AK體-基於Threads演算法的優化skill

ak-threads-booster remains the internal skill id for routing and installation.

Use this as the single entry point for the AK Threads workflow.

This skill is for creators who want to pick stronger topics, write posts with clearer upside, and improve over time using their own Threads history.

It is not a viral-post guarantee engine. It is a decision system:

find topic angles with real demand
avoid obvious repetition and red lines
draft from the user's own voice and history
review outcomes and feed the learning back into the tracker

Intent Routing

Classify the user's request first, then open and follow one primary module unless the task clearly needs a short sequence.

Each sub-skill is located via Glob **/skills/<name>/SKILL.md so resolution works regardless of where the plugin is installed. Do not assume an absolute path or environment variable.

Setup / import / initialize / backfill history -> Glob **/skills/setup/SKILL.md
Refresh tracker / update metrics / scrape own profile -> Glob **/skills/refresh/SKILL.md
Analyze a finished post / inspect / AK-review / optimize / improve / 檢查 / 優化 / 診斷 a post the user has written -> Glob **/skills/analyze/SKILL.md
Draft / write from a topic / 起草 / 寫文 (user has not written anything yet; generate from a topic) -> Glob **/skills/draft/SKILL.md
Predict likely 24-hour performance / expectation check -> Glob **/skills/predict/SKILL.md
Review actual post performance / compare against prediction -> Glob **/skills/review/SKILL.md
Mine next topics / topic suggestions / 選題 -> Glob **/skills/topics/SKILL.md
Build brand voice / voice analysis -> Glob **/skills/voice/SKILL.md
Local visual panel / dashboard / data cockpit / open panel -> Glob **/skills/panel/SKILL.md
Check skill updates / update AK-Threads-Booster / install weekly auto-update / 更新 skill / 自動更新 -> Glob **/skills/update/SKILL.md
Optimize the skill itself / compound pass / 優化skill / 自我優化 / 閉環 (turn threads_skill_learnings.log misses into rule edits) -> Glob **/skills/optimize/SKILL.md

Routing Rules

Do not answer from this file alone when a module exists for the request.
Open the matched module and follow its workflow.
If the user asks for a combined task, use the smallest valid sequence:
- first-use account workflow -> setup, then voice if needed
- write from historical data -> setup or voice first if data is missing, then draft
- post decision flow -> analyze, then predict if the user asks for a range
- post-publication learning flow -> review
Keep outputs grounded in the user's own tracker whenever available.
If threads_daily_tracker.json is missing, do not pretend the work is data-backed. Ask for fallback history or use the setup path.
Analyze vs Draft routing discipline. If the user pastes their own text — no matter whether they say "analyze", "check", "optimize", "improve", "幫我看一下", "幫我優化" — route to /analyze. /analyze gives pointed diagnosis and preserves the user's format; it does not rewrite the post. Route to /draft only when the user has no existing text and wants something generated from a topic.
Brand voice scope. brand_voice.md is a composition driver only in /draft. Every other module treats it as observation-only — for flagging drift, never for rewriting the user's submission toward a voice template.
Proactive update offer. On first setup, installation help, or any repo/skill maintenance conversation, proactively tell the user AK體 can install an opt-in weekly GitHub update checker. Ask whether they want to enable it. Do not install it by default, and do not interrupt normal content tasks like /draft or /analyze with this offer unless the user asks about maintenance.

Working Data

Look in the working directory for:

threads_daily_tracker.json - canonical machine-readable tracker
style_guide.md - produced by /setup
concept_library.md - produced by /setup
brand_voice.md - produced by /voice, referenced by /draft
posts_by_date.md - human-readable archive
posts_by_topic.md - human-readable topic index
comments.md - human-readable flat comment log
threads_freshness.log - audit log for /draft and /topics freshness gates, read by /review
threads_refresh.log - audit log for /refresh runs, read by /review

Low-token runtime also looks for derived compiled memory in compiled/:

compiled/account_wiki.md
compiled/account_state.md
compiled/personal_signal_memory.md
compiled/next_move_queue.md
compiled/post_feature_index.jsonl
compiled/cluster_wiki.json
compiled/exemplar_bank.md
compiled/recent_window.md
compiled/voice_fingerprint.md
compiled/voice_fingerprint.json

These files are runtime caches only. threads_daily_tracker.json remains the source of truth. If compiled memory is missing, stale, or contradicts the tracker, fall back to the tracker and recommend rebuilding compiled memory.

If legacy Chinese companion filenames already exist, treat them as equivalent companion files instead of forcing a rename.

If only the tracker exists, continue in tracker-only fallback mode when the chosen module allows it. If the tracker is missing, do not pretend the work is data-backed - ask for fallback history or use /setup.

Tools Surface

This main SKILL.md declares only the read-only tools it actually uses (Read, Glob, Grep). Each sub-skill declares its own allowed-tools in its frontmatter; some require more:

/draft adds Write, WebSearch, WebFetch
/review adds Write, Edit
/voice, /setup, /refresh each extend the surface as needed
/update uses git through shell commands and may create an opt-in Codex automation when the user explicitly accepts weekly auto-update

When auditing permissions, inspect the union of all sub-skill frontmatters, not just this file.

Persistent-State Policy

Any sub-skill that writes to threads_daily_tracker.json, style_guide.md, concept_library.md, brand_voice.md, or threads_booster_config.json must follow templates/FAILSAFE.md (backup + atomic rename + keep last 5 backups). Append-only logs (threads_freshness.log, threads_refresh.log, threads_skill_learnings.log) follow the append-only rules in the same file.

Compiled memory files under compiled/ are rebuilt views, not hand-edited state. Rebuild them with scripts/build_compiled_memory.py after tracker-changing runs.

Shared Knowledge

Red-line (R) and signal (S) definitions live in knowledge/_shared/red-lines.md — the single source of truth for both /analyze and /draft. Do not inline R-lists in sub-skill SKILL.md files.

Next-post direction lives in knowledge/_shared/next-move-engine.md. Any "next move" recommendation must pass red-line filtering first, name the S signal it is trying to strengthen, and avoid becoming a formula bank or case-study imitation.

Runtime depth, compiled-memory behavior, and output-mode defaults live in knowledge/_shared/runtime-budget.md. In lite and standard, use knowledge/cards/* before full knowledge/*.md files.

If runtime.token_mode is absent or "ask", ask the user to choose 低 token 版 or 高 token 版 before heavy reading. The question must clearly state the tradeoff: low token is faster and cheaper but less detailed; high token is deeper but slower and more expensive.

Compound loop schema — threads_skill_learnings.log — is defined in knowledge/_shared/compound-log-format.md. /review writes misses to it; /optimize reads it, proposes rule changes, and appends supersedes entries when the user approves edits. No other sub-skill touches the log.

Precedence and Conflicts

When guidance conflicts, use this order:

This main SKILL.md — routing and global discipline
Sub-skill skills/<name>/SKILL.md — module-specific workflow
knowledge/_shared/*.md — definitions referenced by

ak-threads-booster