Reading Lens

Reading Lens

Active-reading scaffolding for nonfiction books. Given an EPUB, produce a structured set of markdown files in the user's Obsidian vault: book-level overview (0. General.md), per-chapter pre-read briefs (what to watch for), and per-chapter post-read syntheses (what it meant + contrarian takes + actionable experiments). The user takes their own notes in each chapter file — this skill is scaffolding, never a replacement for the user's own thinking.

Why This Exists

Reading nonfiction passively leaks most of its value. This skill pre-frames each chapter so the reader knows what to watch for, then produces a synthesis with per-lens takeaways, a contrarian take (what the book overclaims), and concrete experiments worth running. Everything lands in Obsidian, wikilinked, so the vault accumulates into durable knowledge.

Model & Thinking

• Default model: Opus 4.7 (claude-opus-4-7). This skill does qualitative analysis — use the strongest reasoning model available. Check with /model, switch if needed.
• Use extended thinking (<think> / ultrathinking) for three steps specifically: (1) generating 0. General.md, (2) generating a post-read synthesis, (3) generating a contrarian take. The quality bar is high — platitudes are a failure mode.

Sub-commands

If the user's intent clearly matches a sub-command but they didn't type the exact syntax, run the matching sub-command. Don't stall on syntax pedantry.

Configuration

Global settings live in {SKILL_DIR}/config.yaml. On first use, verify vault_path is set to a real directory — if it still says /path/to/your/obsidian/vault, abort and tell the user to edit it.

Config fields:

• vault_path — absolute path to the Obsidian vault (required)
• books_folder — subfolder name (default Reads)
• global_lenses — list of {role, interests} objects
• notion_parent_url — optional, for :to-notion

Vault Structure

{vault_path}/{books_folder}/{Book Name} ({Author})/
├── 0. General.md                 # Book-level overview (the "orientation" file)
├── 1. {chapter 1 title}.md       # Chapter notes: brief | my notes | synthesis
├── 2. {chapter 2 title}.md
└── n. {chapter n title}.md

Filename rules. For both the book folder and chapter filenames, sanitize thus:

• Replace : with - (space-hyphen-space) — preserves readability in titles like Start: A Parable → Start - A Parable
• Remove / \ * ? " < > |
• Collapse multiple spaces to one, trim leading/trailing whitespace
• Do NOT slugify — preserve capitalization and spaces (this is Obsidian, not URLs)

Examples.

• Book The Lean Startup by Eric Ries → folder The Lean Startup (Eric Ries)
• Chapter Start: A Parable → file 1. Start - A Parable.md
• Chapter What/When/How → file 3. WhatWhenHow.md (slashes removed)

Lens Selection Flow

Executed at the start of every /reading-lens:setup after EPUB parsing.

1. Read global_lenses from config.yaml.

2. If the list is empty: tell the user "No global lenses configured yet. What lenses should I use for this book?" Ask for role + interests pairs. Offer to save them globally after the user defines them.

3. If the list is non-empty: present the globals and ask:

   You have {N} global lenses:

   1. {Role 1} — {interests}
   2. {Role 2} — {interests} ...

   For {Book Title}, do you want to: (a) Use all global lenses (b) Choose a subset of globals (c) Define new lenses just for this book (ignoring globals) (d) Combine a subset of globals + new book-specific lenses

4. Based on the answer, resolve the final lens set. Confirm back to the user before proceeding: "Using these lenses for this book: [list]. Continue?"

5. Store chosen lenses in the book's 0. General.md frontmatter under lenses:. These become the book's frozen lens set — they override globals for this book going forward. Changing globals later does not rewrite this book.

Update-Don't-Delete Safety

Before writing any file, check if it already exists. Never overwrite silently. Specific rules:

On :setup when the book folder already exists:

1. List what's there: 0. General.md, which chapter files, and for each chapter file read the frontmatter to report brief_generated and synthesis_generated state.

2. Present status and ask:

   Book already exists at {path}. Current state:

   • 0. General.md [generated {date}]
   • 1. {title} [✓ brief, ✓ synthesis]
   • 2. {title} [✓ brief, — no synthesis]
   • 3. {title} [— untouched]

   What would you like to do? (a) Skip — leave everything as-is (b) Fill missing pieces only (generate briefs/syntheses for chapters that lack them) (c) Regenerate 0. General.md only (keeps per-book lenses + chapter files) (d) Add chapters that aren't in the vault yet (EPUB may have more than the folder) (e) Review and decide file-by-file

3. Never delete files. Never overwrite ## ✍️ My notes sections (the user's own writing).

On :brief N or :synthesize N when the section is already filled:

• Show the user the current content of ## Pre-read brief (or ## Post-read synthesis), say when it was generated, and ask: "Overwrite, keep, or append?" Default to "keep" on any ambiguous response.
• If the user chooses "append", add the new version under a ### Re-generated {ISO-date} subheader so both versions are preserved.

On :overview-redo:

• Show current 0. General.md summary (first paragraph + lens list). Ask: "Regenerate 0. General.md? Per-book lenses will be preserved. Overwrite? (y/n)". Only proceed on explicit yes.

Workflow

/reading-lens:setup <epub-path>

1. Validate. Check the path exists and ends in .epub. Abort with a clear error if not.

2. Read config. Load config.yaml from the skill directory. If vault_path is the default placeholder or not a directory, abort with: "Edit {SKILL_DIR}/config.yaml first — set vault_path."

3. Parse EPUB. Run python3 {SKILL_DIR}/scripts/parse_epub.py <epub-path> and capture JSON. On error, surface stderr to the user.

4. Build book folder path. {vault_path}/{books_folder}/{Book Name} ({Author})/. Sanitize as per filename rules.

5. Check for existing book. If the folder exists, run the Update-Don't-Delete flow above.

6. Resolve lenses. Run the Lens Selection Flow above.

7. Detect language for this book. Check the user's setup invocation message for a language directive (e.g., "do it in Turkish", "yap bunu Türkçe", "auf Deutsch"). If detected, set language: <code> for this book (e.g., tr, de). Default: en. Confirm the detected language back to the user before proceeding: "I'll generate this book's analysis in {Turkish}. Confirm?" Wait for confirmation, then continue.

8. Create folder + cache. Create the book folder. Save parsed EPUB JSON to {SKILL_DIR}/.cache/{book-folder-name}.json for fast re-use by :brief and :synthesize.

9. Generate 0. General.md AND extract book-level context for subagents. Use extended thinking. Source material: title, author, TOC (all chapter titles in order), and sample text (first 6000 chars of chapter 1 + first 6000 chars of chapter 2). Fill the Book Overview Template below. Before returning, hold onto three transportable artifacts for step 12 to hand to subagents:

   • Book thesis — the core argument in one sentence
   • Author-voice register — 2–3 sentences characterizing the author's voice (see close-reading extraction step 4)
   • Frozen lens list — as already resolved in step 6

10. Create chapter stubs. For each chapter, write a file with frontmatter + the three empty section headers (Pre-read brief, My notes, Post-read synthesis). Include synthesis_has_notes: false in the frontmatter.

11. Warn the user about duration. Print: "Generating pre-read briefs + post-read syntheses for all {N} chapters in parallel via subagents. This typically takes 5–15 minutes — longer for books with very long chapters or many chapters. Syntheses are generated WITHOUT your reading notes (you haven't read yet) — after you read a chapter and take notes in ## ✍️ My notes, re-run /reading-lens:synthesize <N> to upgrade that chapter's synthesis with your framing." Do not block on a confirmation — proceed.

12. Generate pre-read briefs AND post-read syntheses for all chapters — in parallel via subagents. Use Claude Code's Task tool to dispatch multiple subagents concurrently. Batching rule: aim for each subagent to receive no more than ~100,000 characters of chapter text total (to stay well under subagent context limits). A reasonable default: 4–8 subagents for a 32-chapter book (e.g., 32 chapters → 8 subagents handling 4 chapters each; 16 chapters → 4 subagents handling 4 chapters each; <8 chapters → one subagent). If the book has unusually long chapters (>6k words), drop to 2–3 chapters per subagent. Each subagent receives, in its prompt:

    • Its assigned chapters' titles, numbers, and full text (inline, pulled from the cached EPUB JSON)
    • The book's author-voice register extracted in step 9
    • The book's thesis extracted in step 9
    • The frozen lens list from step 6
    • The book's language (from step 7) — subagent generates output in this language
    • The full Pre-read Brief and Post-read Synthesis templates from the Templates section above
    • The Close-Reading Extraction Discipline instructions
    • A direct write mandate — each subagent writes the completed brief and synthesis straight into the corresponding chapter file using the Edit tool

    The orchestrator (main agent) issues all subagent Task calls in a single message to trigger concurrent execution. See superpowers:dispatching-parallel-agents for the fan-out pattern.

    Each subagent, for each assigned chapter:

    1. Runs the close-reading extraction pass (thesis, anchor phrases, defined terms, voice register refinement, structural role)
    2. Fills the Pre-read Brief template
    3. Fills the Post-read Synthesis template, prepending the > [!info] Generated at setup — no reader notes yet callout
    4. Updates the chapter file's frontmatter: brief_generated: <ISO-date>, synthesis_generated: <ISO-date>, synthesis_has_notes: false, status: synthesized
    5. Returns a short status line: chapter number, success/failure, any warnings

    After all subagents return, the orchestrator verifies every chapter file has both sections filled and reports aggregate completion.

13. Report. Tell the user: book title, author, N chapters parsed, briefs generated, syntheses generated (all flagged synthesis_has_notes: false), full folder path. Recommend: "Start with chapter 1. After each chapter, fill ## ✍️ My notes then re-run /reading-lens:synthesize N to upgrade that chapter's synthesis."

/reading-lens:brief <N>

1. Resolve current book. Find the most recently modified 0. General.md under {vault_path}/{books_folder}/. If multiple books have been touched recently, ask which one.
2. Read language from 0. General.md frontmatter. Generate output in that language. Preserve author-voice register across translation. Fallback: if the language: field is missing (pre-change book), default to en and proceed — don't prompt the user unless they asked for a language switch in their invocation.
3. Load chapter N. Read cached EPUB JSON from {SKILL_DIR}/.cache/. If missing, tell the user to run :setup again.
4. Check for existing brief. If ## Pre-read brief in the chapter file already has content, run the overwrite flow.
5. Generate brief. Using the chapter text (truncate to 40000 chars if longer) and the book's per-book lenses (from 0. General.md frontmatter), produce the Pre-read Brief matching the template below.
6. Write. Replace the content under ## Pre-read brief (or append per user choice). Update frontmatter: brief_generated: <ISO-date>, status: briefed.
7. Report. Print a 3-line summary and the path to the chapter file.

/reading-lens:synthesize <N>

1. Resolve current book. Same as :brief.
2. Read language from 0. General.md frontmatter. Generate output in that language. Preserve author-voice register across translation. Fallback: if the language: field is missing (pre-change book), default to en and proceed.
3. Load chapter N text and read the user's content from ## ✍️ My notes section if present.
4. Check for existing synthesis. If ## Post-read synthesis is filled:
   • If frontmatter says synthesis_has_notes: false AND the user has written real content under ## ✍️ My notes (more than just the placeholder *(take your notes...)*), offer a notes-aware upgrade: "Current synthesis was generated at setup from chapter text only. You've since written notes — upgrade the synthesis to weave them in? (y/n)". On yes: overwrite the synthesis, remove the setup > [!info] callout, set synthesis_has_notes: true, regenerate using notes.
   • Otherwise, run the standard overwrite flow (keep/overwrite/append).
5. Generate synthesis. USE EXTENDED THINKING. If ## ✍️ My notes has real content, reference it ("you noted X — this connects to..."). Produce the Post-read Synthesis matching the template below.
6. Write. Replace the content under ## Post-read synthesis. Update frontmatter: synthesis_generated: <ISO-date>, synthesis_has_notes: <true if notes were used, else false>, status: synthesized.
7. Report. Print a 3-line summary + flashcard count + path + whether this was a notes-aware generation.

/reading-lens:overview-redo

1. Resolve current book.
2. Detect language override. If the user's invocation includes a new language directive, update the language: field in frontmatter. Otherwise preserve the existing value. If the field is missing entirely (pre-change book), insert language: en as part of the regenerated frontmatter unless a new directive says otherwise.
3. Run overwrite flow for 0. General.md.
4. Preserve existing per-book lenses (from current frontmatter).
5. Use extended thinking. Regenerate the overview with the same source material as setup (TOC + sample from first two chapters).
6. Write, report.

/reading-lens:lenses [action]

Manage global_lenses in config.yaml.

• /reading-lens:lenses or /reading-lens:lenses list → print current globals
• /reading-lens:lenses add → prompt for role + interests, append to config
• /reading-lens:lenses edit <role> → show current values, prompt for new
• /reading-lens:lenses remove <role> → confirm, then remove
• /reading-lens:lenses reset → confirm, replace with built-in defaults (PM / CEO / Entrepreneur)

Edit config.yaml in place. Preserve comments and formatting.

/reading-lens:to-notion

1. Confirm Notion MCP is connected. If not, tell the user to enable it in settings.
2. Resolve current book.
3. Use notion_parent_url from config, or ask the user for the parent page.
4. Create a parent page from 0. General.md (strip Obsidian wikilinks, convert to plain references).
5. Create one subpage per chapter file.
6. Report Notion URLs.

/reading-lens:help

Print the sub-commands table, vault structure, where to edit config, how to manage lenses, and model recommendation (Opus 4.7).

Writing Voice

Every section of every brief and synthesis is written as the author speaking directly to the reader. First person throughout. Warm. Flowing. Complexity where earned, not performative.

• Pre-read framing — the author is about to walk the reader into the chapter: "Before you dive in, here's what I'm doing in this chapter…"
• Post-read framing — the author is reflecting with the reader: "Now that you've made it through, here's what I hoped landed, and here's where I'd give ground to a sharp critic."

First-person pronouns. Contractions are fine. No em-dash pileups (max one per paragraph; prefer sentence breaks). Plain English when plain English lands harder. Complexity only when the idea earns it.

Author-character channeling

Each book has a distinct voice register. Fitzpatrick reads different from Kahneman reads different from Taleb. The generator identifies the register during the close-reading pass (see next section) and channels it through every section. Examples of what the register looks like in practice:

The generator must not impose a single house style. Each book should feel genuinely different.

One explicit voice break

In the post-read synthesis, the Contrarian Take section briefly leaves author voice for a critic, then returns. Frame it as:

"A sharp reader will push back here, and they'd be right about X —"

then the specific pushback, concrete and biting. Then back to the author, graciously:

"I'd give ground on that. This is where my argument stretches further than my evidence can carry…"

Every other section stays in author voice.

Close-Reading Extraction Discipline

Required workflow step. Before filling any pre-read or post-read template, the generator runs an internal extraction pass against the chapter text. The extraction output stays internal — it is the scratchpad the generator reasons over, and it does not appear in the chapter file.

Each bullet, term, and per-lens paragraph in the final output must be traceable to extracted material.

The extraction pass (per chapter)

1. Thesis — the one sentence from the chapter that, reduced to ~25 words, survives. Verbatim or close paraphrase.
2. Anchor phrases — 8–12 verbatim phrases, metaphors, vivid images, canonical examples. The specific words the author uses ("bulldozer", "toothbrush", "fragile", "gold dust", "earlyvangelist", etc.). Every Watch-For bullet must attach to one.
3. Defined terms — 4–8 terms the author actively defines in the chapter, along with the author's own working definition (not a paraphrase). Pulled from the text, not invented.
4. Author-voice register — 2–3 sentences characterizing the author's voice in this chapter. Colloquial / academic / polemical / warm / etc. Specific turns of phrase. This register drives every section's voice.
5. Structural role — 1–2 sentences on what this chapter does in the book's argument. Example: "It's the 'why rigor matters' beat before the rules in Ch 4."

Enforcement rule

Do not proceed to template filling until extraction is complete. Each bullet, term, and per-lens paragraph in the output must be traceable to an extracted item. If a Watch-For bullet has no anchor, if a Key Term has no source in the text, or if a per-lens paragraph has no extracted insight to ground on, return to the extraction pass before writing.

Where extraction runs

• /reading-lens:brief N — the main agent runs extraction + generation for the single chapter.
• /reading-lens:synthesize N — same.
• /reading-lens:setup — each subagent runs extraction + generation for its assigned chapter batch. The book-level author-voice register (identified during 0. General.md generation) is passed in as seed context; subagents refine it per chapter.

Templates

Book Overview — content for 0. General.md

---
type: book
title: "{title}"
author: "{author}"
source_epub: "{absolute-epub-path}"
status: reading
started: {YYYY-MM-DD}
chapters: {N}
language: {ISO 639-1 code, default: en}
lenses:
  - role: "{Role 1}"
    interests: "{interests}"
  - role: "{Role 2}"
    interests: "{interests}"
tags: [book, reading-lens]
---

# {title}
*by {author}*

> **{one-liner — 1 sentence on what this book is really about, not the marketing blurb}**

## Thesis

{2–3 sentences: the core argument the book is making}

## Why it matters to a builder

{2–3 sentences: why a reader with the configured lenses should read this specifically}

## Key terms this book teaches

- **{term}** — {working definition under 20 words}
- ... (6–10 items)

## Focus prompts — hold these while reading

- [ ] {active-reading question to hold across the whole book}
- ... (4–6 items, checkboxes)

## Per-lens learning

### {Role 1}
{1–2 sentences: what this reader, in this role, stands to gain from this book specifically — not generic "will learn about strategy"}

### {Role 2}
...

(one H3 per lens)

## Converses with

- {other book, framework, or resource} — {1 sentence on the link}
- ... (3–5 items)

## Chapters

- [[1. {chapter 1 title}|1. {chapter 1 title}]]
- [[2. {chapter 2 title}|2. {chapter 2 title}]]
- ...

Chapter Stub (written at setup, sections filled later)

---
type: chapter
book: "[[0. General|{book-title}]]"
chapter_number: {N}
chapter_title: "{title}"
word_count: {wc}
status: unread
brief_generated: null
synthesis_generated: null
synthesis_has_notes: false
tags: [chapter, reading-lens]
---

# {N}. {chapter_title}

## Pre-read brief

*(run `/reading-lens:brief {N}` before reading)*

---

## ✍️ My notes

*(take your notes here while reading — this section is never overwritten by the skill)*

---

## Post-read synthesis

*(run `/reading-lens:synthesize {N}` after reading)*

Pre-read Brief — fills ## Pre-read brief

Voice: First person. The author is about to walk the reader into the chapter. See the Writing Voice section above. Channel the author-voice register extracted in close-reading step 4.

Notation: [...] marks a placeholder filled during generation. Placeholder descriptions are plain text — rendered output is plain prose, not italic, unless *…* is shown literally.

> **[Author-voice one-line thesis — their own framing, not a
> summary of what the chapter covers]**

[Paragraph 1: 2–3 sentences. Author's voice. A personal opener —
"here's where I'm about to take you." Introduces the chapter's
concerns in the author's register.]

[Paragraph 2: 2–4 sentences. More explanatory. Gives the reader
enough context to understand what they're about to encounter.
Places the chapter in the book's arc — e.g., "I made the
asymmetric-risk argument back in Chapter 3 — this chapter builds
on that by showing you the three species of bad data I keep
seeing…"]

**Watch for**
- [Phrase-anchored bullet, 1–3 sentences, author voice. Cites a
  specific word, metaphor, anecdote, or moment from the chapter.
  Example: "When I drop the bulldozer metaphor in a bit, I mean
  something very specific — the founder who asks 'do you think
  this is a good idea?' and treats the nod they get as data."]
- [3–5 bullets total. Each grounded in an extracted anchor phrase
  from close-reading step 2.]

**Key terms**
- **Term** — Author-voice definition, 1–2 sentences. The author's
  own working vocabulary. No academic paraphrase. Sourced from
  close-reading step 3.
- [3–6 term→definition pairs total.]

**Questions to hold**
- [ ] [Open philosophical question the chapter raises, framed as
  the author asking the reader to wrestle with it. Not a personal-
  action question — the chapter's unresolved tensions.]
- [ ] [3–5 questions total.]

**Per-lens angles**

*{Role 1}.* [3–5 sentences of flowing author prose, directly
addressed to this lens. Names what the chapter asks of this role
specifically. Depth enough that the reader recognizes themselves
in it. Not a bullet disguised as a paragraph.]

*{Role 2}.* [Same depth, for this lens.]

*{Role N}.* [Same depth, for this lens. Always include every
configured lens as a separate paragraph — never collapse two
lenses into one entry.]

Flexibility rules:

• If the chapter is structurally a comparison (two patterns, two failure modes, before/after dialogue), render Watch For or Key Terms as a 2-column markdown table instead of a bulleted list.
• If the chapter describes a sequential process, Watch For may become a numbered list.
• Per-lens angles never collapse — always one paragraph per configured lens in the declared order, even if lenses share overlap.
• Inline wikilinks allowed when the chapter references an earlier chapter explicitly: [[3. Talking to customers is hard|Chapter 3]].

Post-read Synthesis — fills ## Post-read synthesis

Voice: First person. The author is reflecting with the reader after the chapter. Channel the author-voice register. The Contrarian Take section is the one explicit voice-break — see Writing Voice above.

> **[Author-voice one-line core idea — the sentence they'd want
> to survive this chapter]**

[Paragraph 1: 2–3 sentences in author voice. Reflecting on what
was just read together. Example: "Now that you've made it through,
here's what I most hoped landed, and here's why I built the
chapter the way I did."]

[Paragraph 2 (optional, chapter-dependent): 2–3 sentences. Where
the chapter fits in the book's running argument now that it has
done its work.]

**Key takeaways**
- [Substantial bullet, 1–2 sentences per bullet, author voice.
  4–6 total. No clipped labels — each bullet reads as a complete
  thought.]
- [Final bullet names the mental model the chapter builds, as a
  takeaway sentence. Example: "The model I hope you walk away
  with: false positives are asymmetric bets against your own
  runway."]

**Per-lens takeaways**

*{Role 1}.* [3–5 sentences of flowing author prose, directly
addressed to this lens. Monday-morning specific — what changes
in their week because of this chapter. Never a single sentence.]

*{Role 2}.* [Same depth.]

*{Role N}.* [Same depth, one paragraph per configured lens.]

**Contrarian take**

[2–3 paragraphs. Voice break: the critic speaks first.

"A sharp reader will push back here, and they'd be right about X —"
then the specific pushback, concrete and biting. Name the actual
claim that overreaches, not a vague "some might disagree."

Then back to author voice, graciously: "I'd give ground on that.
This is where my argument stretches further than my evidence can
carry…" and a short concession of where the critique lands.]

**Actionable experiments**
- [Concrete experiment a builder could run this quarter, 1–2
  sentences, author voice. 2–3 total. Scaled to a small team's
  reality.]

**Connections**
- **[Other book / framework]** — [1-sentence link in author
  voice. Example: "*Lean Startup* (Ries) — he wrote the
  learning-loop theory; I'm writing the script for what to say
  inside the loop."]
- [3–5 connections total.]

**Flashcards**
- **Q:** [Sharp question that tests the chapter's core move]
  **A:** [Under 30 words.]
- [3–5 pairs total.]

Quality Rules

These rules are not optional. They convert the reading-lens output from a competent book summarizer into a genuinely useful reading companion. Enforce them on every generation.

Voice and flow

• Author voice throughout. First person. Channel the author-voice register identified in close-reading step 4. See the Writing Voice section.
• Flow over friction. Sentences must read aloud smoothly. No em-dash pileups (max one per paragraph). No noun-stacked Latinate phrases when plain English does the job. Contractions are fine. Complexity only when the idea earns it.
• No generic MBA platitudes. Every takeaway must be specific enough that a reader who skipped the chapter could still lose an argument with one who read it. Bad: "focus on customers." Good: "Ries argues the unit of progress is the validated learning cycle — not revenue, users, or ship velocity."

Grounding

• Every Watch-For bullet must cite a specific phrase, metaphor, anecdote, or moment from the chapter text. No unanchored observations. If there's no anchor, cut the bullet.
• Every Key Term must come from the author's own vocabulary. Pulled from the chapter text (close-reading step 3), not invented.
• The close-reading extraction pass is mandatory. Do not proceed to template filling without it.

Per-lens rigor

• Per-lens angles never collapse. Every configured lens gets its own paragraph (3–5 sentences) in the declared lens order. Never combine two lenses into one line.
• Lens paragraphs must be role-specific. If a PM paragraph would read identically as a CEO paragraph, both are wrong. Rewrite until they diverge.
• Monday-morning specificity in post-read takeaways. A PM should be able to name one thing that changes in their week because of the chapter. Same for CEO and Entrepreneur.

Contrarian rigor

• Contrarian takes must bite. "Some readers might disagree" is not a contrarian take. Name the specific claim and say why it overreaches, ignores something, or is wrong.
• The voice-break pattern is required. The critic speaks first, then the author graciously concedes. See Writing Voice → One explicit voice break.

Other

• Actionable experiments must fit a small startup. Not "hire a data science team" — "add an event to track X in the existing analytics for Y feature, run for 2 weeks, compare to the Z cohort."
• Never fabricate. If the chapter text is thin or the topic is outside your confidence, say so in-line rather than inventing.
• Respect the user's notes. Never modify ## ✍️ My notes. When generating a notes-aware synthesis upgrade, read the user's content and weave it in — their framing is data.
• Never delete files. Ever. When in doubt, ask.
• Ask before overwriting. Every regeneration is a deliberate choice, confirmed by the user.

What This Skill Is NOT

• Not a reading app — reading happens in your EPUB reader of choice.
• Not a summary generator — it's scaffolding around the user's own active reading.
• Not a research agent — it works on one EPUB at a time. Don't web-search unless verifying a specific proper noun.
• Not a general note-taker — other skills (e.g., no-brainer) handle general summarization.

Help Output (/reading-lens:help)

When this sub-command is invoked, present:

Reading Lens

Active-reading scaffolding for nonfiction books. Parses an EPUB → generates a structured book analysis in your Obsidian vault.

Where things go

{vault_path}/{books_folder}/{Book Name} ({Author})/

• 0. General.md — book overview
• 1. {chapter title}.md ... n. {chapter title}.md — chapter files (brief | your notes | synthesis)

Commands

Recommended

Use Opus 4.7. Check with /model.

Setup

Edit config.yaml in the skill folder — set vault_path and (optionally) edit global_lenses.