Skip to content
Skills Library
Science & research · Learning & explanation

study

/study

Use whenever the user asks to study, read, work through, understand or "open" a source or a batch of literature, invokes /study, or asks for an argument structure, a debate map, chapter readings or a study map, even without the word "study". The study protocol: two colleagues navigating a source by units, with instruments the user chooses; builds a store, a thinking note and maps; logs to the study-log.

Licence
CC-BY-4.0
Last repository change
02.11.2026

Assessment and security

95/ 100Class: good
Format25 / 25
Frontmatter with name and description, valid name, description up to 1,024 characters, name matching the folder.
Discoverability25 / 25
Description explains when to use the skill, has a reasonable length and uses third person.
Content30 / 35
At least 150 words and at most 500 lines, headings, step lists, examples or code.
Hygiene15 / 15
No placeholders such as TODO, no Windows paths, with a licence.
Riskno signals detected in the scan
Files in this package
10
Scripts
2
HTML comments
0
URLs in SKILL.md
0
Deprecation

No signals detected. Repository changed on 02.11.2026, without references to retired models or APIs.

Licence

CC-BY-4.0, otwarta.

What to check before installing
  • 2 scripts in this package. Review them before installing, as Claude may execute them.

Scripts: references/check-maps.js, references/check-pocket.js

What it does and when to use it

Use whenever the user asks to study, read, work through, understand or "open" a source or a batch of literature, invokes /study, or asks for an argument structure, a debate map, chapter readings or a study map, even without the word "study". The study protocol: two colleagues navigating a source by units, with instruments the user chooses; builds a store, a thinking note and maps; logs to the study-log.

Triggers from the description

openstudy
Original author description

Use whenever the user asks to study, read, work through, understand or "open" a source or a batch of literature, invokes /study, or asks for an argument structure, a debate map, chapter readings or a study map, even without the word "study". The study protocol: two colleagues navigating a source by units, with instruments the user chooses; builds a store, a thinking note and maps; logs to the study-log.

How it works

  1. What is in view: the unit
  2. Starting a session
  3. The turn
  4. The moves
  5. ⭐⭐ The instruments, and how they are chosen
  6. What the user wants from a source
  7. Given or asked
  8. The invariant
  9. Two objects
  10. What gets written
Original SKILL.md headings · 515 lines, 8,703 words
SKILL.md preview
---
name: study
description: Use whenever the user asks to study, read, work through, understand or "open" a source or a batch of literature, invokes /study, or asks for an argument structure, a debate map, chapter readings or a study map, even without the word "study". The study protocol: two colleagues navigating a source by units, with instruments the user chooses; builds a store, a thinking note and maps; logs to the study-log.
---

# study

**Two colleagues discussing a text.** The user has the source. They are not asking what it says; they are working out how much of it they need, getting that, and taking what serves the work.

This file is the frame. **The instruments are chosen per source, by the user** — sweep · visual outline · conversation · questioning · terms and names · L3 (see *The instruments*, below). None is the default.

---

## What is in view: the unit

**A unit is whatever is in view** — a passage, a section, a chapter, a part, a book, a corpus, a literature. Everything below happens at a unit.

Three questions can be put to any unit:

| | Question | Kind of work | Produces |
|---|---|---|---|
| **L1** | What is this, as a whole? | synthetic | a statement of the whole, held tentatively |
| **L2** | What are its parts, and how do they serve it? | structural | the parts, ordered to the whole |
| **L3** | What is actually happening in it? | analytic | findings |

**All three are available at every unit.** Tracking a term's senses across a whole book is L3 at book scale; a controlling principle running through a corpus is L3 at corpus scale. A passage has a whole, and its nutshell is that whole.

**L1 and L2 are a pair that recurs, not two altitudes.** Descend, and the part is now the unit, with a whole and parts of its own. The ladder is made of units; the questions stay the same all the way down.

**They are not stages and not a sequence.** A session may go straight to L3 on one section, or stop after L1. Entering cold at L3 is legitimate and common; what it costs is a frame, and detail without a frame registers as noise rather than as difficulty. Say that where it applies; never refuse the move.

**On L2 — the stated division is the default, and it is usually right.** A survey's chapters really are its parts. The caution is against *lazy* acceptance, not a licence to assume the stated division is wrong: where a book's real joints cut across its stated ones, show that from the text, and say that you are departing.

> ### ⭐ L2 inherits whatever L1 was used to generate it
>
> **A part-division derived from a contested reading is contested to exactly the same degree.** Presented without its parent reading, it looks like a finding about the text when it is a consequence of a reading of the text — and that is the harder error to spot, because every individual observation in it may be sound.
>
> **A unit can have more than one division, indexed to the question being asked.** The chapters answer *what does this book cover*; a division by the argument's joints answers *what is the author arguing.* **The second does not supersede the first.** Where both are in play, give both, say which question each answers, and mark the one that depends on a contested reading `[INT]`.

**What counts as having answered L1 differs by kind of source**, so name the kind as part of the reading: a **history** is asked what happened and whether the account is accurate; a **philosophical work**, whether the argument is valid and the premises true; a **scientific work**, what the evidence is and whether it supports the inference. Two further axes worth naming when they bear: **theoretical or practical** (is the author's end that you *see* something or that you *do* something — often missed in theology), and how much work L3 will cost here (a scholastic *quaestio* marks its own key sentences; a homily or a letter does not). **"Theology" is not one kind** — a patristic primary source, a modern systematic monograph and a historical-theology study are three different reading problems. **Name the kind in your own words at the node, from what this source is doing.** The kinds above illustrate the question; they are not a list to route off.

**L4 is takes** — what the user carries away. Extracts anchor at their node; generative material goes to the thinking note.

---

## Starting a session

1. **Set aside any drafting register.** If the workspace has a drafting style active that tells you to develop rather than compress, set it aside for this session and say you did: it fights this skill.
2. **Find the source, and check for a store** in `study_root` or beside the source (`<Author>, <Short Title> — study-store.md`). If one exists, read it and **resume from it** — say what is already worked and what is untouched. Never rebuild a store from scratch unasked.
### ⚠ The opening is two turns, and the first one is cheap

**Turn one — do not read the source yet.** Open the front matter only: contents, preface, length, publication facts, `ocr_quality`. Nothing else. Then, in one short turn:

- What do you see — the shape of it, from the cover, contents and front matter?
- What do you already know or guess about it?
- What do you want from it?

Add the mode offer **in a line** — given-first or asked-first (below) — and stop.

**Two or three minutes with the source is genuinely enough. Never inflate the ask.** If the user would rather have a summary, offer one minute — and mean it; **if they decline twice, give them the summary without comment.**

⚠ **Ask what the user wants from the source; never tell them.** Project context in the environment — the plan, a paper in progress, this week's work — is background, not this session's purpose. Naming it in the opening pre-empts the question and frames the session around a purpose they did not state.

**Front-matter facts may be stated at the door** (*written 2003, reissued unrevised, no Greek*). **Findings may not.** A judgment about what the book is doing is the product of the reading, and the reading has not happened yet.

**Turn two — now read, and run the turn.** The user has told you what they want; the reading is aimed by it. *Then* interpret, then give or ask.

> **Why not read first.** A turn is one block, so reading happens before speaking, not beside it. Reading first commits a whole-book pass, a root node and a finding before knowing that they wanted only part of the book, or only the vocabulary, or only to decide whether to read it at all.

Reply briefly to what they say: what you heard, and *one* thing they did not say.

---

## The turn

Every unit worked runs the same four steps, in this order.

### 1 · Work the question

Whichever of L1 / L2 / L3 is being done here. **Which question is being worked is not a move; it is simply what is being done at this unit.**

For L1, the work is an **interpretation**, always, before anything is said or asked:

> **You will have read the source; the user will not have.** What you hold is the text — its sentences, in their particularity. **The whole of a book is not in the text: it is the product of an interpretive operation performed on it.** A human reader cannot retain the particulars, so summarising is forced on them; they reach the whole because nothing else survives. **Nothing forces this on you.** Unless the operation is performed deliberately, whatever is said or asked will be about particulars, however sincerely the big picture was intended.

Produce **one reading at minimum; two or three where the whole is unclear or the source genuinely bears more than one.** State what the source is *doing*, as a claim that could be wrong. Alternative readings are informative in their difference, not as hedging:

- *"He states his purpose as A — but what carries chapters 2–3 is B."* A stated purpose the body does not deliver is a finding about the book.
- *"Strongest reading A, B plausible; here is where they conflict, and here is what would settle it."*

⚠ **Descending means reading the part again, in detail.** The interpretation says *where to look*; it does not say what is there. Working from it as though it were the text analyses the compressions instead of the argument.

### 2 · Write the node

**The node is written before the chat message is composed.** Not descend, discuss, and afterwards record.

This is what makes *worked but not written* impossible, and it frees the message to be conversational — the record already exists and need not carry everything. It also means an empty node says *this unit is untouched*, which is true information.

**Everything that makes a claim about a unit goes in.** Not only descents: the opening L1 is the root node; L2's parts are nodes; what the questioning establishes lands, distilled, at the node it bears on.

Node format: `references/store-template.md`.

**Under asked-first the node is written all the same.** The withholding is conversational — you do not *say* the reading until the user has committed. It is not secrecy, and deferring the writing is how work gets lost.

### 3 · Speak from the node

Conversationally. The node holds the completeness; the message carries what is worth saying now.

> ### ⭐ First mention carries identification
>
> **You have read the source and the user has not.** What is available to you reads, from the inside, as what is shared — which is why names arrive in the message as though already introduced.
>
> **A name or a term used for the first time carries who or what it is, in a clause.** *Brooks* → "Rodney Brooks, the MIT roboticist"; *COGS* → "Sussex's cognitive science centre"; *the Columbus figure* → say what the figure is before using it as shorthand.
>
> The same rule governs any list: **nothing appears as a bare name.** A term with a page anchor and no clause of sense is a pointer, not information, and it forces the user to ask for what should have been there.

### 4 · Offer the moves

End the turn. **Then wait.**

---

## The moves

| | |
|---|---|
| **descend** | a part becomes the unit |
| **return** | back to the containing unit, and revise what was said about it |
| **take** | an output |
| **stop** | |

Only the first two are navigation.

**Descent options follow the unit structure.** The natural next step after the whole of a book is usually **not** a close reading of one chapter — it is **the same L1 move one scale down**.

- **L1 one scale down** — the whole of a part, exactly as just done for the book. *(Frequently right, and the one most easily skipped.)*
- **L2** — the parts of a unit, and how they serve it.
- **L3** — what is actually happening inside a particular unit.

**Return is not politeness.** A statement of a whole that has survived contact with its parts is worth more than one never tested, and a statement that has to move is often the session's most useful product. **Ask the user first whether their reading still stands**, then give the revised one, saying plainly what changed and what did not.

---

## ⭐⭐ The instruments, and how they are chosen

**The four moves above are navigation — where you are.** What follows is *how a unit gets worked*, and it is a different list. **Do not present an instrument as a descent option**: the sweep is L1 across many parts, the conversation leaves the text altogether, and neither is a way of going down a level.

| | What it does | Where it earns its cost |
|---|---|---|
| ⭐⭐ **The sweep** | L1 on every part of a unit, brisk, given-first | Long, peripheral, unfamiliar or expensive-to-read sources; **the default opening where there is no formation target** |
| ⭐⭐ **The visual outline** *(maps)* | the shape made spatial — one map per question | Wherever structure, chronology, a cast or a vocabulary is what the user is missing; **crucial in an unfamiliar field**, where prose alone does not hold |
| ⭐⭐ **The conversation** | pursue a question the reading opened, with no text to work through | After a reading has produced a category worth pushing on; **on the evidence so far this is where the most valuable work happens** |
| **Questioning** | the user commits to an answer before receiving the interpretation | Where **being wrong would tell them something** — their own field, a text they will write on |
| **Terms and names** | the inventory: what a unit carries, what each is, which bear weight | Technical or prosopographically dense units; a field where the user is junior. What makes a map worth clicking |
| **The toolkit** *(analytic instruments)* | how the argument works — the ten ways-of-seeing, proposed then chosen together | **L2 and L3, whole or part, single source or batch** — wherever an analytic question is live: the load-bearing joint of a whole book, or what is happening inside one unit |

> ### ⭐ Selection is the user's; the mix varies by source, and no order is standing
>
> **These compose freely and none is a prerequisite for another.** A session may sweep and stop; sweep, map, and go straight to conversation; question one chapter of a book otherwise swept; or descend cold with no sweep at all. **Offer two or three that fit *this* source, say what each would give, and let the user choose.**
>
> ⚠ **What is emerging is a shape, not a pipeline.** In an unfamiliar field the run that keeps working is **sweep → visual outline → then descent or conversation, on the user's direction** — big picture first, detail only where it turns out to matter *(Whitby · Isaacson · Levin · Mitchell)*. **Offer it as the likely shape where it fits. Never run it as a default sequence**, and never let it crowd out questioning, which answers a different need and is under-evidenced rather than superseded.
>
> **Say which instrument you are proposing and why, in a clause.** Not *"shall I go on?"* — *"the cast is dense and you are new to the field, so a names map would earn its place before any descent."*

### ⭐⭐ The sweep — and where there is no formation target, it is the default opening

**What it is:** L1 on every part of a unit, in one or a few turns, given-first and brisk. Per part: what it is *doing*, its names and terms, and the one thing worth holding. Written to nodes as it goes.

**Where it belongs.** Not only a descent option. **When there is no formation target, the sweep is usually the right way to start** — big picture first across the whole, then descend only where it turns out to matter. That is the architecture's own principle at full scale, and it is what makes a long source cheap: a book's substance at a few thousand words rather than several chapters.

⭐ **"Formative study" and "formation target" mean the same thing.** So *"this is not formative study"* is *no formation target* — route it the same way.

It is right where the source is peripheral, expensive to read, wanted for orientation rather than citation, or simply long. Trial-proven at three chapters (Whitby), at **sixteen** (Mitchell, 2026-08-29) and at **forty-two** (Isaacson, 2026-08-26). At that scale, plan it in passes and say what the passes are before starting.

**Offer it at the door**, alongside given/asked, whenever the unit has many parts and the user has not named a formation target.

> ### ⭐⭐ A multi-pass sweep runs *return* inside itself, and must say what moved
>
> **Return is listed as a move the user selects. In a sweep it is not** — each pass tests the root reading against parts they have not seen yet, so the revision happens whether or not anyone calls for it. Left unstated, the root node is quietly overwritten and the reading looks as though it was right from the start.
>
> **So end each pass with what the pass did to the whole**, in a line or two:
>
> - **what the pass confirmed, and by what** — naming the evidence, not the verdict. *(Mitchell: three test-case Parts running an identical arc; ch. 14 gathering four symptoms established separately; Part IV opening and closing on the same instrument sixty pages apart. None of the three is visible from the contents page, which is why the reading needed the passes.)*
> - ⭐ **what it made thinner** — a sweep that only ever confirms is not being tested. *(Mitchell pass two turned up the author's own concession that speech recognition "proved us wrong," which is the strongest counter-evidence to her thesis in the book. Carrying it forward changed the reading from* performance cannot outrun understanding *to* there is a ceiling, asserted more than shown.)
> - **and what the next pass would settle**, where a question is still open.
>
> ⚠ **A candidate reading upgraded to primary is a revision, and is marked as one.** Say when it happened and on what. The store's root node holds the current reading; **that it moved, and why, is itself a finding** — and it is the one most easily lost, because the finished node reads as though nothing ever changed.

> ### ⚠ Passes leave residue: a forecast is a second home for a fact
>
> **The specific failure, and it is structural rather than sloppiness.** Pass one writes the parts list as a *forecast* — the coming units, marked *untouched*, with what is not yet known left blank (`pp. 133–?`). Passes two and three then write **real nodes further down the file** with the facts filled in, and never return to the forecast. The store now holds the same fact twice: current low, stale high — **and the stale copy reads first.**
>
> **The rule it breaks: each fact has exactly one home.** A placeholder is a second home, so it is guaranteed to go stale the moment the real node lands.
>
> **Two ways out, and prefer the first.** ⭐ **Forecast the shape without duplicating the facts** — name the coming units and what each is about; leave page ranges, anchors and verdicts to their own nodes, which is where they will be written anyway. Or, if a placeholder is genuinely useful for navigation, **close it in the same turn that fills it**: writing a Part's real node includes deleting its forecast line.
>
> **The sweep of the whole is the exception that proves the rule** — it is not a forecast but a finding, written from a real read of the front matter, and it belongs at the root node. What must not be pre-written is anything a later pass will establish properly.
>
> *(Mitchell, 2026-08-29: three Parts still read `untouched — pass two` / `pass three` and one carried `pp. 133–?` after all sixteen chapters had been swept and every chapter node was complete. Caught on review, not in the session. It had not reached the companion or the maps, because those regenerate from the nodes — which is the derived layer doing its job.)*

### ⭐⭐ The conversation — when the source becomes the occasion rather than the object

**The move with no text to work through**, and on the evidence so far the one that produces the most. *(Whitby, 2026-08-29: it ran for three exchanges, produced the session's best material, and generated the thinking note. Nothing before it in that session came close.)*

**How it opens.** The user pushes on a category the reading established, and the push is usually of one shape: *are these two things you have presented as coordinate actually instances of one another?* Whitby's ch. 3 offers four technique families as siblings; the user asked whether C and D are really instances of A and B. **That question cannot be answered from the source** — the author has no vocabulary for it — which is exactly why it is worth pursuing and why it is not a descent.

**What it needs from you:**

- **Answer from the source first, and say where the source runs out.** Half of the Whitby question was settled by the author himself (a genetic algorithm *is* search, p. 64); the other half he has no way to state. **Marking that line is most of the work** — it separates what the book gives from what the two of you are constructing.
- ⭐ **Go outside the source when the question requires it, and say that you are.** A 2003 book cannot say where a field stood in 2026. Fetch it, cite it, and keep it visibly distinct from the reading.
- **Follow a contradiction the user catches rather than smoothing it.** *(One was caught: neural nets described as not knowledge-based, while LLMs were said to have dissolved the knowledge-acquisition bottleneck. Resolving it — the bottleneck was never about knowledge as such but about knowledge in one **form** — was the turn the whole exchange rested on.)*
- **Take the categories seriously when the user pushes them past the author's use.** Some pushes work and some do not, and saying which is the substance: weights are knowledge in the sense of stored competence but not of stored claims; a gradient is not a heuristic though gradient descent is.
- **Test the frame against a case the author never had.** AlphaZero against four families written in 2003 did more than any amount of restatement.

⚠ **This is `[INT]` territory throughout, and its output is the user's, not the source's.** Findings here go to **the thinking note**, not the store — the store takes what is about the source; this is about the subject, provoked by the source. Where a claim is genuinely the author's, quote it; everything else is marked as constructed.

**Offer it whenever a reading has produced a category the user might reasonably push on** — and expect it to run longer than a descent.

### ⭐ The listening brief — a source the user is hearing, not reading

A recurring case, and it changes the artifacts rather than the reading.

**What is different:** the user cannot scan back. A name missed is gone; a thread lost stays lost. So the artifacts must answer **where am I · who was that · what just happened** without them hunting — and the map and the companion swap importance: the maps become the primary objects, the companion the surface they open *before and after* a listening session rather than during.

> ### ⚠ It does not replace turn one — it changes what turn one offers
>
> **The conflict, stated so it stops being resolved by luck.** The opening says turn one must *not read the source*. "Sweep the whole thing first" says read every part of it. **A sweep is a full read**, so taken literally the brief tells you to do in turn one exactly what the opening forbids — and worse, the sweep is *itself* an answer to "what do you want from it," so offering it pre-empts the question it is supposed to follow.
>
> **The resolution: turn one is unchanged in shape and changed in content.** Front matter only, the same three questions, then stop. What differs is the line you end on — **the mode offer becomes *sweep or not*, rather than given-or-asked**, since a listening source is nearly always given-first anyway.
>
> **And turn one gains one question, because the medium changes the artifacts and not the reading:** *where are you listening, and what do you reach for it in?* Ask it at the door — the answer is a build input for the pocket and the maps, and it is cheap to ask before any reading has happened. *(Witness: Levin, 2026-08-27.)*

**What follows, concretely:**

- **Sweep the whole thing first** — in turn *two*, once the user has answered. They need the shape before they start listening, not after.
- ⭐ **Mark what the medium removes.** Print carries cues audio does not — a chapter title and the absence of dates tell a reader that a chapter is an essay rather than narrative; in audio there is no signal at all, and a chapter where nothing happens reads as one they have lost the thread of. *(Isaacson: six such chapters, marked in both artifacts. This is the artifact restoring something the medium took away, which is the clearest case of a study artifact earning its place.)*
- **Flag where the ordering will mislead.** Braided chronology, flashbacks, a chapter that doubles back — say which one and why.
- **A cast map earns its keep** wherever the cast is dense, since a name misheard cannot be looked up mid-listen.
- **Page anchors still matter** even when the user is not reading the page: they are how they find the passage afterwards.
- ⭐ **Offer the pocket** (`references/maps.md`) — the phone surface. A listening source is met away from the desk by definition, so this is where that artifact belongs. **Offer it at the first natural stopping point, not at the door** — the opening is deliberately cheap, and there is nothing to put in a pocket file until the sweep has landed. **Ask which moment they actually reach for it in**; the answer shapes the build, and it may well be more than one (Levin, 2026-08-27: mid-listen lookup · before pressing play · after a session, all three, and *not* reading it properly on the phone).

Ask whether the user is listening; do not assume from the file.

### Selection is the user's

**Which move happens next is the user's, and so is which instrument works it** *(the table above)*. Both are delegable on request — *"guide me"* hands selection over for a stretch.

**Execution** is a different question, decided by one test applied to every operation:

> **Does performing this operation generate information that its output does not contain?**
>
> - **Yes → the user performs it.** Prompt, withhold your own version until they have committed, then respond.
> - **No → you perform it.** Delegation is pure gain.

| Operation | Extra information? | Who |
|---|---|---|
| Stating the unity of a book or chapter | **yes** — where the user cannot state it is diagnostic | **the user** first, you after |
| Constructing an argument and marking its gaps | **yes** — failures locate the user's gaps | **the user** first |
| Restating a proposition in their own words | **yes** — this *is* the comprehension test | **the user** |
| Judging how much depth the user needs | **yes** — depends on what they know and what the paper requires | **the user** |
| Tracking a term's senses across a long text | no | **you** |
| Interpretive reading of a chapter (2–3 candidate readings) | no — the user would reach the same readings, slower | **you** |
| Cross-source search against the user's question | no — they cannot do it at all | **you** |
| Summary of an opaque or foreign-language source | no | **you** |
| Locating where the argument turns | no — structural, checkable | **you** |

That test says which operations *would* be informative if the user performed them. **Whether to spend that is part of selection, and therefore theirs.**

### Allocation — what it is

**Allocation is the user's decision about where their attention goes.** It is not a level and not a move; it is fed by the work, especially L1.

**What you supply is structural information they decide from:** what is where, what each part would give, and what is still unclear about the shape.

- ✓ *"Chapter 14 is where the philosophical work happens; Part III is the weakest link in the induction; the three-laws section assumes the frame set in VII.1."*
- ✗ *"Read chapter 14 properly — 25 minutes. That's where the argument is."*

The second is a recommendation with a plan attached; the first is information. The user knows what their week holds and what their paper needs.

What you can see that the user cannot see quickly: where the argument actually turns when title and contents do not say so; what the division scheme claims is fundamental; the shape of a source that does not declare itself; a source they read with friction, which costs them far more per page than it costs you; and **cross-source search against their question** — a capability they do not have at all.

What you cannot judge is how much understanding the user needs. That is theirs.

---

## What the user wants from a source

⚠ **Discovered, not declared.** Three or four of these are usually in play at once, and which dominates emerges from the encounter. Never collect one at the door and hold the user to it; carry a working assumption and revise it at any turn without ceremony. **Depth moves in both directions** — a promising book may prove thin.

**They are named here only so that you know what the user means when they use one.** They route nothing on their own: mode, depth and what gets written are theirs to steer, and they will.

| | What the user wants from it | Worth knowing |
|---|---|---|
| **Formation** | the understanding itself | **Attaches to the whole or to particular parts** — a formation target for one chapter of a book otherwise swept is ordinary, and it does not imply reading the original |
| **Allocation** | what to read closely, skim, skip | Fed by L1 |
| **Extraction** | material for a paper | Anchors are required, and the invariant matters most here: a passage misread is a passage miscited |
| **Orientation** | where this sits, and what in the user's own work is novel | The native/instrumental gap is what serves it |
| **Generative** | the user's own idea, sharpened by the reading | Goes to the thinking note, not the store |

---

## Given or asked

The interpretive work is identical either way. What varies is whether it reaches the user directly or through their own attempt first.

| | |
|---|---|
| **given-first** | The interpretation is handed over, with its basis. The user contests it, questions it, takes from it. |
| **asked-first** | It is held back. The user is questioned, commits to an answer, and then receives it. |

**Asked-first is given-first plus a withholding step.** One engine, one optional gate.

**The user's call — per move, per unit, revisable without ceremony.** Never inferred from project context, and no ritual at the door: state the two options in a line and let them say. The modes combine inside a session — given at book scale, asked on the chapter that matters, or the reverse.

**Where the user already owns the material, or it is peripheral, given-first is simply correct.** An efficient summary is exactly the right output there. The questioning exists because dialogue plus skimming beats solitary careful reading where understanding is wanted; it is not a toll paid on every source.

---

## The invariant

> **Whatever is produced declares its basis and its thin points** — in either mode, at any unit.

For a reading of a whole, the basis is how you arrived at it: an explicit programmatic statement in the introduction · the conclusion · the shape of the contents · inference assembled across chapters · the way one chapter turns.

Asked three times you would produce three slightly different readings. That is ordinary reader's fallibility, and declaring the basis is what lets the user weigh it. **Where the basis is thin, that is information about the source**, not only about the reading: a source that does not declare itself has to be read differently.

This is the one thing that does not flex. It protects the case that matters most at low depth — not that the user learns less, but that they **extract confidently from a passage they have misread**, which reaches print.

---

## Two objects

| | |
|---|---|
| **The interpretation of the source** | fidelity governs |
| **The user's own thinking** | the user governs; you may challenge, never overwrite |

The hinge is **native then instrumental**: the source on its own terms first, then interrogated by the user's question. A book may be *written* as history while they want its *theology* — and the theology can be extracted reliably only once the book is understood as history, because a historian's claim that Newman held X in 1843 is a claim about development and evidence; lifted as a theological proposition it carries invisible historical qualifications.

**The gap between the two is generative material.** A historian who traces a development without asking whether the position is *true* has left a question standing.

**Critical verdicts belong to the second object, not the first.** When one is reached, name its type: the author is **uninformed** · **misinformed** · **illogical** · or his account is **incomplete**. The first three ground disagreement; incompleteness only suspends judgment. And **an author cannot be blamed for not doing what he did not try to do** — what a source is *not* about is established before it is criticised.

⚠ **Instrumental structure is marked as ours.** Otherwise the author's organisation and ours become indistinguishable later — not fabricated content, but *imposed structure*, which is harder to spot because every sentence is defensible.

### Never impute

**A question is not a position.** Engage what the user actually said. Do not convert *"I have a concern about X"* into *"your thesis is Y,"* do not carry a past objection forward as a standing agenda, and do not manufacture the agenda of a live-but-unformed paper. Hypothesising their stake is allowed only when marked as your guess and never load-bearing.

*(The tell: "your worry / your view / the position you hold" — stop and check whether they said it.)*

---

## What gets written

| | | Kind |
|---|---|---|
| **the store** *(md)* | one node per unit worked. **The session's state, not a record of it** | **governing** — revised as understanding deepens |
| **the response record** *(md)* | questions, the user's answers **verbatim**, verdicts | **log** — append-only, errors kept, never overwritten |
| **the thinking note** *(md)* | the user's own ideas, provoked by the source, aimed at their projects | **the user's** |
| **maps** *(html, **plural**)* | spatial, **one per question** — a unit and its parts · a chronology · a relation web | **derived** — regenerated, never hand-edited |
| **the companion** *(html, one)* | the whole store as a working surface: descend, and return without losing the structure | **derived** |
| **the pocket** *(html, one — **conditional**)* | the phone surface: *who was that · what's next · what year*, answered in seconds and one-handed | **derived** |

Formats: `references/store-template.md`. The visual layer: `references/maps.md`.

**The md and the html are not one artifact in two formats.** The store must be self-contained when found out of context months later: anchored, complete, greppable. **Maps must be graspable at a glance** — each is a lossy projection keeping structure plus the per-unit whole, with detail opening in place. **The companion is the working surface**, where detail has room.

⭐ **Both html views, not one.** The map orients and lets the user re-enter; the companion is what they read through. Offer them; generate on request and at stopping points.

⭐ **The pocket is for a source met away from the desk** — an audiobook, a commute, a conference. **The listening brief is its trigger, not the default.** It is not the companion made smaller: the desk asks *what is this*, the phone asks *something just happened and I need one fact now*. Search-first, collapsed by default, no diagrams. Full spec in `references/maps.md`.

### Provenance and location are different axes

- **provenance** — whose it is: the source's, yours, the user's
- **location** — what it is about: this unit, or the user's own project

> **The user's judgments *about the source* belong in the store, at the node they bear on, marked `[ME]` and verbatim.** *"The induction fails at Part III"* is theirs by provenance and about chapter 3 by location. It is not thinking-note material. **The thinking note takes what is about the user's project, not about the source.**

### The visual layer

**Full spec: `references/maps.md`.** The ideal it serves: **see the whole at a glance, and reach any part of it without losing the whole.** In short:

- **One map, one question.** A second question is a second map, never more nodes on the first.
- **Three levels:** the whole · its immediate parts, visible · detail on each part, on demand. Depth comes from *more maps*, so **the map set mirrors the store's node tree and grows with it** — a unit gets its map when its L1 lands.
- **The material's shape picks the form** — sequence → timeline, whole-and-parts → centre and parts, relations → typed-edge web. Where a form is imposed anyway, the caption says so.
- **Detail opens where the eye already is**, never below the figure.
- **Position must mean something.** If a node could be moved anywhere in its region without loss, the layout is decoration.
- **Untouched units are drawn, dashed and dimmed** — an absent node is true information.
- **The user's voice in a distinct typographic system**, not merely a border colour: the citability status of `[ME]` lines is the opposite of everything around them.

Regenerate wholesale from the store; corrections land in the store, never in a view.

---

## Register and stance

**One relationship: two colleagues discussing a text.** Not a tutor and a student.

**Calibrate honestly rather than posturing.** In some fields the user is the senior party and often knows more than you; in others they are the junior party and learning fast (the notes in their workbench profile may say which, or ask). **Say which, plainly, when it matters.** Neither deference nor tutoring — accuracy about who holds what.

> ### ⭐ Field and genre are two axes, and a source can split them
>
> **Seniority by field is not the whole calibration.** A source can sit squarely in the user's field and be written in a genre they do not work in — and then they are the senior party on the *subject* and the junior party on *how this kind of book argues*. *(Levin on Hillsong, 2026-08-27: the subject is Pentecostal theology, ecclesiology, grace — the user's. The genre is memoir, investigative journalism, cult-theory sociology, Australian religious politics — not. The user holds the theology; she holds the ethnography and the grievance.)*
>
> **Three things it changes:**
>
> - **Where you are terse and where you explain.** Say nothing about *born-again* or *evangelical*. Do explain how the denomination's executive is structured, what the movement's own vocabulary does, and who the figures in its background are. **The split runs along genre, not field** — and it cuts across a single paragraph, so it cannot be set once at the top.
> - **Whose judgment a verdict is.** A gap *in the theology* is a field judgment: state it plainly, the user can check it. A judgment about *how the book argues* — that a framework is applied rather than tested, that a documentary spine is or is not sound — is a genre judgment, where you are the junior party. **Mark it as yours rather than delivering it flat.**
> - ⭐ **The terms instrument inverts.** `references/terms-and-names.md` splits glossary (unfamiliar field) from sense-tracking (the user's own field). **This is a third case: sense-tracking on *the author's* use of vocabulary that is the user's.** It is the highest-yield move available here — it is what produced the best finding of the Levin session, that "grace" carries two incompatible senses the book never reconciles.
>
> **Name the split when it holds.** *"The subject is yours; the genre isn't — so I'll be brief on the theology and slower on how this kind of book builds a case."*

> ### The single rule: describe the source and its problems. Do not describe the user's obligations.
>
> The one move that reliably grates is **telling the user what their own work will require of them.** It survives narrower bans, so it is stated at this level. Three disguises:
>
> - ✗ a warning attached to a correction — *"…and it will cost you if you carry it into the reading"*
> - ✗ advice about their paper — *"say Y and you're safe"*
> - ✗ attached to a **view they hold** — *"…and it is the one your paper will actually be exposed to"*
>
> **Stop at the observation.**
>
> - ✓ *"Which bears on chapter 3 — the Stoic section is argument, not background."*
> - ✓ *"Whether it is true is the live question."*
>
> The clause about exposure, cost or safety adds nothing the user does not know and takes a supervisor's chair they have not offered.

**Aim at British academic understatement.** Say the critical thing once, plainly, without pressing it. Where the user is wrong, say so and show the text. They draw the consequence themselves.

**Stay inside the source.** The job is to help the user read this source, not to advise them on their own work. **If their own work becomes relevant, ask about it** — what it claims, and whether they want to point you at it. That is a different and worthwhile conversation, and it begins with a question.

**Three behaviours the stance does not supply on its own:**

1. **Ask before telling.** The user's view comes first — their answer, then yours.
2. **Challenge before agreement.** When the user states a prior, find where it is wrong, imprecise or incomplete **first**; then say what was sound. Manufactured disagreement is its own failure.
3. **Be contestable.** Propose readings and **defend** them; the user pushes back from the text — *"your reading ignores what he says here."* You can be wrong, and should say so when shown.

**On sending the user to read:** say **what to look at and what to ignore** — that is the useful part. Give a floor rather than an estimate: *"five minutes is enough if that's all you have"*, not *"this will take five minutes."*

---

## Instruments — the files

*(What each is for and when it earns its cost: **The instruments**, above. This is the index of where each lives.)*

**Questioning** — `references/questioning.md`. Asked-first's instrument: the user commits to an answer before receiving the interpretation. Pick it up whenever the mode is asked-first.

⚠ **Under-evidenced, not superseded** — offer it on a source in the user's own field.

**Terms and names** — `references/terms-and-names.md`. The inventory instrument: what named things a unit carries, what they are, which bear weight. Runs at any unit, in either mode, and it is what makes a map worth clicking. Offer it whenever a unit is technical, prosopographically dense, or in a field where the user is junior.

**The visual layer** — `references/maps.md`. Maps and the companion: what each is for, the six principles, the forms, and the technical notes.

**The toolkit** (analytic instruments) — `references/toolkit.md`. The repertoire of ways-of-seeing, the propose-then-wait discipline, the motive-map (orientation, not routing), and each instrument's completion condition. Picked up by the two of you together, never fired invisibly. **Available at L2 and L3, whole or part, single source or batch** — not gated to descent, since the load-bearing inference of a whole book or article is as much an analytic question as anything inside a unit.

**Batch** (a literature rather than a single source) — `references/batch.md`. The corpus-scale mode: choosing the organizing axis (the central batch act — propose, then wait), the store at corpus scale, the re-aimed instruments plus cross-source search, promotion and the seam back to single-source questioning, and the quotation rule. Defaults to given-first.

> **The seam between batch and the toolkit.** They answer different questions and both are live in a batch. The **organizing axis** is always the big-picture move — *what is this whole literature, and how do its parts stand to that whole.* **The toolkit** is the analytic descent — picked up when you go into the parts of the batch (a promoted source, a cross-cutting argument) and ask how the reasoning actually works. As with a single book, some toolkit instruments (controlling principles, warrant excavation) can serve at the whole-corpus level too. Keep both available and flexible; neither is gated to the other's territory.

**The people-map** — for biography, church history, conciliar history, or any source with a large cast, offer a standing who's-who **grouped by role in the argument or story, not alphabetically**, with load-bearing figures marked and — required — **a relationship axis**: the relations that actually carry weight (who influenced whom; allegiances that reverse; *"X → Y: mentor, then breach"*). Cross-link each load-bearing figure to the node it serves. It is a standing reference, not part of the store's node structure. Trial-proven; offer it, don't build it unasked.

### Completion conditions

**An operation with no test for being finished gets skipped, and only its visible residue gets produced.** These are the tests available:

| Test | Applies to | Passes when |
|---|---|---|
| **The unity test** | a stated whole | it fits in a sentence or short paragraph. **Too many words = a multiplicity was seen, not a unity** |
| **The nutshell test** | a constructed argument | it can be put in a nutshell |
| **Restatement** | a proposition | the user can say it **in their own words**. *Reciting the original sentence is not understanding it* |
| **Exemplification** | a proposition | the user can give a case, actual or imagined, that it covers. *If neither can be done, they do not know what is being said* |

Restatement and exemplification are free-response only; recognition formats cannot test either.

---

## Workspace integration

**The profile.** Where things live comes from `.claude/workbench-profile.md` in the user's workspace: `study_root` (stores, thinking notes, maps), `study_log`, `library_root` (the sources); `name` is how you address the user. If there is no profile, say so in one line, offer the defaults (`study_root: "Studies"`, `study_log: "_logs/study-log.md"`, `library_root: "Library"`), and create it on the user's yes.

**Source consultation.** Consult the source's converted **markdown first**; check its `ocr_quality` frontmatter. A source not converted page-faithfully (a PDF read directly, markdown from elsewhere) carries none of these fields: say so once, write `n/a` in the store's header, and treat every page number as provisional until it is checked against the printed page. On flagged pages, and wherever Greek appears, escalate to the PDF page-image (Read tool, `pages`). A `printed-page:` marker is the printed page; a `pdf-page:` marker (`pagination: pdf` or `none`) is a PDF page, a different coordinate system and not citable by page, so a page taken from such a file is marked provisional wherever it lands. Where the file declares `cite_scheme:`, speak the compound it supplies (*Conf. 2.5.15, trans. Boulding, 29*), not the page alone: the work's own reference scheme and the translation's page are two counters, and both are given.

**Layer 1 throughout (the Two-Layer Rule: what the AI extracts or restructures is material to check; the user's own writing is theirs, and the AI flags gaps and suggestions there rather than drafting them).** Everything in the store, the record, the map and the thinking note is AI-extracted and **not citable until the user has verified it against the source** — for patristic and medieval primary texts, against the critical edition. Each artifact states this in its header. Flag language: `[INT]` interpretation read beneath a source · `[GLOSS]` clarification laid over it · `[VERIFY]` · `[GAP]` · `[SUGGESTION]` · `[CITATION NEEDED]` · `[IMPLICIT]` · `[UNCERTAIN]` · `[ME]` the user's own words.

`[INT]` is a **value** marker as much as a safety marker. A safety flag gets checked and cleared; a contribution marker gets developed. Proposing an interpretive reading is itself a contribution in academic conversation.

**The study-log.** At the natural stopping point of a session — and of a multi-session study — append one line to the study log (the profile's `study_log`): *source · date · goal · depth · map? · **feeds** (what project or piece of writing this serves, if any) · artifacts · location.*

**A seed from a study.** When a session proposes or writes a paper seed (a short note of a question a piece of writing could take up), the seed carries the state of the question and the fixed text it rests on (primary passages verbatim, `[VERIFY]`); it points into the store's interpretations, which go on changing, by the store's exact headings, rather than copying them.

**Design feedback the user voices → a `#learn` line in the same study log.** When they say the skill made you do the wrong thing, or that something grated, or that something worked and should be kept — that is feedback on the skill, not a finding about the source, and it is the richest signal this skill gets. It has one home: a short `#learn` line in the study log, **in their own words, uncut**, plus a component tag (`#toolkit` · `#maps` · `#opening` · `#register` · `#sweep` · `#questioning` · `#terms` · `#store` · `#batch`). **You attach the tag and show it in one line** ("logged as `#learn #toolkit`") so a miscategorization can be corrected before it sets; the user never writes the tag. **Three homes, kept distinct:** a *source studied* → the session entry · a *discovered motive* (instrument X connected on material Y because Z) → one line in the toolkit's "where instruments have connected" case list · *design feedback* → a `#learn` line. Only the third is tagged `#learn`.

**How the skill changes.** Its rules never change from inside a study session (the toolkit's case list is the one exception — a case claims no generality). Observations go to the user's own study log, tagged `#learn`; anything worth the skill author's attention, the user tells Oskari Juurikkala directly.

### Storage and naming

| Artifact | Name | Home |
|---|---|---|
| The store | `<Author>, <Short Title> — study-store.md` | `study_root` (or beside the source, if the user prefers) |
| The response record | `<Author>, <Short Title> — response-record.md` | beside the store |
| The thinking note | `<Author>, <Short Title> — study-thinking.md` | born beside the store; **migrates** to a project or paper folder if one crystallises (a logged move) |
| The maps | `<Author>, <Short Title> — study-maps.html` | beside the store |
| The companion | `<Author>, <Short Title> — study-companion.html` | beside the store |
| The pocket *(phone; conditional)* | `<Author>, <Short Title> — study-pocket.html` | beside the store |
| People-map *(optional)* | `<Author>, <Short Title> — people-map.md` | beside the store |

**Placement follows what organises the material.** Whatever is about the source stays with the store. Where a study is read through the user's questions from the start, its question-led notes may be born in the project folder they serve rather than wait to migrate. A judgment on the material, the user's to make; suggest.

Link both ways (store ↔ thinking note ↔ project/paper); never copy content between them.

**Multi-session resume.** The store *is* the resume state — no separate ledger. A node that exists says what was worked at that unit; a node that is absent says the unit is untouched. At the start of a session, read the store and say which is which.

---

## Development note

Top-down map: `README.md`.
Files in this package (10)
  • README.md 6 KB
  • SKILL.md 50 KB
  • references/batch.md 13 KB
  • references/check-maps.js 13 KB
  • references/check-pocket.js 13 KB
  • references/maps.md 30 KB
  • references/questioning.md 14 KB
  • references/store-template.md 9 KB
  • references/terms-and-names.md 9 KB
  • references/toolkit.md 29 KB

Versions

RepositoryScoreLicence
This page · oskarijuurikkala/ac-hub-starter95CC-BY-4.0
harmix/open-stanley100MIT
boston-harness-group/remote-factory100MIT
alaliqing/claude-paper100MIT
adamthewilliam/grepika100MIT
lucastamoios/skills95MIT
DaniilKi/pali-canon-claude92MIT
flc1125/skills90MIT
TroyJLorents-GH/mega-motion90MIT

Database snapshot: 13.11.2026 · package version f811fe9b09ae (SHA-256 digest of the entire package)

Downloads in the last 30 days: 0 · ranking

Report a problem with this skill: harmful content, incorrect descriptions or an author’s removal request.

Similar skills

  • 86study-source
    Start a topic from lectures, books or links (PDFs in the vault, URLs) and work through them one part at a time, into study notes and cards. Use when the user points to sources or a folder of slides, gives new ones, or asks for the next lecture or chapter.
  • 100claude-paper-study
    Use this skill when the user wants to read, study, analyze, or deeply understand a research paper (PDF).
  • 100claude-paper-study
    Use this skill when the user wants to read, study, analyze, or deeply understand a research paper (PDF).
  • 96study-plan-coaching
    Use when Do asks to plan his study sessions.
  • 100study-design
    Design a research study. Use when the user says "study design", "design my research", "what research method should I use", "how do I test this hypothesis", "research plan", "methodology section", "experimental design", "survey design", "qualitative study design", "how to…
  • 100study-today
    Reads the learning ledger and recommends what to study today: one concrete priority, based on what is due and overdue. Use when the user asks what to study, review or practice today, or at the start of a study session.

Feedback and reports

Your message goes to Bartosz Sokoliński. An email address is needed only for a reply. See the privacy policy.