docs: write the pages in neutral third person

The docs pages carried 36 attributed opinions — "Matt's own answer",
"his position is", quoted replies from the author. A page is a
technical document about a skill, so the substance of each finding
stays and the attribution goes: "the fix is a direct instruction: …",
"the split comes down to session count".

Quotes from *users* stay, anonymous as they already were — those are
evidence about the skill in the wild rather than the author's view.

Records the rule in .agents/writing-docs.md so new pages don't
reintroduce it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

docs/productivity/teach.md

@@ -21,7 +21,7 @@ Reach for it when the learning is the project: a language, a framework, a codeba
 
 ## Prerequisites
 
-`teach` builds a directory rather than producing a file, and the skill assumes one mission per workspace — so run it somewhere you are happy to give over to a single topic. Matt's own guidance, when someone proposed routing the output to a global `~/.learnings/` folder, was to keep it out of the project you are working in: "you can just put them in a separate repo." A dedicated repo also makes the lessons committable, which is how teams have shared them.
+`teach` builds a directory rather than producing a file, and the skill assumes one mission per workspace — so run it somewhere you are happy to give over to a single topic. Keep it out of the project you are working in: a separate repo is the recommended home, rather than a global `~/.learnings/` folder or the working project itself. A dedicated repo also makes the lessons committable, which is how teams have shared them.
 
 What accumulates in that directory:
 
@@ -59,10 +59,10 @@ Lessons are built from **components** in `assets/`: stylesheets, quiz widgets, s
 A real, open bug ([#377](https://github.com/mattpocock/skills/issues/377)). `SKILL.md` uses `./` for two different roots at once: `./MISSION-FORMAT.md` and its siblings really do sit next to `SKILL.md` in the installed skill, while `./lessons/`, `./reference/`, `./learning-records/` and `./assets/` are meant to be in your directory. An agent that resolves the first kind against the skill's install directory goes on to resolve the second kind there too, and writes your course into the skill folder. Check where the first lesson landed before you build on it, and name the directory explicitly when you start rather than relying on "the current directory" being understood.
 
 **Do I stay in one session, or start a new one per lesson?**
-Asked which of three approaches he uses — stay in the same session, re-invoke `/teach` in a new session, or open a new session in the same folder — Matt's answer was "All three will work." Asked whether each lesson is its own invocation, he confirmed: "Yes, each session." The folder is the continuity, not the conversation. Common practice is to open a fresh session in the workspace and say `/teach next lesson for <topic>`.
+All three approaches work — staying in the same session, re-invoking `/teach` in a new session, or opening a new session in the same folder. Each lesson is its own invocation. The folder is the continuity, not the conversation. Common practice is to open a fresh session in the workspace and say `/teach next lesson for <topic>`.
 
 **How do I know it isn't teaching me something it made up?**
-You don't, on the skill's word alone. Asked exactly this, Matt's answer was "You read the primary sources" — and asked whether that made `teach` reliable, "Of course not, see my video 'never trust an LLM'." The grounding machinery — `RESOURCES.md`, citations in every lesson, one recommended primary source per lesson — exists to make verification cheap, not to remove the need for it. The failure is not hypothetical: one user learning a 2x2 Rubik's cube was given fabricated move sequences that don't solve it. Matt's diagnostic questions in that thread were "Model, harness, effort? And, what was the source?", which is the right checklist. Risk is highest in procedural domains with precise notation, and lowest where the output is immediately verifiable, like code you can run.
+You don't, on the skill's word alone. You read the primary sources. `teach` is not reliable enough to trust unchecked, and no skill built on an LLM is. The grounding machinery — `RESOURCES.md`, citations in every lesson, one recommended primary source per lesson — exists to make verification cheap, not to remove the need for it. The failure is not hypothetical: one user learning a 2x2 Rubik's cube was given fabricated move sequences that don't solve it. The diagnostic checklist for a case like that is model, harness, effort — and what the source was. Risk is highest in procedural domains with precise notation, and lowest where the output is immediately verifiable, like code you can run.
 
 **The correct quiz answer is always the first option.**
 Confirmed by several people, on Sonnet, on Opus and on GLM, and still unfixed. `SKILL.md` now requires every answer to be the same number of words, which kills a different tell — the correct answer used to be the only fully-reasoned one — but says nothing about position. One contributor tested an instruction-level fix for position and reported the correct answer still landing in slot A 33 times out of 33 across nine lessons ([#335](https://github.com/mattpocock/skills/issues/335)), which points at a shuffling quiz component in `assets/` as the real fix rather than better wording. Until that ships, treat answer position as meaningless. Your `assets/` directory is yours to change, so asking for a component that shuffles at render time is a legitimate local fix.