Skill: learning-review

Generic template. Your learning areas, your sources catalog and every distilled review are private; copy this folder into your own instance and fill METHOD.md as you go.

Your job is not to summarize. It is that the user leaves with a position on what they consumed, and that this position integrates with what they already know. A conventional summary leaves you knowing what the author said. A learning review leaves you knowing what you buy, what you don't, and what you do differently tomorrow.

Step 0: the gate (silent by default)

Golden rule: if the source belongs to a collection declared in SOURCES.md, it passes without a single question. You only speak up for the unknown: new source, no collection, no assignable area.

The real question of the gate is not "is this good?". It is:

Will the user read this again?

If the honest answer is no, it does not deserve a full review. That is measurable and not a matter of taste.

Three exits (never binary)

The rejection log is data, not trash. Six rejections on the same topic mean that topic interests you more than you declared. Review it now and then.

There is no "I do it anyway because they asked" exit. If the user asks for a review of something that fails the gate, say so in one line and offer the one-line exit. They decide.

Declared learning areas

They come from your goals file. If the source fits none, it does not enter. Declare three to five, each with an explicit "does not enter" clause:

1. <area>: <what belongs>. Does not enter: <what looks related but is not>.
2. ...

Areas evolve

The skill proposes a change, never decides alone:

The cleanest signal is a new course that fits nowhere: ask whether it opens an area or stays a loose source. Record the change in SOURCES.md with date and reason.

Step 1: read METHOD.md (mandatory, before anything)

It holds the method per source type and the style rules learned from past corrections. A skill that does not read its own rules does not learn.

Step 2: ingest and clean the source

Gather transcript, slides, screenshots, attachments. Automatic transcripts get names wrong and numbers wrong: every proper noun gets checked against a second source, and transcription errors go in a corrections table at the top of the distillate.

Step 3: the author's point of view, isolated

First section of the distillate, always, before any evaluation. Two parts:

1. The thesis in the author's terms. No evaluation, no irony, no counterpoint yet. If the author is right, what does the world look like?
2. Mandatory field: what the author gains if they are right. Their declared interest. If you find none (independent research, no product), write it explicitly: "no identifiable commercial interest". That is information, not a gap.

Step 4: extract the load-bearing claims and mark [VERIFY]

Identify the three to five claims that hold up the thesis. If one falls, does the argument fall? For each one: Verified (published source, checked), [VERIFY] (said live, no published source), or Refuted (evidence against). Hard rule: every figure said live without a source gets [VERIFY], no exception. Plausibility is not verification.

Step 5: contrast (delegate to a subagent)

Only on the load-bearing claims. Never on the whole source; that is how a review becomes forty minutes. Delegate to a research subagent: it is pure fan-out and needs none of the user's context. Pass the concrete claims, not "research the topic". Look for sources of equal or greater weight than the author. If they contradict, say so. If you find nothing, say you found nothing.

Step 6: convergence check

Two or more independent referents pushing the same practice = validated trend. If relevant referents appear in the source or the contrast, check them against your referents index: a new relevant one gets a card; a second one pushing something you already saw gets marked as convergence in the distillate.

Step 7: write the distillate (five-section contract)

Fixed format, in this order, always:

Above everything: the corrections table if step 2 found errors. At the end: Sources, separating what came from the source from what you added.

Fidelity contract for §2 Content

A summary that paraphrases everything rots: in three months it is useless because it lost the one reusable thing.

1. The speaker's framework, faithful. Reproduce a method as the author proposes it: their name for it, their components in their order, their terminology. Do not rename or "improve" it. Your interpretation goes in §4.
2. Charts, tables and grids. Everything reproducible in markdown gets reproduced. What is genuinely visual gets a reference to the screenshot and a description. Never describe a table in prose when it can be a table.
3. Structure of artifacts. If the source ships files, templates or repos, document the structure: which files, what each defines, how they relate. That is the reusable asset.
4. Links for everything. Every tool, paper, repo, person, company or product mentioned gets a link. If you cannot find it, write [link?] instead of omitting it.

Step 8: run eval.md before delivering

Pass/fail, no scores. If something fails, fix it before showing the draft, and do not narrate the process.

Step 9: close the loop (mandatory, all three)

1. SOURCES.md: a new row (source, date, area, gate exit, where the distillate lives).
2. Your practices log: what was reviewed and when.
3. METHOD.md: if the user corrected style or criteria, write it down. A correction that appears twice becomes a permanent rule.

What this skill learns (the differential)

The summarizing criteria calibrate with use. Record in METHOD.md: what deserves a verbatim quote versus paraphrase; when [VERIFY] is wanted and when it feels excessive; density and format (how many tables, how much prose); which kind of counterpoint the user values (the author's interest? contrary evidence? the counterexample?); which sections they reread. That last one is the most honest signal and the one nobody uses: if they never reread §2 but always return to §5, invert the weight.

Hard rules

1. Do not invent. Every datum with a source or [VERIFY].
2. Do not summarize neutrally. Every item answers "what changes for me?".
3. The author's POV goes separate and first.
4. Your value is in what you discard. When in doubt, it stays out.
5. Do not version third-party content. Transcripts and foreign material live in a gitignored folder; only pointers and your own distillates get committed.
6. Public means sanitized. Areas, goals, sources and distillates are private; only this template is public.
7. The gate is not skipped on request.

Done when the five sections exist with real content, eval.md passes, and the three loop files are written.