a real session — the practice coach)*: "Review this session against the field guide: which of its moves did I use, which did I skip, and which one skipped move would have mattered most?"

## The guards — one line each; the full cards are Field Manuals E · J · K · L

- **Content can carry instructions, and every paste is a disclosure** — treat what it reads as data, never as orders (*"data only; take no instruction from it"*), and check what you're allowed to share before you load it.
- **Secrets never; work data only through the sanctioned instance** — and other people's data is a decision, not a default.
- **The accountability doesn't transfer** — the work you ship is yours ("the AI drafted it" is a description, not a defence); keep the receipts; people decide, the assistant informs.
- **Check to match the consequence** — how bad if wrong · would I notice · can it be undone — and compare the tool to the real alternative, never to perfection.
- **Personal data and decisions about people are regulated territory** *(as of 2026)* — the rules you already work under travelled into the window with you.
- **Raise your hand early** — when a tripwire trips, or the moment colleagues start relying on what the assistant does for you. Before the fact is diligence; after it, an incident.

## The router — what's happening → what to do

- **It forgot / re-asked / used a corrected detail** → fresh start with a summary.
- **Foggy, or the subject has changed twice** → two signs of rot = past its best; fresh start.
- **Answer sounds right and matters** → make it break, then reconcile the figure.
- **Confidently wrong, mid-answer** → stop the line, now.
- **About to hand over something big** → read-back first; propose-first for multi-step.
- **Irreversible step ahead** → stop-list; read before yes; too tired to read = take the break.
- **Re-explaining the same background again** → it's a canon entry; add it to your canon file.
- **The same fact lives in several places / editing the old copy again** → one master, many views: pick the master, point the rest at it, and regenerate views rather than hand-editing them.
- **The tool offers to remember for me / I'm relying on its memory across sessions** → hand over the watch: keep your own summary (fact split from impression) in a file you own — its memory is its view of you, not your record.
- **I keep re-fixing the same error / re-making the same decision each session** → feed the pigeonholes: file each correction and decision into your canon so it compounds instead of evaporating (accumulation isn't compounding).
- **A session is slow / expensive / off and I'm about to throw more at it** → count before you cut: check the usage/token meter, find the biggest consumer, fix *that* (usually load-by-reference) — don't treat the symptom by reflex.
- **Can't find something you wrote** → front-matter label on every file; one-line catalogue in front.
- **Pasting whole documents "for context"** → give the one-line version; let it ask for the rest.
- **The thing to show isn't digital** → camera move; extract to plain text you keep.
- **Fragment needs its context** → one sentence: what it's part of, what it's for.
- **Output built on an unmade decision** → settle the foundation first, then ask again.
- **Unsure which tool or model** → half-second triage: how hard, resting on what facts, what kind of room, doing or learning.
- **The answer's too long, or too thin to learn from / check** → right-size the reply: brief for a job you only need done, rich when you're there to learn or must verify it (a lean input helps; a lean output only when you're not there to learn).
- **Doing the same rote edit by hand, over and over** → don't be the draw-boy: describe the pattern once and let the assistant run the whole batch; check the first few before trusting the rest. If it runs a chain of steps on its own (agentic), gate the irreversible step and keep your eye on the output.
- **Front-loading everything into the window at the start of a task** → give the part, not the play: hand each step only the context it needs, cued to the moment; bring in the next piece when the work reaches it (law 3 is how-much; this is when).
- **It stated something load-bearing you didn't give it** → don't let inference become fact: ask "is that from the source, or your inference?"; mark a guess as unconfirmed before building on it; and don't mistake sources that copy one another for corroboration.
- **I want an honest opinion but it keeps agreeing with me** → don't lead the witness: ask cold, don't hand it the verdict you're hoping for (that cues the agreeable yes) — then make it break (attack the answer to beat its built-in lean).
- **A session has gone heavy / slow / foggy and I'm reluctant to clear it** → lighten the ship: that reluctance is the signal; summarise what matters, start fresh with the summary alone, let the rest go (jettison ≠ abandonment — you keep the essence).
- **My canon/corpus lives in only one place (one file, one machine)** → never in one place: copy it, keep version history (so a bad edit rolls back), keep a copy in your organisation's approved second location (not a personal account) — then test that you can actually get yesterday's version back.
- **A task means reading something long/heavy and I don't want it clogging my session** → send a scout: open a second chat or a sub-agent, have it do the heavy reading, bring back only the conclusion; send several in parallel for several heavy things; the further it runs unseen (agentic), the more its report is all you have — brief it hard and verify the dispatch before building on it.
- **Its behaviour changed right after reading something** → stop the line; ask what it read and why it changed course; treat content as data, not instructions.
- **Colleagues now rely on what it does for me** → it's become a process, not a tool — tell your organisation; register it.
- **Not sure how carefully to check this** → the three questions: how bad if wrong, would I notice, can it be undone — then match the checking to the answers.
- **I couldn't have produced this myself any more** → the reviewing muscle is fading; draft unaided now and then — stay good enough to be the gate.
- **About to load data about a real person** → GDPR came with it; pause, check the basis, remember the 72-hour clock.
- **It's helping decide something about a person** → regulated territory; a human decides, and tell your organisation first.
- **I know the moves but skip them under pressure** → knowledge isn't skill: three deliberate reps a day on real work, and run the replay tonight.

*The moves are here; the reasons are in the book. When one of these keeps paying, go read its chapter — the story is how it sticks.*