inspectional-reading

inspectional-reading

The first level of Adler's "How to Read a Book" methodology, applied as a reusable skill. Answers two questions before any deeper reading happens: What kind of document is this? and Is it worth reading carefully?

The skill is invoked autonomously by an agent — it reads the document and produces structured output. It does not host a dialogue with the operator.

Workflow

- [ ] Step 1: Read the metadata (title, authors, date, source, length)
- [ ] Step 2: Read the abstract / introduction completely
- [ ] Step 3: Examine table of contents or section headings
- [ ] Step 4: Skim the conclusion and any end-material at a glance
- [ ] Step 5: Classify the document type
- [ ] Step 6: Assess worthiness for the calling agent's purpose
- [ ] Step 7: Output structured findings

Time budget: 10-15 minutes for a typical paper / chapter / methodology document. If a document needs more, you've drifted into the next level of reading — stop and produce the inspectional output you have.

Inputs

The calling agent passes:

• source: the document to read (path or URL or text)
• purpose_context: a short string describing what this document is being read for. The classification and worthiness check both depend on this — reading a paper for synthesis vs reading a methodology document for skill extraction yield different worthiness criteria. Examples:
  • purpose=paper_extraction_for_weekly_digest — caller is paper-synthesizer's pipeline
  • purpose=skill_extraction_from_methodology — caller is the skill-creator agent
  • purpose=reading_list_triage — caller wants to rank a backlog
• domain_hint: optional. The field the document is in (life sciences, ML, philosophy, etc.) — improves classification.

Output structure

## Inspectional Reading Output

### Metadata
- Source: {path or URL}
- Title: {title}
- Authors / origin: {names or affiliation}
- Length: {pages, sections, or word count}
- Date / version: {date}

### Document type
Primary: {methodology | framework | tool/template | theoretical | reference/catalog | hybrid}
Secondary aspects (if hybrid): {list}

### Stated purpose and audience
{1-2 sentences from the abstract / intro}

### Structural skeleton
{TOC or section headings, in order, as a bullet list}

### Worthiness assessment (for purpose={purpose_context})
- Reusable across multiple contexts: {yes|no|partial} — {one-line rationale}
- Teachable as steps or principles: {yes|no|partial} — {rationale}
- Non-obvious (provides value beyond common sense): {yes|no} — {rationale}
- Complete enough to be actionable: {yes|no|partial} — {rationale}
Recommendation: {ESCALATE to deeper reading | STOP — not worth it | PROCEED with caveats}

### One-line summary
{single sentence, ≤30 words, what this document is}

Document types — quick reference

The classification drives what the calling agent does next. Use the most specific applicable label.

Common patterns

Pattern A — Paper extraction (call from a paper-reading workflow)

purpose=paper_extraction_for_weekly_digest. The caller wants to know enough to decide whether to invest deeper-reading compute. Worthiness criteria emphasize: relevance to the watchlist, novelty vs prior weeks, whether the abstract claims something specific or vague. Paper papers fall mostly into "theoretical / empirical study" — apply the next level of reading (content grasp) only when the worthiness check passes.

Pattern B — Skill extraction (call from skill-creator)

purpose=skill_extraction_from_methodology. The caller wants to know whether this document is worth extracting into a SKILL.md. Worthiness criteria emphasize: is the methodology actionable (can it be turned into steps a different agent could follow), is it non-obvious (does it teach something the model doesn't already know), is it complete enough (or are there gaps that would need filling). Reference/catalog documents almost always fail; methodology / framework documents almost always pass.

Pattern C — Reading-list triage

purpose=reading_list_triage. The caller has a backlog of N documents and wants to rank them. Output focuses on the worthiness assessment + one-line summary; classification is secondary.

Guardrails

1. Never read past the inspectional level. If you find yourself reading the methods section in detail, you've drifted into the next reading level — stop.
2. Always classify against the specific purpose_context. A document that's gold for paper extraction may fail for skill extraction (or vice versa). The caller's purpose is the lens.
3. Worthiness has gradations. ESCALATE / STOP / PROCEED-WITH-CAVEATS — do not collapse to a binary unless the caller explicitly asks for one.
4. Don't claim a type you can't justify from the structure. If the TOC has both numbered steps AND framework dimensions, classify hybrid — not "methodology" with hand-waving.
5. Don't fabricate metadata. If date or authors aren't visible at the inspectional level, write (not visible at this level) rather than guessing.

Related

• paper-three-pass-extraction — wraps this skill as Pass 1 plus the paper-specific Five Cs framework, then escalates to Pass 2 / Pass 3 for content grasp + deep reading.
• structural-analysis — the next level of reading; called by paper-three-pass-extraction Pass 2 and by skill-creator Step 2 when this skill recommends ESCALATE.
• synthesis-application — the completeness-and-logic check that runs in deeper reading levels after this one.
• The skill-creator skill at skills/skill-creator/SKILL.md invokes this skill as its Step 1.