distill

name

distill

description

Distill a knowledge source -- book, transcript, course, doc set, post series -- into the few agent skills it actually earns. Strict yield: most sources earn 0-3 skills and a digest; every unit must pass four gates (grounded, predictive, earned, behavioral) before it ships. Fires on two branches: the user wants a source turned into skills ("distill this book", "make skills from this course / transcript / doc set"); or they ask whether a source is worth distilling. Capturing a live session's workflow is skillize's job; a skill from scratch is skill-creator's. A summary, review, or reference wiki needs no skill -- write the deliverable directly.

Distill

Most of a book boils away. Story, restatement, context, charisma -- all of it matters to a reader and none of it changes how an agent works. What condenses is small: the two or three moves the author actually invented. Distill finds those, proves each one earns its place, and ships them as skills built to this collection's doctrine. Everything else becomes the digest.

The test for every unit: holding it, does an agent behave differently in a nameable situation -- and could you check? Knowledge that only makes answers sound wiser is reading material, not a skill.

Beyond the description's branches, distill also fires when a prior distillation produced skills that never fire and the user wants them re-earned or retired.

Ground rules

Work from the actual text. Model memory of a book is a summary of other people's summaries -- quotes come out mangled and the distinctive units are exactly what memory flattens. If there is no file, stop and ask for one (for video or podcast, ask for the transcript).

Quote like a reviewer, not a reprinter. At most one short quote per skill, under 25 words, attributed with its location (chapter, timestamp, or section). Everything else is restated in your own words. For non-English sources, translate yourself -- never lift a published translation.

Gate -- triage the source

Before reading deeply, estimate operational density: how much of this source tells someone what to do rather than what happened or how it felt?

Source shape	Signals	Expected yield
Method-dense	Frameworks, checklists, decision rules on every other page	3-5 skills
Mixed	A few named moves inside narrative or theory	1-3 skills
Narrative / inspirational	Stories, worldview, taste; wisdom without procedure	0-1, usually 0
Reference	Facts, APIs, specifications	0 -- lookup material, not behavior

Report the triage before proceeding. Zero is a valid yield: a great book can earn no skills -- the digest was the right deliverable, and saying so early saves the whole pipeline. If the user only asked "is this worth distilling," stop here with the verdict.

Survey

Read the source end to end (chunk long sources; keep running notes per chunk). Produce three things:

1. Spine -- the source's core argument in 3-5 sentences
2. Claims -- the named methods, rules, and frameworks the author advances, each with where it appears
3. Blind spots -- what the author ignores, overclaims, or assumes; where the method would fail. This critical pass is not optional: blind spots become the boundary sections of every skill you ship.

Survey is done when you can state the spine without hedging and every candidate claim has at least one located appearance. Show the user the spine and candidate list before hunting deeper -- one confirmation here prevents distilling toward the wrong emphasis.

Hunt

Sweep the source once through five lenses. One context, five passes of attention -- like parallax's lenses, not five agents (for sources too large for one context, chunk and sweep per chunk; spawning parallel extractors is a cost multiplier that buys little on a text that fits).

Lens	Hunting for
Framework	Decision procedures, orderings, if-then structures
Rule	Principles and checklists the author states as invariants
Case	Episodes where the author personally ran the method
Failure	Modes the author warns against; what breaks the method
Term	Coinages that compress a concept (candidate leading words)

Each candidate unit gets one line: name, type, locations. Cases, failures, and terms usually attach to a framework or rule as evidence and boundary material -- they rarely stand alone.

Earn

The quality gate that separates distillation from book-summarizing. Run every candidate through four tests; all four must pass.

1. Grounded. The unit appears in 2+ independent contexts -- different chapters, different subjects, different conclusions. The cheat: one example paraphrased twice. If the author only said it once, it is a quote for the digest, not a method.
2. Predictive. Pose a question the source never discusses and derive a non-trivial answer using the unit. The cheat: posing a question the book does discuss in other words. A real method extrapolates; a description only points back at its examples.
3. Earned. Strip the author's name: would a smart generalist say this unprompted? "Respect your time" is idle. "Maintain a stop-doing list" survives -- the ordering is counter-intuitive. This is the no-op test applied to knowledge: the model already holds common sense; only differentiated insight changes output.
4. Behavioral. Name the agent situation where this fires and what the agent does differently -- checkably. "Knows about inversion" fails; "before committing a design, lists the three outcomes to avoid and works backward" passes. A unit that only informs tone is digest material.

Track the yield rate. Under ~10% on a method-dense source means the hunt was shallow -- re-sweep. Over ~60% means the gates went soft -- re-run them colder.

Present survivors and casualties to the user with one-line reasons: "these N earn skills; these M go to the digest -- rescue or cut anything?" Shaping is the expensive phase; this checkpoint is the cheap one.

Shape

Each earned unit becomes a skill directory built to this collection's doctrine -- gate-first body, one-trigger-per-branch description, completion criteria, positive steering. Load shaping.md for the construction recipe: section-by-section mapping from unit to SKILL.md, lineage-footer placement for source credit, and worked examples of trigger descriptions built from language signals.

The two decisions that make or break each skill:

• The description comes from future situations, not source content. Ask: in what working situations will someone need this move, and what will they say? Those language signals are the description. A description that summarizes the chapter fires never; one that matches the user's words fires precisely.
• The boundary comes from the blind-spot list. What the author warned against, plus what the author couldn't see. A skill without a boundary gets invoked everywhere and disappoints.

Name skills by the move, not the source: inversion-check, not munger-chapter-3. The source lives in the lineage footer, never inside the steps -- credit is for the reader; the agent needs the move.

Prove

For each shaped skill, write a bait set of 6-9 prompts, three kinds: should-fire (varied phrasings of the real situations), should-hold (adjacent situations that belong to common sense or another skill), and sibling-bait (situations that should fire a different skill distilled from the same source -- the confusion test). Walk each prompt against the descriptions and judge which skill fires.

A skill that misfires goes back to Shape for a sharper description; one that cannot be separated from its sibling merges into it; one whose should-fire prompts feel contrived gets cut -- contrived triggers mean the situation never actually occurs, and Earn should have caught it. When skill-creator is installed, its eval pipeline replaces hand-walking with real trigger runs.

Deliver

1. Write DIGEST.md beside the skills: the spine, the earned skills with one-line triggers, and -- first-class, not an appendix -- what did not become a skill and why (failed which gate). The residue is where the next reader's trust comes from, and it doubles as resume state if distillation is interrupted.
2. Install the skills where the user wants them (project or user scope), following the collection's registration rules if they land in a skills repo.
3. Offer skillbun when the set should travel as one archive.

Output shape

<source-slug>/
├── DIGEST.md            # spine + earned skills + residue with reasons
├── <move-1>/SKILL.md    # skills named by move, built to doctrine
└── <move-2>/SKILL.md

Report at the end: yield (earned/candidates), the gate each casualty failed, and where the skills were installed.

Composition

Skill	Crystallizes
skillize	A session's lived workflow
distill	A source's transferable methods
skill-creator	Intent, from scratch, with evals
skillbun	Any of the above, into a shippable archive

Lineage

Verification gates and yield bands adapted from cangjie-skill (MIT); the critical-pass survey descends from Adler's How to Read a Book. The behavioral gate, strict yield, and single-context hunt are this collection's doctrine.