@sellable/mcp 0.1.140 → 0.1.141

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sellable/mcp",
-  "version": "0.1.140",
+  "version": "0.1.141",
   "type": "module",
   "description": "Sellable MCP server for Claude Code and Codex campaign workflows",
   "main": "dist/index.js",

package/skills/interview/SKILL.md

@@ -40,6 +40,30 @@ Choose the lightest mode that can produce useful memory.
 Use quick mode when source material is strong or the user wants a fast pass.
 Ask roughly 10-20 high-leverage questions after source prefill and correction.
 
+## Draft-First Voice Calibration Mode
+
+Use draft-first voice calibration when the user wants the skill to capture
+voice/taste and already has some product or identity memory. Do not ask the
+user to invent examples from a blank page. Instead:
+
+1. Pick one concrete context, such as a LinkedIn post, sales call answer,
+   website line, investor answer, cold outbound note, reply, or objection
+   response.
+2. Draft how you think the user would answer based on current memory:
+   "Based on what I know, I think you'd say it like this..."
+3. Ask the user to fix what feels wrong. Encourage rough dictated feedback,
+   audio-transcript paste, or voice-note transcript. The correction can be
+   messy; do not require polished writing.
+4. Extract only behavior-changing calibration from the correction: phrases to
+   keep, phrases to ban, tone shifts, proof-safety rules, structure rules,
+   decision rules, and context-specific voice laws.
+5. Save the raw correction in the run archive, update the smallest durable core
+   sections, then continue with the next calibration draft.
+
+Use choices only to reduce friction, not as a substitute for the user's
+correction. Good prompts look like: "What feels off: too polished, too harsh,
+too vague, wrong proof, wrong phrase, or something else?"
+
 ## Application Mode
 
 Use application mode for a specific writing context such as founder profile,
@@ -105,6 +129,24 @@ proof, stories, taste, anti-style, context modes, reference examples, transcript
 coverage, or reusable answers. Ask compact questions from
 `references/question-bank.md`; do not dump the whole bank.
 
+When the gap is voice/taste, prefer draft-first calibration over open-ended
+questions. Show a specific answer, post, reply, or paragraph that current memory
+would produce, then ask the user to correct it. The correction itself is the
+interview answer.
+
+Audio-friendly workflow:
+
+- If the user wants to record audio, tell them to dictate or record their
+  reaction to the draft and paste the transcript or rough notes back into the
+  session.
+- If the host provides an audio transcript, treat it as user-supplied source
+  material and save it in the raw archive with date, context, target speaker,
+  and sensitivity.
+- If an audio file is attached but no transcription is available, ask for a
+  transcript or rough bullet notes instead of pretending to transcribe it.
+- Preserve the user's rough wording when it teaches voice, but compile only the
+  smallest behavior-changing rules into core files.
+
 ## 3. Save Raw Truth
 
 Every run writes a raw archive:
@@ -185,6 +227,17 @@ Before finalizing, run the silent audit pass in `references/anti-ai-audit.md`.
 Capture banned AI-ish patterns, user-authentic exceptions, context-only rules,
 negative-parallelism limits, proof hygiene, and anti-overfitting warnings in
 `.sellable/configs/core/anti-ai-writing-style.md`.
+
+## 9. Continue After Saves
+
+Do not end a calibration run immediately after writing memory. After each save,
+give a compact progress update:
+
+1. what changed
+2. what voice gap remains
+3. the next specific calibration draft or question
+
+Then continue unless the user explicitly says stop, compile only, or pause.
 </process>
 
 <idempotency_rules>

package/skills/interview/references/voice-capture-method.md

@@ -18,7 +18,9 @@ materials are provenance; do not expose them as separate user-facing workflows.
    argument shape, story habits, proof style, and recurring refusals.
 5. Run the gap interview: ask only for missing identity, company truth, proof,
    taste, context modes, transcript topics, answer-bank candidates, and
-   reference curation.
+   reference curation. For voice/taste gaps, use draft-first calibration:
+   produce the answer/post/reply you think the user would write, then ask the
+   user to correct it.
 6. Do durable compile: raw archive first, then stable core files, answer bank,
    transcript library, proof/story/rule updates, and reference manifests.
 7. Audit/correct: run the anti-AI audit, show compact candidate memory, ask for
@@ -32,6 +34,10 @@ materials are provenance; do not expose them as separate user-facing workflows.
 - Natural samples and transcripts as evidence, with target-speaker filtering.
 - Observed language patterns over generic self-description.
 - Correction loops: "here is what I think I know" before a questionnaire.
+- Draft-first calibration: "based on current memory, I think you would say it
+  like this" before asking the user to fix voice/taste.
+- Dictated or voice-note transcript corrections. Rough spoken feedback is valid
+  source material when it has target speaker, date, context, and sensitivity.
 - Raw archives separate from curated transcript entries.
 - Approved reusable answers separate from raw answers.
 - Compact compiled memory that changes future behavior.
@@ -49,6 +55,11 @@ materials are provenance; do not expose them as separate user-facing workflows.
 - Do not require exact sample-size gates. A large corpus helps, but it is not a
   blocking prerequisite.
 - Do not do theatrical impersonation or invent personality quirks.
+- Do not make the user create examples from scratch when current memory can
+  generate a concrete calibration draft first.
+- Do not stop after each memory write during calibration; report what changed,
+  name the remaining voice gap, and continue with the next specific draft unless
+  the user pauses.
 - Do not optimize for AI-detector avoidance at the expense of accurate,
   natural, user-authentic writing.
 - Do not use giant questionnaires as the default. The 100-question version is