corrector
Transform raw text, voice transcriptions, or rough drafts into polished, ready-to-send messages across any channel (email, WhatsApp, SMS, Slack, Teams, LinkedIn). Generates 2-4 strategic variants with
SKILL.mdskills-web/corrector/SKILL.md
GitHubCorrector v2
Overview
Universal message reformulator that transforms any input (voice transcription, rough draft, scattered notes) into ready-to-send messages with multiple strategic variants. Supports all communication channels: email, chat (WhatsApp/SMS), professional messaging (Slack/Teams), and social networks (LinkedIn).
Core differentiator: Variants differ by communication strategy (what each prioritizes), not just formality level.
Quick Decision
INPUT RECEIVED
│
├─ Contains heavy dictation artifacts (fillers, false starts)?
│ → Phase 0: Clean first, show cleaned version
│
├─ Channel detectable or specified?
│ │
│ ├─ YES → Phase 1: Analyze intent + stakes
│ └─ NO → Ask ONE question to clarify channel
│
└─ All clear → Phase 2-3: Generate variants via message_compose
Workflow
Phase 0 — Transcription Cleanup (conditional)
Trigger: Input has ≥3 dictation artifacts (fillers + false starts + self-corrections).
When triggered:
- Clean using patterns from transcription-patterns.md
- Display: "Here's what I understood: [cleaned version]" before variants
- If input is too confused to interpret → ask for clarification instead
When NOT triggered: skip silently to Phase 1.
Phase 1 — Input Analysis
Extract from input:
| Element | How |
|---|---|
| Channel | Explicit cue ("pour un WhatsApp", "mail à") or infer from context. See channel-adaptation.md |
| Intent | Action verb: inform, follow up, decline, propose, confirm, negotiate, thank, apologize, escalate, request... |
| Recipient | Name, relationship, context. If a known profile exists in Claude's memories → enrich automatically |
| Stakes | Transactional / Relational / High-tension (drives variant count) |
| Key info | Facts, dates, amounts, expected actions — NEVER invent these |
Stakes Detection
| Stakes level | Signals | Variant count |
|---|---|---|
| Transactional | Simple info, confirmation, thanks, routine update | 2 |
| Relational | Follow-up, request, proposal, invitation, soft decline | 3 |
| High-tension | Hard decline, bad news, negotiation, conflict, escalation, boundary-setting | 3-4 |
Phase 2 — Recipient Profile Enrichment (conditional)
If a recipient name is mentioned AND a profile exists in Claude's memories:
- Retrieve: relationship type, preferred channel, preferred tone, context
- Auto-adapt variants accordingly
Profiles are stored via memory_user_edits, not in this skill. The skill only leverages them when available.
Phase 3 — Variant Generation
Variant Strategy
Each variant must have a goal-oriented label (2-4 words) describing what it prioritizes or trades off. Never use generic labels like "Formal" or "Informal".
Reference: strategy-labels.md for label catalog by intent × stakes.
Channel Adaptation
Apply channel-specific rules for length, structure, and sign-off formulas:
| Channel | Length | Structure | message_compose kind |
|---|---|---|---|
| WhatsApp/SMS | 2-6 lines | None, fluid prose | textMessage |
| Slack/Teams | 3-10 lines | Light (dashes if needed) | other |
| Free | Full (hook → body → close) | email (with subject) |
|
| LinkedIn/Social | 3-15 lines | Strong hook, short paragraphs | other |
Full rules: channel-adaptation.md
Output Format
Always use the message_compose tool:
message_compose(
kind: "email" | "textMessage" | "other",
summary_title: "[Short context]",
variants: [
{ label: "Strategy A", body: "...", subject: "..." }, // subject only for email
{ label: "Strategy B", body: "..." },
{ label: "Strategy C", body: "..." }
]
)
Phase 4 — Notes (conditional)
Display notes in conversational text AFTER the message_compose widget only if:
- Critical information is missing (date, amount, name)
- Ambiguity detected in the input
- Risk identified (inappropriate tone, sensitive info, contradiction)
Do NOT display notes if everything is clear and variants are self-sufficient.
Critical Rules
- Language: Output messages in the SAME language as user input
- No invention: NEVER add facts, dates, amounts, or commitments not present in input
- Preserve intent: Messages must convey exactly what the user wanted to say
- Always use message_compose: Never output variants as raw markdown blocks
- Labels over tones: Variant labels describe communication strategy, not formality
- Channel respect: Adapt length and structure strictly to channel constraints
- Audience adaptation: Technical jargon only for technical recipients
- Actionable output: Every message must be usable without modification
Triggers
| French | English |
|---|---|
| "corrige", "reformule" | "correct", "rephrase" |
| "aide-moi à tourner" | "help me word this" |
| "rends ça plus pro" | "make this more professional" |
| "message pour [nom]" | "message for [name]" |
| "mail à", "email pour" | "email to", "mail for" |
| "réponds à ça" | "reply to this" |
| "transforme ça en [canal]" | "turn this into a [channel]" |
| "corrige =>" | "correct =>" |
Edge Cases
| Situation | Behavior |
|---|---|
| Channel not detectable | Ask one question (prefer ask_user_input_v0) |
| Input < 10 characters | Ask for context |
| Input extremely confused | Phase 0 + confirmation before variants |
| Multiple recipients | Produce group message OR suggest splitting |
| Input is a reply to received message | Understand received message context to adapt response |
| User requests additional variant | Generate it if stakes justify |
| Input already well-written | Still produce variants with different strategies |
Limitations
This skill does NOT:
- Translate messages between languages
- Write long documents (reports, specifications, proposals)
- Generate code or technical documentation
- Send messages automatically
- Handle attachments beyond mentioning them
- Maintain conversation threads across sessions
References
- Transcription Patterns — Dictation artifacts to clean
- Strategy Labels — Label catalog by intent × stakes
- Channel Adaptation — Rules per communication channel
Version History
| Version | Date | Changes |
|---|---|---|
| 1.0.0 | 2025-12-11 | Initial release — email-only, migrated from GPT "My Corrector v2.1" |
| 2.0.0 | 2025-02-05 | BREAKING: Multi-channel support, strategic variants via message_compose, conditional cleanup, recipient profiles |