Review

/scribetronic:review

Multi-focus review of a draft. The orchestrator dispatches one subagent per focus in parallel (single message, multiple Task calls), each with a narrow brief and a copy of the user's writing-style override. Each subagent returns findings as a structured table; the orchestrator merges them into a single severity-grouped report.

This is the parallel counterpart to /scribetronic:editing-pass + /scribetronic:ai-slop-check. Two modes:

• Full mode (default) — review the entire draft against 5 focuses. ~10s wall time. Use as a pre-publish gate or when you want a fresh full read.
• Delta mode (--since-last / --since <revspec> / --since-staged) — review only the lines that changed since the last snapshot or git ref. 3 focuses (voice, slop, continuity), ~5s wall time, ~80% fewer tokens. Use when you've just written a paragraph or two and want fast feedback before continuing.

When to pick:

• Mid-draft, just wrote 2 paragraphs → delta mode (--since-last)
• Mid-draft, "let me check the whole thing" → full mode
• Pre-publish gate → full mode (sometimes with --with-evidence)
• On a worked snippet from a long-form → full mode

Voice resolution

Before dispatching, load the voice base in this order:

1. scribetronic/style/writing-style.md in the project root, if it exists — user's personalized override.
2. Bundled writing-style/SKILL.md template otherwise — generic baseline.

Pass the resolved voice content to every subagent so they all check against the same rules.

Inputs

--since-last, --since, --since-staged are mutually exclusive — passing more than one is a usage error.

If <draft-path> doesn't exist, refuse with a hint to check the path.

Focuses

Each focus is one Task subagent invocation. The subagent reads the draft + the resolved voice base + the focus-specific brief.

voice

Brief: "Read the writing-style rules below. Then read the draft. Report every line where the draft violates a writing-style rule. Cite the rule from the style guide and quote the offending line. Severity: HIGH if the violation breaks a 'never' rule (e.g. 'not just X, but Y'), MEDIUM if it's a tic-overuse, LOW if it's borderline. Skip everything that doesn't violate a rule."

Output shape:

HIGH  L23  "...not just a number, but a vision"  ← rule: "not just X, but Y" banned
MEDIUM L41  third rhetorical question in 5 paragraphs  ← rule: max 1 per piece
LOW   L7   adverb stack ("incredibly powerful")

structure

Brief: "Read the draft. Evaluate it on these structural axes: (1) thesis in first 2 sentences? (2) descriptive vs clever H2s? (3) section length variance? (4) bullets used for enumeration only or padding? (5) closer declarative or trailing? Report problems only."

Output shape:

HIGH  L1-2   thesis missing — opener is "I want to share what happened this week" (no point until L8)
MEDIUM ## Why  H2 is "Meet the new kid on the block" — clever, not descriptive
LOW   §3-§4   two consecutive 200-word sections — vary length

slop

Brief: "Run the ai-slop-check detection list (HIGH/MEDIUM/LOW patterns) over the draft. Report every hit. Be conservative on LOW — only flag if it's definitely lazy."

This subagent essentially re-uses /scribetronic:ai-slop-check's detection table. Same output shape as that skill.

hook

Brief: "Read only the first 3 sentences of the draft. Apply the 5 hook patterns from the writing-style guide (number reveal, contrarian assertion, unexpected admission, list promise, specific scene). Verdict: does the hook earn the scroll-stop? If not, propose one rewrite that fits the post's actual subject."

Output shape:

HIGH  L1-3  hook is neutral framing ("This week was about shipping")
            → propose: "I shipped 4 things this week. Three flopped. Here's why I'm doing more of one of them."

closer

Brief: "Read only the final 2 sentences of the draft. Apply the 4 closer patterns (repeat-thesis-as-command, specific-next-action, one-liner reframe, sign-off with date/CTA). Verdict: does the closer commit to something? If it trails off, propose one rewrite."

Output shape:

HIGH  L156-157  soft closer ("Onwards. Let's see what next week brings.")
                → propose: "Next week: ship the publishing pipeline. Newsletter on Sunday. Same time."

continuity (delta mode only)

Brief: "Read the BEFORE block (last 1-2 paragraphs of the draft as it was) and the AFTER block (the same 1-2 paragraphs plus what was just added). Evaluate ONLY whether the AFTER continues the BEFORE coherently. Check: tone consistency (no genre shift), argument flow (does the new content build on or pivot from what was there?), no orphan references (does the AFTER name a noun the BEFORE didn't introduce?), tense/perspective consistency. Report breaks only."

Output shape:

HIGH  L78-82  AFTER shifts from past tense to present, mid-paragraph (BEFORE was past throughout)
MEDIUM L84   "the framework" — first mention; BEFORE never introduced one
LOW   L80   tone tightens (BEFORE conversational, AFTER didactic) — intentional?

This focus is only run in delta mode. In full mode it doesn't fire (the whole draft is the content; there's no "before" to compare against).

factual (opt-in via --with-evidence)

Brief: "Compare every numeric claim, named person, named tool, dated event in the draft against the evidence file . Report contradictions only. Never invent corrections — only flag the discrepancy and quote both."

Skip silently if no evidence file is found.

Output shape:

HIGH  L34  draft says "8 free signups, zero paid" — evidence (notes.md L12) says "12 free signups, 1 paid"
MEDIUM L51 draft says "MakerOps launched on August 16" — evidence has no date for that launch

Workflow

Phase 1 — Resolve and validate

1. Resolve <draft-path>. If missing → refuse with error: draft not found at <path>.
2. Resolve voice base (scribetronic/style/writing-style.md → bundled template).
3. Determine mode:
   • If --since-last, --since <revspec>, or --since-staged is present → delta mode.
   • Otherwise → full mode.
   • More than one delta flag → usage error.
4. Parse --focus list (default depends on mode):
   • Full default: voice,structure,slop,hook,closer
   • Delta default: voice,slop,continuity
   • In delta mode, silently drop structure, hook, closer from --focus if present (they don't apply to deltas) and emit a note:.
5. If factual is in the list, locate the evidence source. If absent, drop factual silently and warn (note: skipping factual focus — no scribetronic/calendar/<week>/notes.md found).
6. Delta mode only — compute the delta:
   • --since-last: read .scribetronic/snapshots/<basename>.md if it exists; if not, snapshot the current draft, emit note: first --since-last on this file — snapshotted, falling back to full mode this run, then run as full mode.
   • --since <revspec>: git diff <revspec> -- <draft-path> (must be in a git repo; error otherwise).
   • --since-staged: git diff --staged -- <draft-path>.
   • The delta is the unified diff. Extract the AFTER hunks (lines prefixed +, excluding the +++ header) and locate them in the current file with line numbers.
   • If the delta is empty (no changes since the reference), emit note: nothing changed since <reference> — nothing to review and exit 0 successfully.
   • Build the BEFORE/AFTER context per added hunk: 1-2 paragraphs of unchanged content immediately above and below, used as continuity anchor.
7. Delta mode only — at the end, update the snapshot (only for --since-last): write the current draft to .scribetronic/snapshots/<basename>.md so the next --since-last invocation diffs against this run.

Phase 2 — Dispatch in parallel

Spawn one Task subagent per focus, all in a single message (parallel tool calls).

Full mode prompt skeleton:

You are running the "<focus>" review pass for scribetronic.

## Draft
<paste full draft content>

## Writing-style rules (your reference)
<paste resolved voice base>

## Your brief
<focus-specific brief from the table above>

## Output format
Return a markdown table with columns: Severity | Location | Finding.
Severity: HIGH | MEDIUM | LOW. Location: line number(s) or section heading.
Finding: 1 sentence description, optionally followed by "→ propose:" rewrite.
Return ONLY the table, no preamble, no closer. Empty table if nothing to flag.

Delta mode prompt skeleton (per-focus):

You are running the "<focus>" review pass for a delta in a scribetronic draft.

## What changed (AFTER block — review THIS only)
<lines added in the diff, with their line numbers in the current file>

## What was there before (BEFORE block — context, do NOT flag issues here)
<1-2 paragraphs of unchanged content immediately above the AFTER, plus 1 paragraph below if any>

## Writing-style rules (your reference)
<paste resolved voice base>

## Your brief
<focus-specific brief — for continuity, the brief explicitly compares BEFORE to AFTER>

## Output format
Same as full mode. ONLY flag findings located in the AFTER block.
Empty table if the delta is clean.

The BEFORE block exists to give the subagent context (so it doesn't flag a pronoun whose antecedent lives in BEFORE) but is NOT itself reviewed.

Subagent type: general-purpose for all focuses. Model: haiku is fine for voice/slop/hook/closer (pattern-matching tasks); use sonnet if factual is enabled (cross-referencing requires more reasoning).

Phase 3 — Aggregate

Collect each subagent's table. Merge into one table grouped by severity, with a Focus column added:

HIGH
| Focus     | Location | Finding                                              |
|-----------|----------|------------------------------------------------------|
| voice     | L23      | "not just X, but Y" construction (rule violation)    |
| closer    | L156-7   | soft closer — propose: "Next week: ship..."          |
| factual   | L34      | claim contradicts notes.md (8 vs 12 signups)         |

MEDIUM
| Focus     | Location | Finding                                              |
|-----------|----------|------------------------------------------------------|
| voice     | L41      | 3rd rhetorical question — rule: max 1 per piece      |
| structure | ## Why   | clever H2 ("Meet the new kid on the block")          |

LOW
| Focus | Location | Finding                                                  |
|-------|----------|----------------------------------------------------------|
| voice | L7       | adverb stack ("incredibly powerful")                     |
| slop  | L98      | "Pro tip:" call-out — consider cutting                   |

Phase 4 — Verdict

Below the tables, print a one-line verdict:

verdict: SHIP — 0 HIGH, ≤2 MEDIUM
verdict: REVISE — 1+ HIGH or 3+ MEDIUM

The thresholds match /scribetronic:ai-slop-check's "If 1+ HIGH or 3+ MEDIUM, revise and rerun" rule.

If --quiet is set, only print the verdict and the HIGH table (skip MED/LOW).

Output discipline

• The aggregated report goes to stdout (it's the data).
• Per-focus progress lines (e.g. dispatching voice... (haiku), voice: 3 findings) go to stderr unless --quiet.
• Errors (missing draft, missing evidence) → stderr.

This lets an agent or script do result=$(/scribetronic:review draft.md); echo "$result" and get clean output, while a human running interactively still sees progress.

Anti-patterns

• Don't run focuses sequentially. The whole point is parallelism. If you find yourself awaiting one subagent before dispatching the next, re-read this skill.
• Don't let subagents propose rewrites for HIGH violations they didn't find. Each subagent owns its focus — voice doesn't propose hook rewrites.
• Don't aggregate verbatim findings if multiple focuses flag the same line. Dedupe by (line, finding-essence) and keep the most specific source.
• Don't run factual without an evidence file. Without grounding, the subagent will hallucinate corrections. Drop it silently.
• Don't run full mode on drafts shorter than ~100 words. The overhead of dispatching 5 subagents isn't worth it. Below that, just use /scribetronic:editing-pass.
• Don't run continuity outside delta mode. It needs a BEFORE/AFTER pair. If a user tries --focus continuity in full mode, drop it silently with a note:.
• Don't update the --since-last snapshot if dispatch failed. The snapshot only advances on a successful run, so a failed review doesn't lose its reference point.
• Don't snapshot in any mode other than --since-last. Other delta modes use git refs and don't need a snapshot file.

When to use which

Integration with /scribetronic:write

The write orchestrator currently runs editing-pass then ai-slop-check sequentially in Phase 4. From v0.3.0 onward, when review is available, write calls review instead:

Phase 4 — Quality gate
  /scribetronic:review <draft-path>
  if verdict == REVISE:
    show user the HIGH table
    ask: revise now / mark draft anyway / cancel

This is documented in write/SKILL.md and is opt-out via --no-review if a user wants the old sequential flow.

Limitations

• Requires Claude Code with the Task tool. Won't run in IDEs without parallel subagent support.
• 5 subagents = 5x token spend per review. Default focuses are pattern-matching (haiku-friendly), so cost is low (~$0.01/review at current haiku pricing), but still 5x of one editing pass.
• No streaming. Findings arrive together when the slowest subagent finishes (typically ~8-12s). Can't show partial results mid-review.