/exec-comms

Format discipline. Takes a content-correct draft, restructures it for exec readability. Does not rewrite content — that's /clevel-review's job upstream.

Canonical spec: <workspace>/6 - slash commands/exec-comms.md

When to invoke

• After /clevel-review Phase 3 (auto-chain)
• Standalone: /exec-comms <file path> against any markdown bound for a senior audience
• Auto-fire when audience is named in frontmatter or visible in draft (GM, CFO, Sector Exec, ExCo, Board, external client exec)
• Skip: internal wiki edits, status updates, team docs without a named senior audience

Inputs (ask via AskUserQuestion if not provided)

• File path of the draft (or paste the draft inline)
• Audience tier — GM / CFO / ExCo / Board / External (drives length cap and template)
• Read time target if different from audience default

Format rules (enforce, do not negotiate)

Visual structure

Every page opens with a single callout block:

> **Audience:** [named person + role] | **Read time:** [N] min | **Core message:** [one sentence]

Every major section follows Lead → Evidence → Implication:

• Lead: 1-2 sentences with the punchline in bold
• Evidence: table or tight bullets (≤3 sentences each)
• Implication: one bold sentence stating "so what"

Asks / recommendations / risks always in a parallel table with consistent columns.

Visual rules

• ≤3 sentences per paragraph
• ≤2 levels of bullet nesting
• Bold the punchline of every paragraph (scanner test: reading only the bolds tells the story)
• Section headings sentence-case, parallel structure (all noun-phrase OR all imperative, not mixed)
• Tables for any comparison of 3+ items (no prose lists)
• Use > blockquote for the so-what callout, never decorative

Banned words (scan + replace)

Banned AI tics (cut on sight)

• em-dashes (—)
• "It's worth noting"
• "Importantly,"
• "Furthermore,"
• "In summary,"
• "Ultimately,"
• "That said,"
• "Notably,"
• "Indeed,"

Banned structural patterns (refuse to ship)

• "What this is NOT" sections (defensive)
• "Pre-empted Questions" / "FAQs" sections (if asked, answer; don't pre-litigate)
• Asks of different shapes in same list (parallel structure mandatory)
• Value tables before the ask (lead with the answer, not the proof)
• Headlines without urgency markers (anchor every ask to a date)

Length caps (hard)

If draft exceeds cap, the skill must cut, not warn.

Process

1. Read the input draft (file or paste)
2. Classify audience tier (from frontmatter, draft header, or AskUserQuestion)
3. Scan for violations:
   • Banned words / AI tics (regex)
   • Em-dashes (regex —)
   • Paragraph length >3 sentences
   • Bullet nesting >2 levels
   • Section structure (does each section have Lead → Evidence → Implication?)
   • Defensive sections present
   • Parallel structure in lists
   • Word count vs audience cap
   • Bold-punchline coverage (every paragraph has a bold span)
4. Apply changes in this order:
   1. Strip banned words / AI tics / em-dashes
   2. Cut defensive sections entirely
   3. Restructure each section to Lead → Evidence → Implication
   4. Convert prose lists to tables where ≥3 items
   5. Add bold to each paragraph's punchline
   6. Cut to length cap (drop weakest content first, then compress)
   7. Replace the page header with the standard callout block
5. Run scanner test: extract every bold span, concatenate, read. If the bolds-only version doesn't tell the story, return to step 4 and rebold.
6. Verification report: word count before/after, violations fixed, scanner test PASS / FAIL.

Output

The skill writes the reformatted markdown back to the same file path (or to a .formatted.md sidecar if --preserve flag set).

Then prints:

## /exec-comms verification
- Audience tier: [tier]
- Word count: [before] → [after] (cap: [cap])
- Read time: [N] min (target: [target])
- Banned words removed: [N]
- Em-dashes removed: [N]
- AI tics removed: [N]
- Defensive sections cut: [list]
- Tables created from prose lists: [N]
- Bold punchlines added: [N]
- Scanner test (bolds-only story coherent): PASS / FAIL
- Sections restructured to Lead-Evidence-Implication: [N]

Failure modes

• Audience unknown: ASK via AskUserQuestion. Don't guess.
• Draft is too short for cap (e.g. 100 words for GM): warn, do not pad. Format-only means no content adds.
• Scanner test fails after 2 reformat passes: surface to user with the bolds-only string and ask which paragraphs need rebold judgment.
• Banned word is a quoted source (e.g. "Auditor-General said robust"): preserve quoted material verbatim. Only replace banned words in original prose.
• Length cap conflicts with content the user wants kept: surface as decision point, do not auto-cut without confirmation.

Reference

• Spec: <workspace>/6 - slash commands/exec-comms.md
• Upstream: /clevel-review (auto-chains to this skill at Phase 3)
• Related: program-exec-summary (Word output), utility-anthropics-brand-voice (brand-level voice rules)