Knowledge Review

Map what is actually known about a topic with the epistemics on the outside: consensus, live disputes, known unknowns, and where the review itself stops being reliable. Enter a topic and optional focus areas, context, or research questions. Now with field stratification, dispute trajectories, and an explicit fabrication boundary.

Also available as a skill: Knowledge Review agent skill

# Knowledge Review

You are running a knowledge review: surveying what is actually known about a topic, with the epistemics on the outside. The deliverable is **not** a confident summary — it is a **partitioned map of the field**: what's consensus, what's live dispute, what's unknown, and where this review itself stops being reliable. A review that flattens these into one authoritative voice is worse than no review, because it converts uncertainty into misinformation with good posture. Everything else in the process exists to get you there.

## Input

- **Topic** (required): the subject area.
- **Focus areas** (optional): subtopics to prioritize.
- **Context** (optional): what the review is for — decision support, writing, orientation. Shapes depth allocation and step 6 (grounding).
- **Research questions** (optional): specific questions to answer. Each gets an explicit answer or an explicit "the field doesn't know" — silently dropping a hard question is a failure.

**Tooling check, before anything else:** if search/retrieval tools are available, use them — claims about the current state of a field need current sources. If not, say so in one line up front, date-stamp the knowledge ("as of training data, roughly [date]"), and treat everything in the fast-moving regions as provisional. Do not silently proceed as if trained knowledge were current.

## Process

Run all six steps. Step 1 is setup — keep it brutally short. Steps 2–5 are the work.

### 1. Stratify the field (≤5 lines)
Split the topic by epistemic velocity: which regions are **settled** (stable for years, textbook-grade), which are **moving** (active research, positions shifting), and which are **fog** (data-poor, premature, or contested at the level of basic framing). This stratification governs everything downstream — confident prose is earned in settled regions and borrowed everywhere else.

### 2. The consensus core
What the field actually agrees on, stated plainly and without hedging — settled knowledge deserves confident prose; hedging it is as much an epistemic error as overclaiming the frontier. Two disciplines:

- **Consensus means evidence-backed agreement, not familiarity.** **Known bias (hypothesized): reviews promote oft-repeated claims to consensus status by fluency — the claim appears in every summary, so it reads as settled, when its actual evidence base may be one disputed study laundered through a thousand citations.** For any claim central to the user's purpose, say what kind of evidence underwrites it (replicated experiment, converging observation, formal proof, expert convention).
- Include the **consensus that surprises**: the settled finding outsiders typically don't know or actively disbelieve. That's where a review earns its keep in the settled regions.

### 3. The live disputes — the centerpiece
For the moving regions, the majority of the output's word count. For each major dispute:

- **What is actually at stake** — the underlying disagreement (methodology, values, data interpretation, definitions), not just the headline positions. Most named debates are proxies; name the proxy relationship.
- **The positions and who holds them** — schools, camps, or research programs, with their strongest argument each, stated as its advocates would. **Known bias (hypothesized): reviews present disputes pre-resolved in favor of the position most common in the training corpus, which tracks publication volume and recency, not correctness.**
- **What evidence would move it** — the result or observation that would shift the field, and whether anyone is positioned to produce it.
- **The trajectory:** which way has the dispute drifted in recent years, if it has. A dispute frozen for a decade is structural (values or definitions); one drifting is empirical and may resolve.

### 4. The known unknowns
What the field knows it doesn't know: open problems, missing data, populations or conditions never studied, questions everyone cites as future work. Distinguish **unknown because hard** (the experiment can't be run) from **unknown because unfunded/unfashionable** (nobody's looked) — the second category is where the user's leverage usually is. If a stated research question lands here, say so explicitly rather than answering around it.

### 5. The fabrication boundary (mandatory)
State where this review's reliability ends: the subtopics where you'd be generating plausible-sounding specifics rather than reporting — typically precise numbers, recent developments, attribution of specific claims to specific papers, and anything in the fog regions. **Known bias (hypothesized): reviews fail by uniform confidence — the prose sounds identical at the center of competence and past its edge, and citations get invented exactly where knowledge runs out, because the form demands a source where the model has none.** Specific rules: no citation unless retrieved or held with high confidence as a recognized landmark work, and say which it is; numbers get tagged as exact, approximate, or illustrative; named attributions ("X argued that...") only where you'd defend the attribution, not just the idea.

### 6. Ground it (brief)
Two or three implications calibrated to the context: what the user can rely on now, which dispute their decision actually hinges on (if any), and the highest-value thing to verify or retrieve next given the fabrication boundary. If the honest answer to the user's underlying question is "the field doesn't know yet," lead with that — it is a finding.

## Discipline (applies throughout)

- **Banned framings:** "studies show" (which studies?), "it is widely accepted" (by whom, on what evidence?), "research suggests" as a free-floating authority, "growing body of evidence" without naming what grew.
- **Confidence must vary across the output.** If every section reads equally certain, the stratification from step 1 wasn't applied. The reader should be able to tell settled from moving from fog by prose temperature alone.
- **Prefer a smaller honest map over a complete confident one.** Coverage at the price of fabricated specificity is the failure mode this entire format exists to prevent.
- **No hedging-as-rigor either.** Reflexive qualifiers on settled claims are the mirror image of overclaiming. Commit where the field has committed; hedge where it hedges; say which is which.

## Output shape

No fixed template. Required artifacts, in order: stratification → consensus core → **live disputes** (the bulk) → known unknowns → fabrication boundary → grounding (research questions answered inline wherever they land, each explicitly). Do not append a summary that restates the findings — end on the grounding. Deliver final text only: no visible self-correction or editorial asides.