rl-review-panel

Files (1)

• SKILL.md

SKILL.md (first lines)

# Review Panel

A panel of blind readers for a finished content draft. Each lens asks one question of the whole
piece, from one side. None of them writes a fix. The writer — or the author-voice layer — writes
the fix, so revisions stay in the writer's register instead of drifting toward the panel's.

One rule governs every finding: **it must make the piece a better version of what it already is,
not turn it into something else.** (Robert Gottlieb's test, from his criticism of editors who
rewrote writers toward their own aesthetic.) The panel serves many writers with different voices;
it critiques without homogenizing.

The rules the lenses enforce live in `rl-writing-craft`. A lens names the rule it is applying; it
never restates it. If you find yourself writing a rule into a finding, stop and cite it instead.

---

## Scope: content only

**Reviews:** content meant to be read by an audience — articles, blog posts, essays, newsletters,
emails, case studies, landing and web copy, sales and pitch decks, scripts, press releases,
social posts, announcements.

**Does not review:** strategy documents; brand foundations (brand story, positioning, messaging
frameworks, audience definitions, content playbooks); voice profiles and `AUTHOR-CONTEXT.md`;
briefs, outlines, and plans. Those are inputs the writer decides, not drafts an audience reads,
and the lenses below would judge them against the wrong standard. If asked to panel one, say so
and stop.

---

## When it runs

- **In `rl-content-pipeline`, step 6** — the panel pass, alongside the judge, adversarial, and
  fact-check passes. See that skill for how the passes fit together.
- **Standalone, via `/panel`** — on any content draft.

**Never from the context that drafted the piece.** Self-review by the drafting context is the
weakest possible judge (`rl-content-pipeline`, step 6). Every lens runs in a fresh context.

**Inputs:** the draft; the brief if one exists; `AUTHOR-CONTEXT.md`; the author-voice skill or
`VOICE-PROFILE.md` if present. Nothing about how the draft was produced. The brief goes to the
Keeper only; the lenses get the Keeper's statement of it, so they judge the piece on the page
rather than the plan behind it.

---

## The Keeper (chair, non-voting)

The Keeper owns voice and intent. It does not judge prose quality and never blocks a HARD STOP on
voice grounds. It runs in a fresh context too — never the drafting one. Under `--light`, the one
fresh agent that runs the lenses also plays the Keeper.

**Opens — once per piece, before any lens reads.** A short statement every lens receives:
- **What the piece is:** form, audience (from `AUTHOR-CONTEXT.md` or the brief), purpose.
- **Voice markers:** the deliberate features of this writer's voice, from the voice profile —
  signature phrases, fragment habits, sentence shape, register, profanity, anything the profile
  marks as chosen. Mark aspirational items as such; they are never grounds for a finding.
- **Narrative or not:** whether the Storyteller is seated.

In the pipeline, write it from the brief (step 3) and the voice profile; don't re-derive it.

**Closes — after every lens reports.** Consolidate (see Output), then run the voice check after
revision (see Deferring to Voice).

---

## The Lenses

Each lens is defined by its question, the signature questions it reads with, the rules it cites,
and what it must not do. The signature questions come first: they are how a good editor reads.
The rules are how the finding gets checked.

### Architect — structure, stakes, and promises
**Question:** does every part earn its place, in this order, for this reader?
**Reads with:** What is at stake, and for whom? Where does the reader first learn the point — and
is that where it should be? If this section went, what would the reader lose? Does the title
promise what the piece delivers? Is the best material buried late?
**Cites:** `structure.md` → Stakes · Titles · Opening · Flow and Section Logic · The Close ·
Format Defaults.
**Must not:** impose one template on every genre (answer-first suits a memo, not an essay or a
story); line-edit; change the piece's form or intent.

### Checker — claims and evidence
**Question:** does the evidence support each claim at the strength it's stated?
**Reads with:** Who says so, and what exactly did they say? How could the writer know this? Is
this claim stronger than its source? What's the denominator behind "most"?
**Cites:** `edit.md` → Grounding Rules · `audit.md` → vague attribution · `SKILL.md` → Logical
Consistency (personal vs. sourced knowledge; causal chains that swap their terms).
**Note:** the pipeline's fact-check pass decides whether a claim is *true*. The Checker judges
whether it is *stated* at the strength the evidence carries.
**Must not:** demand citations in genres that don't carry them — it asks for verifiability, not
footnotes; treat labeled opinion as a factual claim.

### Stranger — the reader's side of the page
**Question:** can a smart newcomer follow this, trust it, and not feel talked down to?
**Reads with:** Who is this, and why should I know them? What step did the writer skip because it
felt obvious? Where did I have to reread? Is this one thing or several things with different
names? Where did the writer explain what I already know?
**Cites:** `edit.md` → Reader Access · Orient Before You Move · Audience Calibration ·
`SKILL.md` → Logical Consistency (antecedents, definite references).
**Must not:** flatten specialist writing for a specialist audience — read `AUTHOR-CONTEXT.md`
first; simulate a specific customer persona's reaction (that's an audience-persona review, not
this one).

### Conscience — honest persuasion
**Question:** does this persuade by honest means?
**Reads with:** Would this framing survive the reader learning the whole truth? What happens if
the reader waits a month? Is emotion showing the stakes or replacing the argument? Would the
writer be comfortable if the reader saw exactly how this was built?
**Cites:** `edit.md` → Persuasion Integrity · Manufactured vulnerability · Exclamation points ·
`audit.md` → significance inflation.
**Must not:** ban emotion, story, or strong opinion; moralize at the writer; judge factual truth
(the Checker's job).

### Storyteller — seated for narrative pieces only
Case studies, customer and brand stories, founder narratives, memoir, fiction.
**Question:** does the story work as a story?
**Reads with:** Who wants what, and what's in the way? Where would the reader want to be in the
room? Whose eyes are we looking through? Does the ending pay off what the opening set up? Who

From reventurelabs/rl-content-os, MIT.

More from reventurelabs/rl-content-os