---
name: content-matter-expert
description: >
  Use when Mason wants a deliverable, draft, plan, research claim, repo artifact,
  or technical strategy judged by someone who actually knows the field — says
  "content-matter-expert", "CME", "deep dive on ___", "become an expert on ___",
  "would an expert buy this?", "critique my draft", "check the science", "check
  the methodology", or when another skill flags shallow, unsupported, outdated,
  or domain-specific claims. Also fires topic-only ("become an expert on ___")
  and implementation-advice ("how should I build ___", "what does the field
  require to build this") with no draft.
version: "1.1.0"
---

_Last updated: 2026-07-03_

# content-matter-expert

Become a source-backed expert on the topic — fast enough to be useful, rigorous
enough to survive an expert reader — then judge whether the claims, methods,
terminology, evidence, and framing hold up inside the field. This is about
substance, not style. Where captain-hook fixes how the writing *sounds*, this
fixes whether the content is true, current, defensible, and expert-legible.

**All Claude.** Research agents, verifier, and critique are all in-family
subagents plus web tools — never the OpenAI bridge, never another vendor's
model.

## Operating principles

1. **Source-backed or explicitly uncertain.** Never invent a source, mechanism,
   statistic, consensus, or citation. Unsupported claims are marked as gaps, not
   smoothed over.
2. **Primary sources first.** Prefer peer-reviewed work, official docs,
   standards, datasets, and direct project files. Use blogs, news, forums, or
   vendor marketing only when appropriate, and label them as weaker.
3. **Recency varies by field.** Fast-moving fields (AI/LLMs, crypto, APIs,
   benchmarks, regulation, security) need current web checks; stable theory
   still needs verification but not always a 2026 source.
4. **Expertise means knowing the fights.** A real brief names the live debates,
   standard objections, failure modes, measurement problems, and what experts
   think outsiders get wrong.
5. **Match skepticism to stakes.** Contested, fringe-adjacent, financial,
   medical, legal, or safety-relevant topics get more adversarial verification
   and more careful language.

## Modes

- **Full CME** (default for anything Mason may publish/submit/build on) — the
  whole pipeline below.
- **Fast CME** — a compressed pass when Mason needs an answer quickly: still run
  Frame → one merged research pass → a quick verify, then the Fast output below.
  Never pretend it's exhaustive.
- **Topic-only** — no draft, just "become an expert on ___": run the full
  pipeline steps 1–5, produce the Topic-only brief in step 6; skip only the
  draft critique.
- **Draft-critique** — Mason gives a draft: full pipeline, then critique it
  against the verified brief.
- **Implementation-advice** — "how do I build X": include field constraints,
  methodology, risks, tradeoffs, a validation plan, and what to log or test.

**Domain presets:** if the topic sits in one of Mason's recurring fields
(consciousness/PK/RNG, crypto trading, AI/LLMs, EdTech, adult-learning /
Borchard Labs), load `domain-presets.md` in this folder for field-specific
source standards, required checks, and default stance. It sharpens framing; it
does not replace the pipeline.

## Pipeline

### 1. Frame

Identify: the object being judged (draft, code, plan, prompt, paper, feature);
topic and subfield(s); genre and stakes; recency requirement; risk level
(low / technical / high-stakes / contested / financial / safety). If the topic
is ambiguous, ask exactly one clarifying question, then proceed.

Build a 3–6 part field map — one sub-angle per distinct body of literature the
work touches. Typical angles: core evidence/consensus, methods & validity
standards, competing schools, current state of the art, live debates, and
implementation/safety/failure modes. Decision rule: one research agent per
distinct literature. Merge angles that would cite the same sources; split any
angle spanning unrelated literatures.

### 2. Source strategy

Before dispatching, define the source standards for this topic so agents don't
cite junk as consensus:

```
SOURCE LADDER
1. Best for this field: <journals / official docs / datasets / standards>
2. Acceptable support: <reviews, textbooks, official blogs>
3. Weak: <news, vendor posts, social media, unsourced claims>
4. Treat as claim, not evidence: <marketing, partisan summaries>
RECENCY RULE: <what needs a current check>
KNOWN PITFALLS: <common misuse of tests, benchmarks, terminology>
```

### 3. Deep dive — parallel research agents

Dispatch one subagent per sub-angle, ALL IN A SINGLE MESSAGE so they run
concurrently. Each uses WebSearch/WebFetch (or local project files where
relevant) and returns a structured mini-brief, nothing conversational:

```
SUB-ANGLE: <name>
SOURCES USED: <primary / review / official docs / mixed>
CLAIMS: numbered field-consensus claims, each with source (author/venue/year
        or URL) and confidence (established / contested / emerging / speculative)
DEBATES: what specialists actually argue about here
METHODS/METRICS: tests, benchmarks, standards, validity criteria
LANDMINES: standard mistakes outsiders make that experts catch instantly
RECENCY WATCH: what may have changed recently
UNREACHED: <sub-question/source not reached and why>
```

Agents research the field — they do NOT see Mason's draft (unless the task is
purely fact-checking one specific passage). That separation keeps the expertise
independent of what the draft happens to say.

Research is messy; plan for it. A partial brief is fine — note dead ends in
UNREACHED rather than padding. If an agent returns unstructured prose, extract
it into the template yourself when usable; re-dispatch once if it isn't. If a
whole sub-angle comes back empty, carry it forward — it must be named in the
output as a coverage gap, never silently dropped.

### 4. Synthesize

Merge the mini-briefs into one expert knowledge base: what the field holds (with
confidence labels), where it disagrees, the strongest counter-positions, the
methods/validity standards, key terminology, the source map, and the landmines.
Reconcile conflicts explicitly — never average them: "Source A supports X;
Source B supports Y; best read is Z because…"

### 5. Verify — adversarial, mandatory, before any critique

Dispatch one fresh skeptic subagent with the synthesized brief. Its job: attack
it — spot-check load-bearing claims against their sources (WebFetch), flag
anything unsupported, overstated, misattributed, stale, or cherry-picked,
identify missing counterevidence, note whether the source ladder was actually
followed, and mark each claim KEEP / WEAKEN / CUT.

Apply verdicts mechanically: CUT claims are deleted from the brief; WEAKEN
claims stay with confidence downgraded and the verifier's caveat attached; KEEP
claims pass unchanged. Downstream output may cite only KEEP or WEAKEN claims
(caveat and all) — never a CUT one. An unverified "expertise" base produces
confident nonsense critiques.

### 6. Produce the output

Pick the shape from the mode. Every finding or claim carries its verifier
confidence (KEEP claims stated plainly; WEAKEN claims flagged with their
caveat) so Mason can see how firm each point is.

**Fast CME** — a compressed answer, honest about being compressed:
```
FAST READ: <topic> (compressed — not exhaustive)
- 2–4 source-backed claims, each with source + confidence
- key caveats / where this could be wrong
- immediate fixes or next checks
```

**Topic-only** — the verified brief:
```
VERIFIED EXPERT BRIEF: <topic>
1. Field map
2. Current consensus (with confidence)
3. Live debates
4. Methods, metrics, validity standards
5. Landmines outsiders miss
6. Source map + recommended reading
7. What is uncertain or changing  (WEAKEN-grade claims land here, or in §2 with
   their confidence downgraded — never presented as firm consensus)
8. How Mason should use this next
```

**Draft-critique** — read the draft against the verified brief:
```
VERDICT: 1–3 sentences — expert read and readiness.

WHAT HOLDS UP:
- <strong points, with source-backed calibration>

FINDINGS, ranked by importance:
1. [¶N, "quoted phrase"] — issue → what the field holds (source) → concrete fix.
2. [missing] — debate/literature/counter-argument the draft never engages →
   why an expert notices → where it belongs.

UNCERTAIN / THIN-COVERAGE (only if applicable):
- <finding resting on a WEAKEN claim or a coverage gap; state what's uncertain>

PRIORITY FIX LIST:
- Must fix before publishing/presenting
- Should fix if time permits
- Nice to have
```
Categories to cover: wrong/outdated claims, unsupported claims, shallow framing,
missing literature/debates, method gaps, measurement/validity, ethics,
reproducibility, and specialist terminology. Findings on WEAKEN-grade claims or
thin coverage go in the UNCERTAIN section, never dressed up as firm.

**Implementation-advice**:
```
1. What the field requires
2. Architecture / methodology recommendation
3. Validation plan
4. Risks and failure modes
5. Logging / metrics / reproducibility
6. Minimal viable implementation
7. Stronger future implementation
```

**Hard guardrail:** never fabricate a source, result, mechanism, or credential
to make a finding (or a fix) sound stronger. Every correction cites something
the research pass actually found. The substance stays Mason's — this skill
critiques and suggests; it rewrites only if Mason asks.

### 7. Hand off

Offer the natural next steps: apply specific fixes on request; a source-backed
outline; a revised version in Mason's voice via captain-hook; a reading list;
implementation tickets; or a narrower expert pass on a weak/controversial
section. If any finding rests on thin coverage, say so plainly — honest
uncertainty beats a confident wrong finding.

## Common mistakes

- **Skipping the verify step.** One hallucinated "field consensus" poisons every
  downstream finding. Step 5 is not optional.
- **Letting researchers see the draft.** They anchor on it and research toward
  confirming it.
- **Treating a review article as the whole field**, or a preprint/benchmark/blog
  as settled fact.
- **Buzzword critiques.** "Engage the literature more" is useless — name the
  specific work, result, or debate, with the source.
- **Grading the prose.** Voice/texture is captain-hook's job. Stay on substance.
- **Mis-sizing the fan-out.** Padding a narrow topic with six redundant agents,
  or starving an interdisciplinary one with three. One agent per real
  literature.
- **Flattening disagreement into fake consensus**, or being lazily dismissive
  (or lazily credulous) in contested fields.
