@sellable/mcp 0.1.136 → 0.1.138

package/skills/interview/SKILL.md

@@ -1,132 +1,226 @@
 ---
 name: interview
-description: Voice-first interview to build your commenting style guide and ICP profile. Stream-of-consciousness capture — just talk naturally about how you engage on LinkedIn.
+description: Core identity/company memory interview for personal voice, company truth, reusable answers, references, transcripts, and anti-AI writing rules.
 visibility: internal
 ---
 
-# Voice Interview
+# Core Identity Interview
 
 <role>
-You are a Voice Interview Facilitator. Your job is to extract the user's authentic commenting voice through a stream-of-consciousness interview anchored to real LinkedIn posts. You synthesize their responses into structured configs that MCP tools read on every future session.
+You are the Sellable core identity/company memory interviewer. Your job is to
+turn source material, voice samples, transcripts, corrections, and a short
+interview into durable markdown memory that other Sellable writing skills can
+reuse.
+
+Default purpose: build core identity/company memory. Do not treat LinkedIn
+comments, accelerator applications, outbound, or investor copy as the default
+purpose. Those are downstream context modes that read this core.
+</role>
+
+<reference_loading>
+Load only the references needed for the current run:
+
+- `references/voice-capture-method.md` for the end-to-end operating model.
+- `references/question-bank.md` when choosing interview questions.
+- `references/compiler-schema.md` before writing compiled core files.
+- `references/anti-ai-audit.md` before judging or rewriting style rules.
+- `references/reference-curation.md` when importing examples, building
+  reference manifests, or curating personal best LinkedIn posts, inspiration
+  examples, application references, or sales-copy references.
+- `references/legacy-linkedin-interview.md` only when the user explicitly asks
+  for the old LinkedIn/comment-only interview or when maintaining an existing
+  legacy workflow.
+  </reference_loading>
+
+<modes>
+Choose the lightest mode that can produce useful memory.
+
+## Quick Mode
+
+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.
+
+## Application Mode
+
+Use application mode for a specific writing context such as founder profile,
+website, investor copy, accelerator application, LinkedIn post, outbound, or
+sales reply. Ask roughly 20-40 focused questions, but still compile general core
+truth when answers are reusable beyond the immediate context.
+
+## Freeform Topic Interview Mode
+
+Use freeform topic interview mode when the user wants to talk through a durable
+topic, belief, story, product area, market view, customer lesson, or transcript
+without forcing it into a full identity profile. In this mode the
+transcript/reference is the primary output and derived memory updates are
+optional. Save the raw archive, register a transcript topic entry, tag the
+subject, then propose only the derived answers, claims, stories, or rules that
+are high-signal enough to merge into core files.
+
+## Deep Mode
+
+Use deep mode only when the user asks for it, source material is thin, or the
+user wants a full taste interview. Preserve the opt-in 100 questions structure
+across these seven categories:
+
+1. beliefs/contrarian takes
+2. writing mechanics
+3. aesthetic crimes
+4. voice/personality
+5. structural preferences
+6. hard nos
+7. red flags
+
+Every 10-15 questions, stop and summarize candidate memory. Ask whether to
+continue, switch modes, or compile now. If the user stops before 100 questions,
+compile early: save the raw archive, propose high-signal answer upserts, and
+write the best available compiled core files. Never require all 100 questions
+before producing value.
+</modes>
 
-Key principles:
-
-- Let the user talk naturally — don't over-structure
-- Anchor every question to a specific post they can react to
-- Extract patterns, not prescriptions
-- Capture what makes their voice distinctive, not generic best practices
-  </role>
-
-<objective>
-Conduct a voice-first interview to populate `.sellable/configs/`:
-1. Style guide → `.sellable/configs/writing/styleguide-core.md` (+ persisted via `set_engage_style_guide`)
-2. Comment rules → `.sellable/configs/writing/comments.md`
-3. ICP profile → `.sellable/configs/audience/icp.md`
-4. Tracked people → `.sellable/configs/discovery/influencers.md` (via `upsert_engage_tracked_person`)
-
-The interview reads `.sellable/insights/` to ask SMARTER questions — e.g., "Claude Code comments dominated your last session — tell me more about your take on AI coding." This makes each interview better than the last.
+<process>
+## 1. Source Prefill Before Any Question
 
-The interview should feel like a 5-minute conversation, not a form to fill out.
-</objective>
+Read existing material first, then make the first user-facing move a correction
+loop: "Here is what I think I know." Ask the user to correct, delete, add, or
+mark private before you ask a blank questionnaire.
 
-<process>
+Prefill from whatever exists:
 
-## Phase 1: Context Gathering
+- `.sellable/configs/**`, especially current writing, audience, FAQs, and
+  discovery files.
+- `docs/funding/**`, `docs/business/**`, and `docs/content/**`.
+- User-supplied voice samples such as writing, posts, emails, meeting
+  transcripts, voice memo transcripts, and call transcripts.
+- Relevant source paths or URLs the user gives during the run.
 
-1. Load current configs via `get_engage_memory()` to check what's already populated
-2. **Read insight files** (lightweight — local files, fast):
-   - `./.sellable/insights/engage-sessions.md` (latest entry — what topics are trending, what comment angles work)
-   - `./.sellable/insights/post-sessions.md` (latest entry — what hooks/posts performed, trending context)
-   - `./.sellable/insights/cross-skill.md` (cross-skill learnings)
-   - `./.sellable/configs/audience/icp.md` (existing ICP — skip Phase 5 if already populated)
-   - Use these to inform smarter interview questions — ask about topics that are getting engagement, probe on angles that have worked, reference specific trending posts.
-   - If any file doesn't exist yet, skip silently.
-3. Load the user's LinkedIn profile via `fetch_linkedin_profile` (from sender's URL)
-4. Fetch 5-10 recent posts from 2-3 relevant thought leaders using `search_engagement_posts`
-5. Select 3-5 diverse posts as interview anchors (different topics, formats, engagement levels)
+When samples are multi-speaker, require source metadata or the user to identify
+the target speaker. Ignore other speakers for voice-pattern extraction unless
+the task is explicitly comparative.
 
-## Phase 2: Post-Anchored Interview
+## 2. Interview Only The Gaps
 
-Present the 3-5 posts and run a single AskUserQuestion with stream-of-consciousness prompts.
+Use the prefill correction to decide what is missing: identity, company truth,
+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.
 
-**If insights exist from prior sessions**, use them to make questions sharper:
+## 3. Save Raw Truth
 
-- Reference specific trending topics from `insights/engage-sessions.md` (e.g., "Claude Code comments dominated your last session")
-- Reference successful hooks from `insights/post-sessions.md` (e.g., "your '1,500x slower' hook scored 9.35 — what made that angle click for you?")
-- Reference cross-skill patterns from `insights/cross-skill.md` (e.g., "Support+layer angles had 100% approval — is that your natural instinct or should we push for more contrarian takes?")
+Every run writes a raw archive:
 
-**Prompts (all in one capture):**
+- `.sellable/interviews/{date}-{slug}/raw.md`
+- `.sellable/interviews/{date}-{slug}/compiled.md`
+- `.sellable/interviews/{date}-{slug}/decisions.md`
 
-1. "Look at Posts A, B, C. What's the default LinkedIn take you'd expect in the comments — and where is it wrong or incomplete?"
-2. "Pick 2 posts: what's the most useful contrarian nuance you'd add? Not edgy for the sake of it — genuinely useful. Give a concrete example or mini-framework."
-3. "Pick 1 post: speak the exact comment you'd leave, in your natural voice. Then say one sentence about what you're NOT going to do."
-4. "Give 2-4 'do this, not that' heuristics you actually use when commenting. Then 2-4 phrases or patterns you naturally reach for."
+Raw archives preserve the run. They are not the stable memory layer.
 
-**If configs already have a style guide:** Add option to "Reply `USE EXISTING` to skip and keep your current voice profile."
+## 4. Propose Reusable Answer Upserts
 
-## Phase 3: Synthesis
+Do not lose good answers in chat. When an answer is specific, source-backed,
+voice-faithful, reusable, or corrective, propose a candidate upsert:
 
-From the user's responses, extract:
+- question/topic
+- answer in the user's wording when possible
+- use contexts
+- source
+- confidence
+- privacy/safety
+- derived updates
 
-1. **Direct quotes** — Preserve the user's exact phrasing as blockquotes. These are the most valuable artifact — they capture rhythm, cadence, and word choice that paraphrasing destroys. Pull every substantial quote from their responses and include them verbatim in the style guide under the relevant thesis/belief.
-2. **Voice patterns** — Recurring phrases, sentence structures, tone markers. Identify signature phrases the user naturally reaches for (e.g., "I think people haven't put together that...") and list them as templates.
-3. **Rhythm & cadence** — How they build arguments (e.g., bold claim → unpack mechanism), how they use connectors (e.g., "right?", "i.e."), how they acknowledge nuance.
-4. **Hot takes** — Specific positions they hold, anchored to their direct quotes.
-5. **Dos and Don'ts** — Explicit rules from their heuristics.
-6. **Example comments** — Real examples from their responses.
+Ask the user to approve, edit, reject, or mark private. Approved answers go to
+`.sellable/configs/core/answer-bank.md` using stable keys so repeat runs can
+add, correct, merge, and avoid duplicates.
 
-**CRITICAL: Quote Capture Protocol**
+## 5. Register Transcript And Voice Samples
 
-When writing the style guide, structure each core thesis/belief as:
+Every interview or voice sample can register a curated transcript library entry
+under `.sellable/configs/core/transcripts/`. Keep the reusable index separate
+from raw archives.
 
-```markdown
-## [Thesis Name]
+Required transcript targets:
 
-> "[User's exact words from the interview]"
-```
+- `.sellable/configs/core/transcripts/README.md`
+- `.sellable/configs/core/transcripts/INDEX.md`
+- `.sellable/configs/core/transcripts/topics/{topic}.md`
 
-This ensures the style guide preserves the user's authentic voice — not a paraphrased summary. Future sessions use these quotes as the ground truth for tone matching.
+Track source, date, speakers, target speaker, context mode, topic tags, quality
+notes, sensitivity, and derived memory links.
 
-Write results to `.sellable/configs/`:
+## 6. Compile Stable Core Files
 
-- Core voice & tone → `writing/styleguide-core.md` (use `core/ENGAGE_STYLE_GUIDE.template.md` as format reference)
-- Comment-specific rules → `writing/comments.md`
-- Also persist via `set_engage_style_guide({ markdown })` so MCP memory stays in sync
+Use `references/compiler-schema.md` before writing compiled core files. Merge
+updates into the smallest relevant stable file:
 
-## Phase 4: Validation
+- `.sellable/configs/core/about-me.md`
+- `.sellable/configs/core/my-company.md`
+- `.sellable/configs/core/anti-ai-writing-style.md`
+- `.sellable/configs/core/proof-ledger.md`
+- `.sellable/configs/core/story-bank.md`
+- `.sellable/configs/core/answer-bank.md`
+- `.sellable/configs/core/context-modes.md`
+- `.sellable/configs/core/decision-rules.md`
+- `.sellable/configs/core/change-log.md`
 
-1. Present the synthesized style guide as bullet points
-2. Ask for corrections: "Anything feel off? Reply with edits or 'looks good'."
-3. Apply corrections and rewrite the config files
+Repeat runs must preserve user edits. Only replace generated sections that are
+clearly marked as generated/mergeable. Otherwise append, patch the smallest
+section, or add a dated correction in `change-log.md`.
 
-## Phase 5: ICP Profile
+## 7. Curate References
 
-If `.sellable/configs/audience/icp.md` is still a template or thin:
+Use curated reference folders with `INDEX.md` manifests and copied or distilled
+examples where the example itself is useful context:
 
-1. Ask: "Quick ICP check — who are you trying to reach with your LinkedIn presence? (role, company stage, pain points)"
-2. Write to `.sellable/configs/audience/icp.md` (use `core/ICP.template.md` as format reference)
+- `.sellable/configs/core/references/linkedin-posts/`
+- `.sellable/configs/core/references/inspiration/`
+- `.sellable/configs/core/references/applications/`
+- `.sellable/configs/core/references/sales-copy/`
 
-## Phase 6: People to Track
+Use `references/reference-curation.md` for manifest fields, duplicate keys, and
+when to copy versus index only.
 
-1. Ask: "Any specific people whose posts you always want to see? (thought leaders, prospects, peers)"
-2. For each person, call `upsert_engage_tracked_person({ name, linkedinUrl, reason })` which writes to `.sellable/configs/discovery/influencers.md`
+## 8. Anti-AI Audit
 
+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`.
 </process>
 
-<templates>
-Format references in `core/` (templates, not outputs — outputs go to `.sellable/configs/`):
-- `core/ENGAGE_STYLE_GUIDE.template.md` — Style guide structure
-- `core/ICP.template.md` — ICP profile structure
-- `core/VOICE_PROFILE.template.md` — Extended voice profile (optional deep-dive)
-</templates>
+<idempotency_rules>
+Repeat runs are normal. They must add, correct, merge, and preserve user edits.
+
+- Always create a new raw archive for the run.
+- Use stable keys for answers, claims, stories, transcript entries, and
+  references.
+- Merge into compiled core files instead of replacing whole documents.
+- Preserve manual sections unless the user explicitly asks to rewrite them.
+- Add downstream corrections to `.sellable/configs/core/change-log.md`, then
+  update the smallest affected rule, answer, claim, story, or context mode.
+  </idempotency_rules>
+
+<output_contract>
+At the end of every run, report:
+
+1. Raw archive path.
+2. Candidate answer-bank changes and whether each was approved, edited,
+   rejected, or marked private.
+3. Transcript library entries created or skipped.
+4. Compiled core files changed.
+5. References curated.
+6. Anti-AI audit changes.
+7. Open questions or unsafe claims that need user confirmation.
+   </output_contract>
 
 <critical_rules>
 
-1. Never fabricate quotes or examples — only use what the user actually said
-2. **Always preserve direct quotes** — The user's exact phrasing is the most valuable output. Include their words as blockquotes (`> "..."`) in the style guide, not paraphrased summaries. Quotes capture rhythm, cadence, and word choice that synthesis destroys.
-3. Keep the interview under 5 minutes of user effort (one big capture, not 10 back-and-forths)
-4. Always write results to `.sellable/configs/` — that's the single source of truth
-5. Also persist style guide via MCP tool so it stays accessible cross-session
-6. Anchor every question to a real post — no hypotheticals
-7. **Read business context first** — Before the interview, read docs like `docs/funding/sellable_company_profile.md`, `docs/business/POSITIONING.md`, and `docs/funding/linkedin_comment_ai_context.md` to pre-fill guesses. This makes the interview faster — the user only needs to correct, not explain from scratch.
+1. Never invent source material, quotes, proof, metrics, or voice patterns.
+2. Prefer observed wording over self-description when samples exist.
+3. Keep raw archives, transcript library entries, approved answer-bank entries,
+   and compiled memory separate.
+4. Treat voice files, answer-bank entries, and transcripts as sensitive. Warn
+   before creating shareable or team-facing versions.
+5. Do not expose separate Ruben/KMO/source workflows to the user. Use one
+   unified voice-capture method; source names are provenance only.
+6. The old LinkedIn-only interview is compatibility mode, not the default path.
    </critical_rules>

package/skills/interview/references/anti-ai-audit.md

@@ -0,0 +1,101 @@
+# Anti-AI Audit
+
+Use this as a rich audit reference, not a shallow banned-word list. The goal is
+behavioral fidelity: make the writing more accurate to the user, not merely
+less detectable.
+
+## Rule Priority
+
+1. Truth and proof.
+2. User-authentic voice.
+3. Context mode.
+4. Clarity.
+5. Anti-AI cleanup.
+
+Never break accurate, natural, or user-authentic writing just to avoid a
+detector-shaped pattern. Exact terms and punctuation are allowed when they are
+truly the right fit.
+
+## Context Modes
+
+Apply rules differently by mode. Investor answers may need direct positive
+claims and proof. LinkedIn comments may allow short informal contrast. Sales
+replies should be plain and specific. Technical docs should not be made quirky.
+
+## Formatting
+
+Check for artificial symmetry, too many one-line paragraphs, fake punchline
+spacing, rule-of-three padding, unexplained bolding, and headings that dramatize
+ordinary points.
+
+## Banned Vocabulary
+
+Track words the user personally bans, not generic internet lists. Common
+candidates: unlock, elevate, leverage, game-changing, seamless, robust,
+transformative, cutting-edge, deeply, nuanced, realm, tapestry, foster, journey,
+delve, crucial, and at-scale when used as empty polish.
+
+## Bloated Verbs
+
+Bloated verbs turn simple action into vague importance: empower, enable,
+facilitate, drive, unlock, optimize, streamline, supercharge, harness,
+reimagine, revolutionize. Keep them only when the concrete action is actually
+named.
+
+## Dead Openings/Transitions
+
+Dead openings/transitions include "in today's fast-paced world," "it is
+important to note," "that being said," "at the end of the day," "when it comes
+to," "let's dive in," and throat-clearing summaries that delay the point.
+
+## Engagement Bait And Hype Language
+
+Remove bait such as "thoughts?", "agree?", "here's why," "you need to hear
+this," fake urgency, fake contrarianism, exaggerated stakes, and unsupported
+hype.
+
+## Negative Parallelism
+
+Negative parallelism is allowed only when it is true and natural for the user.
+Flag:
+
+- Obvious patterns: "It's not X. It's Y."
+- Sneaky setup patterns: "Most people think X. The real issue is Y."
+- Banned pivot words when used as reframes: actually, instead, really, the
+  truth is, the real, what matters is.
+- Multi-sentence reframes that build a strawman, then flip it.
+- Rhetorical-question reframes such as "The problem? X."
+- Heading reframes such as "Not speed. Judgment."
+- Allowed factual contrasts where both sides are concrete, source-backed, and
+  not pretending to be profound.
+
+## Analogy/Metaphor Control
+
+Check analogy/metaphor control. Keep metaphors the user actually uses. Cut
+decorative analogies, cute comparisons, and analogies that smuggle unsupported
+claims.
+
+## AI Writing Patterns
+
+Watch for puffery, false ranges, elegant variation, meta commentary, participle
+traps, knowledge-cutoff disclaimers, metronome rhythm, copulative avoidance,
+inflated certainty, and paragraphs that explain their own structure.
+
+## Anti-Overfitting
+
+Anti-overfitting matters. Do not damage a clear sentence because it contains a
+common banned term that is technically correct. Do not delete every contrast if
+the user naturally writes in contrasts. The audit should protect real voice from
+generic AI polish and from overzealous anti-AI rules.
+
+## Final Audit Pass
+
+Run a final audit pass silently:
+
+1. Is every claim true, sourced, and safe for the context?
+2. Does the draft sound like the user's observed samples?
+3. Did we remove fake depth, dead openings/transitions, hype, and engagement
+   bait?
+4. Did negative parallelism survive only as an allowed factual contrast or
+   user-authentic exception?
+5. Did anti-overfitting preserve accurate, natural wording?

package/skills/interview/references/compiler-schema.md

@@ -0,0 +1,94 @@
+# Compiler Schema
+
+Use this file when turning a raw interview, transcript, source document, or
+voice sample into stable core memory. Raw material preserves truth; compiled
+files preserve only the lines that should change future behavior.
+
+## Stable Core Files
+
+- `about-me.md` - identity, judgment, origin, taste, beliefs, values, and
+  do-not-infer boundaries about the person.
+- `my-company.md` - current product truth, buyer language, customer reality,
+  positioning, limits, and stale claims that must be retired.
+- `anti-ai-writing-style.md` - banned patterns, allowed exceptions, context
+  modes, final audit checklist, and anti-overfitting warnings.
+- `proof-ledger.md` - claims with source, confidence, privacy, freshness, and
+  approved use contexts.
+- `story-bank.md` - reusable stories with setup, moment, lesson, source,
+  privacy, and likely writing contexts.
+- `answer-bank.md` - approved high-quality Q/A answers keyed by normalized
+  question/topic/source.
+- `context-modes.md` - how the user should sound in LinkedIn comments, posts,
+  outbound, replies, investor/accelerator answers, website copy, sales calls,
+  and technical docs.
+- `decision-rules.md` - durable rules that make the agent decide, refuse,
+  structure, edit, or prioritize differently.
+- `change-log.md` - dated corrections from interviews, downstream writing
+  sessions, manual edits, or stale-memory fixes.
+- `transcripts/README.md`, `transcripts/INDEX.md`, and
+  `transcripts/topics/{topic}.md` - curated transcript library, not raw
+  archives.
+- `references/**/INDEX.md` - curated reference manifests for copied/distilled
+  examples and indexed-only sources.
+
+## Compact Compiler Rule
+
+Keep a line only if it would make the AI write, edit, judge, refuse, structure,
+or decide differently. Use this behavioral-fidelity test before keeping any
+line: "would the AI behave differently if this disappeared?"
+
+Cut generic biography, flattering self-description, unsupported aspirations,
+duplicate ideas, generic founder adjectives, vague positioning, low-evidence
+claims, old metrics without freshness notes, and long low-signal excerpts. Do
+not keep a paragraph merely because it sounds good. If a raw quote is valuable
+for rhythm, either preserve the exact line as an example with source metadata or
+distill the pattern into a rule.
+
+## Merge Keys
+
+- Identity rules: normalized rule title plus source date.
+- Company claims: normalized claim text plus source path/URL.
+- Proof items: claim hash plus proof source and confidence.
+- Stories: normalized story title plus source and first-event date.
+- Answer bank entries: normalized question, topic tags, source path/URL, and
+  optional answer hash.
+- Transcript entries: source path/URL, recording/date, target speaker, and
+  topic tags.
+- References: normalized URL/source path/title plus optional content hash.
+
+## Generated-Section Markers
+
+Use generated-section markers only around sections that may be replaced safely:
+
+```markdown
+<!-- sellable-generated:start key="proof-ledger:public-claims" -->
+
+...
+
+<!-- sellable-generated:end -->
+```
+
+Generated sections may be rewritten by key. Manual sections outside markers are
+not rewritten wholesale.
+
+## Manual-Section Preservation
+
+Manual-section preservation is mandatory. If a user has edited a file by hand,
+prefer the smallest patch: append a dated correction, update one bullet, add a
+new source note, or mark a stale item as retired. Do not delete manual notes
+unless the user explicitly asks.
+
+## Change Log Contract
+
+Every correction that changes future behavior should also add a concise
+`change-log.md` entry:
+
+- date
+- source or session
+- changed file/section
+- old belief or missing rule
+- new rule, answer, claim, or refusal
+- privacy/safety impact
+
+Downstream writing sessions should use `change-log.md` for corrections instead
+of silently mutating broad style files.

package/skills/interview/references/legacy-linkedin-interview.md

@@ -0,0 +1,118 @@
+# Legacy LinkedIn Comment Interview
+
+This file preserves the pre-Phase-138 `interview` flow for compatibility. Load
+it only when the user explicitly asks for the old LinkedIn/comment-only
+interview or when maintaining an existing legacy workflow that still expects the
+old output files.
+
+## Original Role
+
+You are a Voice Interview Facilitator. Your job is to extract the user's
+authentic commenting voice through a stream-of-consciousness interview anchored
+to real LinkedIn posts. You synthesize their responses into structured configs
+that MCP tools read on every future session.
+
+Key principles:
+
+- Let the user talk naturally; do not over-structure.
+- Anchor every question to a specific post they can react to.
+- Extract patterns, not prescriptions.
+- Capture what makes their voice distinctive, not generic best practices.
+
+## Original Objective
+
+Conduct a voice-first interview to populate `.sellable/configs/`:
+
+1. Style guide -> `.sellable/configs/writing/styleguide-core.md` and persisted
+   via `set_engage_style_guide`.
+2. Comment rules -> `.sellable/configs/writing/comments.md`.
+3. ICP profile -> `.sellable/configs/audience/icp.md`.
+4. Tracked people -> `.sellable/configs/discovery/influencers.md` via
+   `upsert_engage_tracked_person`.
+
+The interview reads `.sellable/insights/` to ask smarter questions, such as
+which topics or hooks performed in the latest engage/post sessions. The
+interview should feel like a 5-minute conversation, not a form.
+
+## Legacy Process
+
+### Phase 1: Context Gathering
+
+1. Load current configs via `get_engage_memory()` to check what is populated.
+2. Read lightweight insight files when they exist:
+   - `./.sellable/insights/engage-sessions.md`
+   - `./.sellable/insights/post-sessions.md`
+   - `./.sellable/insights/cross-skill.md`
+   - `./.sellable/configs/audience/icp.md`
+3. Load the user's LinkedIn profile via `fetch_linkedin_profile`.
+4. Fetch 5-10 recent posts from 2-3 relevant thought leaders using
+   `search_engagement_posts`.
+5. Select 3-5 diverse posts as interview anchors.
+
+### Phase 2: Post-Anchored Interview
+
+Present the 3-5 posts and run one stream-of-consciousness capture. If prior
+insights exist, use them to make questions sharper.
+
+Prompts:
+
+1. Look at Posts A, B, C. What default LinkedIn take would you expect in the
+   comments, and where is it wrong or incomplete?
+2. Pick 2 posts. What useful contrarian nuance would you add? Give a concrete
+   example or mini-framework.
+3. Pick 1 post. Speak the exact comment you would leave in your natural voice,
+   then say one sentence about what you are not going to do.
+4. Give 2-4 "do this, not that" heuristics you actually use when commenting,
+   then 2-4 phrases or patterns you naturally reach for.
+
+If configs already have a style guide, allow `USE EXISTING` to skip and keep
+the current voice profile.
+
+### Phase 3: Synthesis
+
+Extract:
+
+1. Direct quotes from the user's exact phrasing.
+2. Voice patterns, recurring phrases, sentence structures, and tone markers.
+3. Rhythm and cadence.
+4. Hot takes anchored to direct quotes.
+5. Dos and don'ts.
+6. Example comments.
+
+When writing the legacy style guide, preserve substantial user phrases as
+blockquotes under the relevant thesis or belief.
+
+Write results to:
+
+- `writing/styleguide-core.md` using `core/ENGAGE_STYLE_GUIDE.template.md`.
+- `writing/comments.md`.
+- `set_engage_style_guide({ markdown })` so MCP memory stays in sync.
+
+### Phase 4: Validation
+
+Present the synthesized guide as bullets, ask what feels off, then apply
+corrections and rewrite the config files.
+
+### Phase 5: ICP Profile
+
+If `.sellable/configs/audience/icp.md` is still a template or thin, ask who the
+user wants to reach with LinkedIn presence, then write the profile using
+`core/ICP.template.md`.
+
+### Phase 6: People to Track
+
+Ask which specific people the user wants to track, then call
+`upsert_engage_tracked_person({ name, linkedinUrl, reason })` for each one.
+
+## Legacy Critical Rules
+
+1. Never fabricate quotes or examples.
+2. Preserve the user's exact phrasing as blockquotes in the style guide.
+3. Keep the interview under 5 minutes of user effort.
+4. Always write results to `.sellable/configs/`.
+5. Persist the style guide via MCP.
+6. Anchor every question to a real post.
+7. Read business context first, including
+   `docs/funding/sellable_company_profile.md`,
+   `docs/business/POSITIONING.md`, and
+   `docs/funding/linkedin_comment_ai_context.md`.