typo-prevention-cold-drafts

name	typo-prevention-cold-drafts
description	Pre-send typo gate for cold drafts that share a narrative scaffold across multiple recipients. Enforces cross-batch scaffold consistency, canonical-source copy-paste discipline, and domain-noun semantic verification before status flips to sent.
triggers	["drafting cold message in cold_contacts/","cold outreach batch","before marking status: sent","user says send / 보냈어 / 発送 / sent","logging verbatim ## Sent section"]

Typo Prevention for Cold Drafts (cross-batch scaffolds)

The failure mode this skill exists to prevent

When a batch of cold messages share an identical narrative scaffold sentence (the "spine" of the pitch, reused across N recipients), retyping that sentence N times gives N chances to drift. Domain-specific terms in non-English locales are the easiest place to drift: a single mistyped noun produces a body that is grammatically valid but semantically wrong, which a casual review pass will not catch.

Cost of a single semantic typo to a high-value target: the recipient reads a sentence whose noun is incoherent for the surrounding clause and silently downgrades trust. You will not get told.

When to activate

Apply this gate when ANY of:

• Drafting a cold message destined for cold_contacts/{slug}.md.
• Drafting a batch (2+ drafts sharing a narrative scaffold).
• About to flip status: draft → status: sent.
• User says "send" / "보냈어" / "発送" / "sent" / etc.
• Writing a ## Sent (..., verbatim) section.

The four checks

1. Canonical-source copy-paste

Identify the canonical source for any scaffold sentence shared across multiple drafts. Suggested sources, in order of preference:

• A previously-sent draft that was confirmed-received without correction.
• A canonical "ground truth" file you maintain for your domain (e.g., a verbatim quote from a domain expert, a public industry report sentence, a regulatory phrasing).

When a scaffold appears in multiple drafts, copy-paste from the canonical, then edit only the per-recipient anchors. Never retype.

2. Cross-batch scaffold diff

Before marking any draft sent, diff its scaffold sentence against the canonical and against batch siblings. Any token-level divergence is either an intentional per-recipient variation OR a typo. Verify which.

canonical scaffold: <reference sentence>
draft N:            <draft sentence>
                          ^^^^ DIVERGENCE  ← intentional or typo?

3. Domain-noun semantic check

Read each domain-specific noun in context and ask: "Does this noun fit this clause?"

The high-trap categories are language-specific. Maintain a project-specific trap list as you accumulate incidents. Generic categories to watch for:

Trap category	Example shape (any locale)
Confusable homophones / near-homophones	two domain nouns that sound similar but mean different things
Word-spacing changes meaning	one-word fixed term vs. two-word compound (very common in Korean / Japanese / Chinese)
Term-vs-process distinction	a noun for the artifact vs. a noun for the act ("photo" vs "photographing")
Synonym variants in same body	e.g. mixing "ins. company" / "insurer" / "carrier" within one message
Tense / aspect on key verbs	"shoot" vs "re-shoot", "verify" vs "verified"

For terms outside your trap list, the question is always: "If a recipient reads this noun cold, will it parse correctly given the surrounding clause?"

4. Pre-mark-sent verbatim re-read

Before flipping status: sent and writing the ## Sent (verbatim) section:

• Re-read the entire body once.
• Focus on every domain-specific noun.
• Compare against canonical scaffold for shared sentences.
• ANY divergence = halt and surface to user before logging.

Success criteria

• Zero domain-incoherent typos in sent cold drafts.
• All drafts in a batch share a byte-identical narrative scaffold (modulo intentional per-recipient anchors).
• ## Sent (verbatim) body matches what the recipient actually saw.

Anti-patterns

Don't	Do
Retype a shared scaffold sentence from memory for each draft	Copy-paste from a canonical source, edit only anchors
Trust user review to catch body typos	The user edits anchors and skims body; body verification is yours
Mark sent immediately on user "send" signal	Always do step 4 (pre-mark verbatim re-read) before flipping status
Rely on locale spell-check as the gate	Spell-check does not catch domain-semantic errors (the typo is a real word, just the wrong real word)
Treat scaffolds as flexible boilerplate	Scaffolds are templates; copy-paste discipline is non-negotiable

Recovery if a typo escapes

1. Log the typo as-sent (verbatim rule trumps correction).
2. Note the typo in the file's notes field.
3. If a follow-up message is queued, the follow-up can use the corrected term and naturally overwrite. Do NOT send a "correction" message.
4. Add the specific trap to your project's trap list so it is caught next time.

Open questions

• Should this gate run as a Bash diff (e.g., extracting scaffold lines from recently-modified cold_contacts/*.md and diff-ing them automatically)? The current form is a behavioral checklist; an automated diff step is a clean upgrade.
• For projects with multiple working locales, is one multilingual skill sufficient, or do per-locale sibling skills compound better?