@easysociable/cli 0.1.1 → 0.1.2

package/skills/easysociable-content/starter/CLAUDE.md

@@ -0,0 +1,756 @@
+# Content Engine — Agent operating manual
+
+This folder is the local **Content Engine** for EasySociable.
+
+**Human interface:** conversation only. The user does **not** edit markdown by hand. Observatory UI: `workbench/` at `http://127.0.0.1:8765/workbench/`.  
+**Agent interface:** you create/update files; always return a structured result card in chat. After markdown writes, run `python3 workbench/build.py` (leave `serve.py` running if already up).
+
+## Defaults (v1)
+
+- Folder tree: **English** (`Content Engine/`). Do not ask language.
+- Default **account:** first **active** handle in `accounts/_index.md` (starter: `your-handle` — replace it)
+- Default **platform:** `x` (watchlist / catalog filter when unset)  
+- **开一单:** do **not** silent-default platform — **ask**; multi-select → **N runs**  
+- Do not invent metrics, customers, or quotes the user did not provide
+
+### Resolve account → profile (required)
+
+1. Read **`accounts/_index.md`** for the handle (default: first active row)  
+2. Get `profile_id`, `voice`, `platforms`, `default_platform`, `overrides`  
+3. Load **`profiles/<profile_id>/pillars.md`** + **`audience.md`** (core + current platform section)  
+4. Voice: if account `voice` is `profile` or empty → **`profiles/<profile_id>/voice.md`** (includes **Identity** + tone); else load the explicit path. Never invent bio facts not in Identity.  
+5. If `overrides` is **non-empty**, read that path (e.g. `accounts/overrides/<handle>.md`) and apply last  
+6. Craft: **always** load `platforms/<platform>.md` for the active platform (see `platforms/README.md` for id → file; `instagram` → `ig.md`). Missing file → stop and tell the user; do not invent craft rules.  
+7. Catalog filter: swipe/atoms/claims where `platforms` matches **and** (`profiles` is `*` **or** contains `profile_id`)
+
+## File map
+
+| Path | Role |
+|------|------|
+| `CLAUDE.md` | Operating manual |
+| `profiles/_index.md` | Profile registry |
+| `profiles/<id>/pillars.md` | Topic pillars (what to talk about) |
+| `profiles/<id>/audience.md` | Readers + platform emphasis |
+| `profiles/<id>/voice.md` | Identity + tone |
+| `accounts/_index.md` | handle → profile, platforms, voice, overrides |
+| `accounts/overrides/<handle>.md` | Optional account deltas |
+| `platforms/*.md` | Platform craft only |
+| `inbox/_index.md` | Capture registry |
+| `inbox/entries/C-*.md` | Per-capture working notes |
+| `raw/_index.md` | Authoritative source registry |
+| `raw/*.md` | Source bodies |
+| `topics/_index.md` | Topic queue (one row per topic + `item` + src engagement) |
+| `topics/items/T-*.md` | Content Opportunity (structured fields + scores) |
+| `topics/WATCHLIST.md` | Watchlist router (per-profile files) |
+| `discovery/_index.md` | Analyzed watchlist signals and promotion status |
+| `discovery/D-*.md` | Readable Source Digest, source limits, and workflow state |
+| `topics/watchlists/<profile_id>.md` | Profile-scoped radar handles |
+| `topics/watchlists/_shared.md` | Shared radar (`*`, all profiles) |
+| `topics/scoring.md` | Scout gate + opportunity five-dim (context/audience/conflict/insight/evidence) |
+| `swipe/_index.md` | Structures (+ `profiles` column) |
+| `wiki/atoms/_index.md` | Atoms (+ `profiles`) |
+| `wiki/claims/_index.md` | Claims (+ `profiles`) |
+| `runs/_index.md` | Job queue (one row per **platform run**) |
+| `runs/<slug>/` | One platform draft → pack → feedback |
+| `published/_index.md` | One row per **shipped URL** (one run → usually one row) |
+| `exports/<slug>/` | Rendered images for that run (optional `<platform>/` subfolder) |
+
+## Routing
+
+| User intent | Action |
+|-------------|--------|
+| capture / 记一下 (+ link) | **§ Capture** only — **not** topics by default |
+| 加 watchlist / 更新 WATCHLIST | Edit `topics/watchlists/<profile>.md` or `_shared.md` (**ask profile** if unset; set **`platform`**) |
+| 按 watchlist 选题 / 扫 list 选题 | **§ Discovery** — **ask profile if unset**; filter **`platform`** (default `x`); analyze signals before promoting Topics |
+| 扫 TikTok/LinkedIn/… watchlist | Same; only rows where `platform` matches (after profile resolve) |
+| 扫 inbox 选题 | **§ Topics** — from captures without topic yet (max 5) |
+| 扫 watchlist / 发现内容 | **§ Discovery** — create Discovery records; do not create Topics directly |
+| 生成 topic (no source specified) | **§ Topics** — ask for the input/source; never choose an existing Discovery silently |
+| 链接直接选题 / 把 C-xxx 做成选题 | **§ Topics** — create T- (may link capture) |
+| 开一单 / 用 T-xxx 开一单 | **§ Open a run** — **ask platform(s)** if unset; multi-select → **N runs** (bind `topic_id` if from T-) |
+| Quick draft | Resolve profile; selection; draft in chat |
+| 收成 swipe/atom/claim | **§ Library ingest** after approval |
+| 选题后同意进 inbox | Capture that URL/C-link then optional library offer |
+| Post URL after ship | feedback + **§ Published index** |
+| 复盘 / 批准晋级 | Claims promote + published index |
+| Update pillars/audience | Edit **`profiles/<id>/…`** |
+| Update account binding | Edit **`accounts/_index.md`** |
+| 出图 | **§ PagePack + render** |
+
+---
+
+## § Capture (inbox + raw indexes)
+
+### When
+
+User stores an idea, link, competitor structure, or source — not a pure Q&A.
+
+### Always write (scheme B)
+
+1. **`inbox/entries/C-YYYYMMDD-XX.md`** — working notes (our understanding; not necessarily full source text)  
+2. **Row at top of `inbox/_index.md`** — registry (id, type, one_liner, profile, entry path, raw path, library, status)  
+
+Id: `C-YYYYMMDD-XX` sequential per day (`01`, `02`, …).  
+Default `profile`: current account’s `profile_id`.
+
+### Also write raw when
+
+- URL worth revisiting, multi-step structure, media, or user says important  
+
+Then:
+
+1. Body: `raw/YYYY-MM-DD-short-slug.md` (authoritative text/structure/media links)  
+2. Row in **`raw/_index.md`** with `capture_id`  
+3. Set inbox row `raw` column to that file  
+
+### Do not
+
+- Use deprecated `inbox/capture.md` (removed)  
+- Silent-write swipe/atoms/claims  
+- Treat capture as a finished run  
+- Reply “saved” without **Capture Card**  
+- Invent facts not in the user message / source  
+
+### entry file (minimal shape)
+
+```markdown
+---
+id: C-YYYYMMDD-XX
+date: YYYY-MM-DD
+type: structure | link | insight | product | other
+profile: default
+---
+
+# C-… — short title
+
+## One-liner
+## User intent / source
+## Key points (our notes)
+## Links (raw if any)
+## Status
+```
+
+### Capture Card (chat)
+
+```text
+✅ Captured [C-YYYYMMDD-XX]
+
+【One-liner】 …
+【Type】 …
+【Profile】 (resolved profile_id)
+【Key points】
+- …
+
+【Stored】
+- inbox/_index: yes
+- entry: inbox/entries/C-….md
+- raw: no | raw/….md + raw/_index
+
+【Refine】
+A) OK as stored
+B) Refine notes
+```
+
+### Library offer (required if type is structure or link)
+
+```text
+【入库？（需点头）】
+1) 不入库 — index library=none
+2) 收成 swipe
+3) 收成 atom (type: hook|reframe|proof|process|cta|other)
+4) 收成 claim
+5) 组合 …
+
+Reply number(s) or 不入库.
+```
+
+On approval: create catalog files per **§ Library ingest**, set inbox `library` column (e.g. `swipe:S2`).  
+Path A (capture): **no** engagement gates on *your* posts.  
+
+Optional: if user also says「同时选题/打分」, run **§ Topics** for this capture.
+
+---
+
+## § Discovery (watchlist signal analysis)
+
+**Role:** Understand external signals faithfully before deciding whether they become our content opportunities. Discovery is not a Topic queue and does not generate publish-ready copy.
+
+### Flow
+
+```text
+watchlist signal → Discovery → Topic (only when explicitly promoted) → Brief
+```
+
+### Discovery rules
+
+1. Create one `D-YYYYMMDD-XX` record for each meaningful watchlist signal; discard noise before analysis.
+2. Write `## Source Digest` as a short, natural-language note:
+   one sentence, what happened, the author's argument, supporting material, native language, and limits.
+3. Do not add our audience, product, thesis, ownership, or candidate Topic to a new Discovery.
+4. Keep source engagement as context only; it is not evidence for our claims.
+5. When a user explicitly promotes a Discovery, create a Topic and run the Topic opportunity evaluation there.
+
+### Discovery status
+
+`unread` → `read` → `saved` | `promoted` | `rejected`
+
+### Discovery Card
+
+```text
+🔎 Discovery D-YYYYMMDD-XX
+【Source】 @handle · url · platform
+【Source digest】 what happened … · author argues … · support …
+【Limits】 what the source does not establish …
+【Status】 unread | read | saved | promoted | rejected
+【Next】 none / revisit / promote to Topic
+```
+
+### Stop rules
+
+- Do not create a Topic merely because a source has high engagement.
+- Do not force every Discovery into the current product wedge.
+- Do not present a source summary as our own claim.
+- Do not generate candidate opportunities inside Discovery.
+- Do not use 5C as the default reading format.
+
+---
+
+## § Topics (Opportunity evaluation)
+
+**Role:** Evaluate **Content Opportunities** — whether a topic deserves to exist as content. Does not replace capture or runs.  
+**Principle:** Topics are not titles. They are opportunities with a context, audience, conflict, insight, and evidence potential.
+
+### Layout
+
+| Path | Role |
+|------|------|
+| `topics/scoring.md` | Scout → opportunity **15/20/25/25/15** (context/audience/conflict/insight/evidence); ready = opp≥**7.5** **and** aud≥**6** |
+| `topics/WATCHLIST.md` | Watchlist router |
+| `topics/watchlists/<profile_id>.md` | Per-profile radar (**scheme B**) |
+| `topics/watchlists/_shared.md` | Shared rows for all profiles (`*`) |
+| `topics/_index.md` | Queue: one row per topic, **`item`** link, **`src_*` engagement** |
+| `topics/items/T-YYYYMMDD-XX.md` | Content Opportunity (structured fields + scores) |
+| `topics/_template-item.md` | Template |
+
+### Defaults
+
+- Max **5** new **`produce`** topics per promotion session; watchlist discovery itself has no Topic output cap
+- Scout first: `purpose` = `discard` | `library` | `produce`; `scout_label` = `demand` | `ally` | `foil` | `craft_only`
+- Opportunity dims (produce only): `context` `audience` `conflict` `insight` `evidence` (0–10 + note each) → `opp_score`
+- `ready` requires **opp_score ≥ 7.5 and audience ≥ 6**
+- `src_likes` / `src_replies` / `src_bookmarks` (+ views if known) + `engagement_at` on **index** when source is external
+- Do **not** fold `src_*` into opp_score; among **ready** only, prefer replies/bookmarks over views
+
+### Signal paths
+
+| User intent | inbox/raw | topics |
+|-------------|-----------|--------|
+| `记一下` + URL | **Yes** (default) | Only if also asked to score as topic |
+| `按 watchlist 选题` | **Default NO** | Run **§ Discovery** first; create Discovery records, not Topics directly |
+| `扫 inbox 选题` | Existing only | Scout + score ≤5 produce; set `capture_id`; skip captures already linked to a T- |
+| `把 C-xxx 做成选题` | Keep | Create T- with `capture_id` (still Scout + opportunity score) |
+| `链接直接选题` | Default NO unless user agrees capture | Create T- + src engagement |
+| `把 T-xx 收进 inbox` | Yes for that source | Backfill `capture_id` on T- |
+
+After a Discovery batch, report **unread / read / saved / promoted / rejected** counts. Topic promotion is explicit and happens after Discovery, not during the scan.
+
+### Create topic steps
+
+**Input gate:** A Topic must have an explicit input. Accept a user idea/brief, a URL, a Capture ID, or a Discovery ID named by the user. If the user only says “generate a topic” without naming a source, ask which input to use; do not select from the Discovery index, watchlist, or inbox by preference.
+
+1. Hard filters (`scoring.md`); else skip.
+2. **Scout gate** — set `purpose` + `scout_label`; if `discard`, do not create T-.
+3. If `produce` (or user forces score):
+   a. Fill **structured opportunity fields**: context (what_changed, why_now), audience (who, current_belief, icp_checks), conflict (old_model, new_reality, belief_gap), insight (potential_insight, why_nonobvious), evidence (available, type, gaps).
+   b. Score all 5 opportunity dims (0–10 + note each) → `opp_score`.
+   c. Write `topics/items/T-….md`.
+4. If `library`: optional light item with `status: library`, or capture offer only.
+5. Add row on `topics/_index.md` (`item` + `src_*` when external).
+6. Output **Opportunity Card** in chat.
+
+**Important:** the structured fields (current_belief, belief_gap, potential_insight) are not decorative — they are the raw material the Brief Engine inherits. If you cannot articulate a belief gap, the conflict score cannot be above 5.
+
+### Opportunity Card
+
+```text
+📋 Opportunity T-YYYYMMDD-XX
+【Opportunity】 … (one sentence — what's the content opportunity)
+【Source】 watchlist @handle (profile … · platform …) | inbox C-… | user | url
+【Scout】 purpose … · label … · note …
+【Src engagement】 likes … · replies … · bookmarks … (@ time)
+
+【Context】 what changed (1 line) → score …
+【Audience】 who + ICP check (1 line) → score …
+【Conflict】 current belief → reality (1 line) → score …
+【Insight】 non-obvious perspective (1 line) → score …
+【Evidence】 available + gaps (1 line) → score …
+【Opp score】 … · audience floor …
+
+【Profile fit】 pillar … · audience pain …
+【Verdict】 ready | backlog | parked | library
+【Weak dims】 (any dim < 6 → name it + suggest: research / reframe / discard)
+【Suggested selection】 swipe/atoms/claim (none OK)
+【Next】 开一单 / 改 angle 重打分 / 研究 evidence / 放弃 / 收进 inbox / 入库
+```
+
+### Topic → run
+
+On `用 T-xxx 开一单`: **ask platform(s)** if unset (multi-select OK). Create **one run per platform**, same `topic_id`, prefill selection per platform, mark topic `promoted` + list run ids on `topics/_index` `run` column.
+
+### Watchlist (scheme B — per profile)
+
+**Layout:** `topics/WATCHLIST.md` (router) + `topics/watchlists/<profile_id>.md` + `topics/watchlists/_shared.md`.
+
+**Resolve profile before scan or add:**
+
+1. User gives **profile_id** → use it.  
+2. User gives **account** → `accounts/_index` → `profile_id`.  
+3. **Neither → ask** (list profiles from `profiles/_index`; do **not** silently default).  
+
+**Load set:** active rows in `_shared.md` ∪ `watchlists/<profile_id>.md`.  
+**Platform filter:** user-specified, else default **`x`**; `全部平台` = all platforms in the set.  
+**Always set `platform` on each row.** Same brand on two networks → two rows.  
+**Shared across all profiles** → `_shared.md` (ids `#S-01…`). Two profiles only → copy row into both files.  
+**Add handle:** write to the chosen profile file (or `_shared` if user says shared / 全 profile).
+
+### Do not
+
+- Auto-dump watchlist scans into inbox
+- Skip Scout gate or opportunity scoring on `produce` topics
+- Skip structured opportunity fields (context/audience/conflict/insight/evidence) on `produce`
+- Skip per-dimension notes on `produce`
+- Exceed 5 **produce** topics per scan
+- Mark `ready` on high `src_*` or high opp_score alone (need audience ≥ 6 too)
+- Score conflict > 5 without articulating the `current_belief` and `belief_gap` in writing
+- Scan wrong network for a row (e.g. X tools on a `tiktok` handle)
+- Silently assume profile when user did not name profile or account
+
+---
+
+## § Swipe / atom library
+
+- **Catalog:** `swipe/_index.md` / `wiki/atoms/_index.md` are the selection entrypoints (Approach B).  
+- **Hit-status:** `trial` | `working` | `dead` (not claim epistemology).  
+- **Hits cache:** index columns `n` `win` `loss` from **our** `published.result`.  
+- **Use history:** `published/_index.md` (columns `swipe` / `atoms`). **Do not** put an Evidence table in swipe/atom files.  
+- **New item:** template + index row; default `trial`, n=win=loss=0.  
+- **From capture / run:** only via **§ Library ingest**. Never silent seed from runs.
+
+---
+
+## § Open a run
+
+### When
+
+User wants a tracked job: "开一单", "open a run", "按完整流程", or a multi-step post with gates.
+
+### Model (one platform = one Run)
+
+- **One folder = one platform draft** (native rewrite).  
+- **Same Topic, different platforms** → **N runs**, shared `topic_id` — **not** one folder with many platform sections.  
+- `idea.md`: `primary_platform` = this run’s platform; `platforms` = **same single value** (column kept for index compat).  
+- `draft.md`: copy for **this platform only**.  
+- Image deck: **`pack.md`** (or `pack-<platform>.md`); pixels under **`exports/<slug>/`**.  
+- Ship → one `published/_index` row for this `run_id`.  
+- Cross-platform progress lives on the **Topic** (`run` column lists all RUN ids), not inside one run.
+
+### Platform gate (required before folder create)
+
+1. User already named platform(s) → use them.  
+2. Else **ask** (multi-select OK). Offer account’s `platforms` from `accounts/_index` (plus others user names).  
+3. **Do not** silently default to `x` when opening a run (default `x` still applies to other flows, e.g. watchlist filter).  
+4. **N platforms selected → N runs** (same topic/goal/pillar; independent selection/draft/ship).  
+5. Slug: include platform when N>1 or always OK — e.g. `2026-08-26-angle-x`, `…-linkedin`.
+
+### Load platform craft (required — every run, every draft stage)
+
+For **each** run, before brief / packet / draft / editor / rubric / pack:
+
+1. Resolve craft file via `platforms/README.md` (`x` → `x.md`, `instagram`/`ig` → `ig.md`, …).  
+2. **Read** that file end-to-end. Apply length, hooks, anti-slop, quality checklist.  
+3. Also load profile `audience.md` **core + this platform’s section**.  
+4. Cite the craft path in `brief.md` / `packet.md` (not hardcoded `platforms/x`).  
+5. **Do not** paste another platform’s copy unchanged. Sibling runs each re-read their own craft file.  
+6. If the craft file is missing → ask user / create stub; **do not draft** from memory.
+
+### State machine (do not skip)
+
+```text
+RUN_OPEN → BRIEF_WAIT → DRAFTED → EDIT_WAIT
+  → SCORED → SHIP|HOLD → READY_TO_POST
+  → PUBLISHED feedback
+```
+
+Also mirror high-level status on **`runs/_index.md`**:  
+`open` → `brief` → `draft` → `scored` → `ready` → `published` | `hold` | `killed`.  
+(`partial` = **legacy** only; do not use for new multi-platform planning.)
+
+Human gates (chat only):
+
+1. **简报 OK** / Brief OK  
+2. **编辑 OK** / Editor OK  
+3. After SHIP: user posts **this** platform, then URL + result → feedback + published row  
+4. If images: **图 OK** after render (when skill available)
+
+### Folder setup
+
+1. Resolve platform list (gate above). For **each** platform:  
+2. **Load** `platforms/<platform>.md` (required; see above).  
+3. Create `runs/YYYY-MM-DD-short-slug[-platform]/` from `runs/_template/`.  
+4. Fill `idea.md` (pillar, **platform**, **selection ids**, optional `topic_id`).  
+5. **Append row** on `runs/_index.md` (`status=open`).  
+6. If from **T-…**: prefill from topic; mark topic `promoted`; append all new run ids to `topics/_index` `run`.  
+7. Never leave production only in chat when in run mode.
+
+### Selection inside a run (required step; **none allowed**)
+
+Before brief body, evaluate catalogs (Approach B + **profile + this run’s platform**).
+
+Filter every catalog row:
+
+- `platforms` includes this run’s platform (or `*`)  
+- `profiles` is `*` **or** includes current `profile_id`  
+- Swipe/atom: drop `dead`. Claim: drop `retired`.  
+
+Rank:
+
+- Swipe/atom: `working` before `trial`; then higher `win/n` (if n=0, treat as unranked trial).  
+- Claim: **do not** rank by `hypothesis` vs `supported`. Rank remaining by `win/n`. `hypothesis` is valid to pick. `weakened` only if the user wants that tension.
+
+Then: swipe 0–1 → atoms by role → claim 0–1. **`none` always allowed** with why.  
+Sibling runs for the same topic may pick different swipe/atoms (native rewrite).
+
+**Write the pick into `idea.md` (this is the citation).** Chat card alone does not count.
+
+| Field | Required form |
+|-------|----------------|
+| `swipe_id` | `none` or `S1` (one id) |
+| `atoms` | `none` or `hook:A-… cta:A-…` (real `A-` ids) |
+| `claim_id` | `none` or catalog `C-…` — **not** the `hypothesis:` one-off line |
+
+If the user later changes the pick, **overwrite** those three fields. Do **not** bump catalog `n`/`win`/`loss` here — that is ship-only (§ Catalog evidence).
+
+**引用** = how many **runs** currently name this id in `idea.md` (killed runs still count unless selection was cleared). The workbench derives it by scanning runs; it is not a column the agent increments. Observatory numbers update on `workbench/build.py`, not at the moment of chat.
+
+### Brief Engine (decision tree → quality gate → stop rule)
+
+The brief inherits reasoning from the Topic Opportunity and focuses on **what to say and how**.
+
+**Step 1 — Inherit from Topic Opportunity**
+
+When the run is from T-xxx, copy the topic's structured fields into `brief.md` `## Inherited`:
+- `context` (from topic's what_changed / why_now)
+- `audience` (from topic's who / current_belief)
+- `conflict` (from topic's belief_gap)
+- `insight` (from topic's potential_insight)
+
+Sharpen these for the specific platform/angle — do not re-derive from scratch. If T-xxx has no structured fields (legacy topic), derive them before proceeding.
+
+**Step 2 — Walk the decision tree**
+
+```text
+1. Read inherited context/conflict/insight
+2. Can you state a thesis in one sentence?
+   → No → reframe or stop
+3. Can you name specific evidence?
+   → No evidence at all → stop, suggest research
+   → Weak but mechanism is provable → continue
+4. What's the angle? (How does this conflict become content?)
+5. What changes for the reader after consuming this?
+   → "They'll learn something" is not specific enough
+   → Name the belief before and after
+6. → Fill brief → Self-score → Gate
+```
+
+**Step 3 — Fill brief.md**
+
+Fill Claim (core_thesis + supporting_claims), Evidence (sources, type, quality), Communication (angle, format, hook, structure), Outcome (belief, action, emotion), Constraints, Selection.
+
+**Step 4 — Brief quality gate (self-scoring)**
+
+Score `brief.md` quality gate (6 × 1–5, max 30):
+
+| # | Gate | Question |
+|---|------|----------|
+| 1 | Thesis clarity | Can the thesis be one sentence a stranger understands? |
+| 2 | Evidence strength | Is there at least one specific, named source or mechanism? |
+| 3 | Angle originality | Is this angle different from what's already in the conversation? |
+| 4 | Structure coherence | Does the narrative flow from hook → conflict → thesis → proof → outcome? |
+| 5 | Outcome specificity | Can you name the belief change in before/after form? |
+| 6 | Tension preservation | Does the brief still carry the topic's conflict, or did it dilute? |
+
+| Result | Action |
+|--------|--------|
+| **≥ 24** (PASS) | Present Brief Card to human for OK |
+| **20–23** (REVISE) | Flag weak dimensions, self-revise once, re-score |
+| **< 20** (STOP) | Trigger stop rule |
+
+Update `brief.md` frontmatter: `gate_score`, `gate_pass`.
+
+**Step 5 — Stop rule**
+
+When the brief cannot pass the quality gate:
+
+```text
+⚠️ Weak Brief — RUN-…
+【Failing gates】 gate … score … · gate … score …
+【Reason】 one line why
+【Suggestion】 reframe angle / research evidence / kill topic
+Reply: 用 reframe / 研究 / 换 topic / 强制继续
+```
+
+The agent may reject a brief. "强制继续" overrides the gate (user decision). The stop rule is not a hard block — it is structured feedback that makes weakness visible.
+
+### Packet
+
+After brief OK: short `packet.md` — pillar slice + voice + audience (**core + this platform**) + **craft notes distilled from the loaded `platforms/<platform>.md`** + **thesis sentence from brief**.
+
+### Writer → Editor
+
+- Writer **re-reads** `platforms/<platform>.md`, then fills `draft.md` (**this platform only**) + full text in chat.  
+- Editor → `editor.md`; check against that craft file’s quality bar; REVISE until APPROVE or user override.
+
+### Rubric gate
+
+Score `rubric.md` (6 × 0–2, max 12).  
+
+- **≥ 8 → SHIP** — 发布包 in chat; `runs/_index` → `ready`.  
+- **< 8 → HOLD**.  
+- Carousel/slideshow platforms: after SHIP → **§ PagePack + render**.  
+- Pure X text: no multi-page pack.
+
+### Required chat cards (run mode)
+
+**Open card** (one card per run; if N>1, list all opened)
+
+```text
+✅ Run opened  RUN-YYYYMMDD-slug
+【Account】 (from accounts/_index)
+【Profile】 (resolved profile_id)
+【Voice】 profile | path
+【Platform】 x | linkedin | …  → craft `platforms/<file>.md` loaded
+【Topic】 none | T-…  (sibling runs: RUN-…, RUN-…)
+【Goal】 …
+【Pillar】 P?
+【Selection】 (profile + platform; drop dead/retired; prefer working; none OK)
+  swipe: S? | none — why
+  atoms: …
+  claim: C? | none — why
+【Index】 runs/_index updated
+【State】 open → next: brief
+```
+
+**Brief card** (after quality gate passes)
+
+```text
+📋 Brief  RUN-…
+【Inherited from T-…】
+  Context: …
+  Conflict: current belief → reality
+  Insight: …
+
+【Thesis】 one sentence
+【Evidence】 type · quality (weak/moderate/strong)
+【Angle】 …
+【Outcome】 belief / action / emotion
+【Gate】 Thesis … · Evi … · Angle … · Struct … · Outcome … · Tension … = …/30 ✓
+【Selection】 swipe … · atoms … · claim …
+Reply 简报 OK / 改 …
+```
+
+**Draft / Editor / Rubric** — as before; draft is single-platform and must follow the loaded craft file.
+
+**Publish pack** (SHIP)
+
+```text
+📦 READY_TO_POST  RUN-…
+【Platform】 x | …
+【Copy】 …
+【Pack】 n/a | ready | blocked
+【After post】 URL + win/flat/loss → feedback + published row
+```
+
+### Feedback
+
+`feedback.md` + **§ Published index** on ship.  
+Update `runs/_index` → `published`.  
+Then **§ Catalog evidence**. Never auto-`supported` / auto-`working`; never silent catalog ingest.
+
+### § Published index (Step 6)
+
+On confirmed ship for this run:
+
+1. Row on **`published/_index.md`** (date, **platform**, url, run, pillar, swipe, atoms, claim, result, notes).  
+2. Optional stub.  
+3. `runs/_index` status → `published`.  
+4. **§ Catalog evidence** for every cited swipe / atom / claim id (skip `none`).  
+5. Weekly 复盘: start from **published** `_index`. Same topic’s other platforms = other runs / other published rows.
+
+### § Catalog evidence (required on ship)
+
+Hits cache lives on the **index row** (`n` `win` `loss`). This is **not** 引用. 引用 is run selection in `idea.md` (already true once the run exists). `n`/`win`/`loss` only move **on ship**.
+
+**Swipe / atom — no Evidence file.**  
+The ledger is the **published row** (`swipe` / `atoms` / `result` / `notes`). Why a post flopped stays in `feedback.md` or that published `notes` cell. Do **not** add `## Evidence` to swipe/atom bodies.
+
+**Claim — Evidence table in the claim file.**  
+That table answers “does the sentence still hold?” (`vs claim`), which is not `result`.
+
+For each cited id:
+
+1. `n += 1`  
+2. `win += 1` if `result=win`; `loss += 1` if `result=loss`; `flat`/`unknown` bump `n` only.  
+3. Sync file frontmatter `status` if you change index status (only after user 批准).  
+4. **Claim only:** append the run to the claim file **Evidence** table (`result` + **vs claim**).  
+
+Propose status change (user must 批准; never auto):
+
+| Object | Propose | Gate |
+|--------|---------|------|
+| swipe / atom | `trial` → `working` | `n≥2` AND `win≥2` AND `win>loss` |
+| swipe / atom | → `dead` | user says stop, **or** `n≥3` AND `win=0` AND `loss≥2` |
+| claim | → `supported` | ≥3 runs **support the proposition** (not merely `result=win`) AND user 批准 |
+| claim | → `weakened` / `retired` | user 批准; hits do not flip epistemology |
+
+Single win never promotes. n=1 is recorded, not celebrated.
+---
+
+## § PagePack + render (Step 7)
+
+Bridge to **easysociable-slideshow** (when skill exists). Vault owns pack slots; product owns templates + pixels.
+
+### When
+
+Planned platform needs multi-page images (linkedin / ig / tiktok), not pure X text.
+
+### Preconditions
+
+1. Run exists; brief OK; editor OK; **rubric SHIP**.  
+2. Per-platform `page_count` / `deck_hint` sensible.  
+3. Prefer swipe with page roles; else explicit roles from brief.
+
+### Agent steps
+
+1. Write **`runs/<slug>/pack.md`** (from `_template/pack.md`; `pack-<platform>.md` OK):  
+   - this run is already one platform  
+   - slots, language, pillar, swipe_id, claim_id  
+2. **Pack Card** in chat — user may `改 pack：…`.  
+3. **Render:**  
+   - If slideshow skill available: render → files under **`exports/<slug>/`** → set pack `job_id`, `export_path`, status `done`; set `runs/_index` `pack`/`export` to yes.  
+   - If **not** available: status `not_requested` / blocked — **do not fake images**.  
+4. User **图 OK** or revise slots and re-render.
+
+### Pack Card
+
+```text
+📑 Pack  RUN-…  platform=linkedin|tiktok|ig
+【Pages】 N
+【deck_hint】 …
+【Outline】 P1… Pn
+【export_path】 exports/<slug>/
+【Render】 done | blocked (no skill) | failed
+```
+
+### Do not
+
+- Cram multiple platforms into one run / one pack  
+- Pack for pure X text  
+- Private template AST — public slots only  
+- Invent metrics in slots  
+- Skip SHIP before quality pack  
+- Store binaries inside `runs/` (use **exports/**)  
+
+---
+
+## § Library ingest — swipe / atoms / claims (strict)
+
+**Runs never auto-enter catalogs.** Ship → feedback + published + **§ Catalog evidence** if ids already used. New catalog rows still need user approval.
+
+### Two ingest paths
+
+| Path | When | Your-post metrics? |
+|------|------|--------------------|
+| **A. Capture offer** | structure/link capture → user picks 入库 options | **No** (source heat is prior, not our n/win) |
+| **B. Post-publish** | In `published/_index` + gates → Agent proposes → user approves | **Yes** |
+
+### Hard bans
+
+| Forbidden |
+|-----------|
+| Unpublished run → new swipe/atom/claim |
+| Silent seed while scaffolding |
+| Catalog write without user approval |
+| Single win → claim `supported` or swipe/atom `working` |
+| Mixing claim epistemology with swipe/atom hit-status |
+| `## Evidence` table in a swipe or atom file |
+| Auto-changing `status` when bumping `n`/`win`/`loss` |
+
+### Engagement gates (path B only)
+
+At least one of: **replies ≥ 2** | bookmarks ≥ 5 | likes ≥ 10 | user override “这条可收”.  
+These gates are for **new ingest**, not for promoting `working` (that uses our `win`/`loss` counts).
+
+### After user approval
+
+| Ask | Create |
+|-----|--------|
+| 收成 swipe | `swipe/` + index, `trial`, n=win=loss=0 |
+| 收成 atom | `wiki/atoms/`, set `type`, `trial`, n=win=loss=0 |
+| 收成 claim | `wiki/claims/`, `hypothesis`, n=win=loss=0 |
+
+### § Claims promote
+
+- Epistemic Evidence on existing claims when used (separate from hits).  
+- `supported` only if ≥3 runs support the **proposition** **and** user 批准.  
+- A `win` that does not actually support the sentence does **not** count toward `supported`.
+
+### 复盘 card
+
+```text
+📈 Review
+【Run】 …
+【Result】 …
+【Hits updated】 swipe S? n/win/loss … · atoms … · claim …
+【Hit-status propose】 none | S?/A-… trial→working | →dead
+【Path B ingest】 not eligible | eligible → propose swipe/atom/claim
+【Claim promote】 none | suggest C-… hypothesis→supported|weakened
+Reply 批准 … / 否决 / 再观察
+```
+
+### Self-check (Capture)
+
+- [ ] `inbox/_index.md` row + `inbox/entries/C-….md`  
+- [ ] `raw/` file + `raw/_index.md` only if criteria matched  
+- [ ] Capture Card + library offer when structure/link
+
+### Self-check (Topic / Opportunity)
+
+- [ ] Structured opportunity fields filled (context/audience/conflict/insight/evidence) on `produce`
+- [ ] `current_belief` and `belief_gap` articulated in writing (not just a score)
+- [ ] `potential_insight` stated — what's non-obvious
+- [ ] 5-dim scores with per-dim notes
+- [ ] `opp_score` weighted correctly; `audience` floor checked
+- [ ] Weak dims (< 6) flagged with suggestion
+- [ ] Opportunity Card output in chat
+
+### Self-check (Run)
+
+- [ ] Folder exists under `runs/` with template files filled  
+- [ ] `runs/_index.md` row created/updated  
+- [ ] Platform gate done; one run per selected platform; `primary_platform` = that platform  
+- [ ] `platforms/<platform>.md` **read** before brief/draft/editor; cited in brief + packet
+- [ ] Brief `## Inherited` populated from Topic Opportunity (not re-derived)
+- [ ] Brief decision tree walked; quality gate scored (6 × 1–5)
+- [ ] Brief gate ≥ 24 before presenting to human (or stop rule triggered)
+- [ ] No skip of brief / editor / rubric when user asked for full run  
+- [ ] HOLD never presented as final publish pack  
+- [ ] Image pack only for this run’s platform; exports under `exports/<slug>/`  
+- [ ] Each ship → published row; this run → `published`  
+- [ ] `idea.md` has `swipe_id` / `atoms` / `claim_id` matching the Open card (not only chat)  
+- [ ] Cited swipe/atom/claim index `n`/`win`/`loss` bumped **on ship only**; status unchanged unless user 批准  
+- [ ] Swipe/atom: no Evidence section; claim Evidence table updated if claim used  
+- [ ] Selection dropped `dead`/`retired`; did not treat claim `supported` as a hit-rank  
+- [ ] Same topic multi-platform = sibling runs sharing `topic_id`, not one multi-section folder
+- [ ] Packet includes thesis sentence from brief
+- [ ] No invented metrics  
+- [ ] No cross-platform paste without re-applying that platform’s craft file