# AI Hot Brief Telegram Public > Build a Chinese AI news digest for Telegram with a public-safe two-stage workflow. Use when Codex needs to run a bundled Stage 1 collector, do a mandatory second live-web pass, merge both stages into one candidate pool, rank 5 hot items plus 5 watch items, and send the final digest with the bundled Telegram sender helpers from any writable workspace. - Skill: `godgod126/ai-hot-brief-telegram-public` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add godgod126/ai-hot-brief-telegram-public` - Raw SKILL.md: https://api.skillmd.com/api/skills/godgod126/ai-hot-brief-telegram-public/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: godgod126 (https://skillmd.com/u/godgod126) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/godgod126/ai-hot-brief-telegram-public --- # AI Hot Brief Telegram Public Run this skill when the goal is to produce a Chinese AI digest for a general audience and send it to Telegram from any writable workspace. ## Workspace Compatibility Before running, confirm the current workspace provides: - write access, because Stage 1 creates `output/briefings/` - Python 3.10+ with the packages listed in `requirements.txt` - `TELEGRAM_BOT_TOKEN` and `TELEGRAM_CHAT_ID` for the final send step, either as environment variables or via a workspace `.env` file The skill no longer depends on a project-local `main.py` or `scripts/send_latest_brief.ps1`. The bundled sender script is the default send path for public users. ## Workflow 1. Work from the current workspace. Stage 1 writes candidate-pool files into `output/briefings/`. 2. Treat this as a strict two-stage workflow: - Stage 1: run the bundled `scripts/build_brief.py` with `--skip-send`. The script only collects and organizes a candidate pool. It does not decide the final 5 + 5 and it must not send. - Stage 2: Codex must do another live web collection pass, add more candidates into the same pool, and only after that perform the final ranking. 3. Treat `--runtime-budget-seconds` as a hard timeout ceiling, not a fixed collection window. The script should finish as soon as the planned source batches are scanned. 4. Stage 2 is mandatory. Never skip it. 5. The automation-safe path must avoid browser MCP tools. Use only: - the Stage 1 script output - native live web search - direct webpage reading via non-browser web access 6. During both Stage 1 and Stage 2 collection, only add to the candidate pool. Do not decide the final 5 + 5 until collection is complete. 7. After both collection stages are done, rank the full combined pool with: - quantitative signals from `latest.json` - Codex qualitative judgment about public interest, freshness, spread, and mainstream relevance 8. Ensure the final brief has exactly 2 sections: - `过去 2 小时最值得发的 5 条` - `接下来可能爆的 5 条` 9. Ensure every item in both sections contains exactly `标题`、`时间`、`简介`、`来源`、`链接`. 10. Keep the output in Chinese except for necessary product names, company names, platform names, and a few unavoidable AI terms. 11. Only send after Stage 2 final ranking is complete and `latest.txt` has been rewritten as the final brief. 12. The final brief must be mostly Chinese. An English draft is invalid and must be translated before sending. ## Stage 1 Command Run this first to generate the Stage 1 candidate pool. Resolve `` to the installed root of this skill, not the workspace root: ```powershell python "/scripts/build_brief.py" --cwd "." --window-minutes 120 --runtime-budget-seconds 600 --x-budget-seconds 180 --hot-count 5 --watch-count 5 --skip-send ``` ## Stage 2 Collection And Ranking After Stage 1: 1. Read `output/briefings/latest.json` first. 2. Treat `candidate_pool` as the current combined pool baseline, not as the final answer. 3. Always do a second live-web collection pass before ranking: - search the web again for fresher coverage - inspect the pages listed in `stage2_web_source_hints` - open and read direct story pages through native web access when needed 4. Check high-value pages such as: - Techmeme - Wired AI - CNBC AI - Financial Times AI - Product Hunt AI 5. Add Stage 2 findings into the same candidate pool. Do not bypass the pool and write the final 5 + 5 directly. 6. Only after the pool is complete, rank it with both: - quantitative fields already computed in `latest.json`, such as `review_score`, `heat_score`, `breakout_score`, `source_count`, and `observed_sources` - qualitative judgment about whether a story is actually broad, timely, and interesting for ordinary readers 7. Then write the final selections into: - `output/briefings/latest.txt` - `output/briefings/latest.json` - use UTF-8-safe file editing only - prefer direct file editing tools such as `apply_patch` - do not use shell redirection or PowerShell `Set-Content` for the final Chinese brief text 8. Before sending, update `latest.json` so it reflects: - the final combined candidate pool - final `hot_items` - final `watch_items` - `final_selection_ready = true` ## Send Command After Stage 2 is complete, send the final brief: ```powershell python "/scripts/send_brief.py" --message-file "./output/briefings/latest.txt" ``` ## Output Rules - The script writes: - `output/briefings/latest.txt` - `output/briefings/latest.json` - Stage 1 output is only a candidate-pool note. - Stage 2 must rewrite `latest.txt` into the final brief before sending. - Stage 2 must rewrite the final brief in Chinese, not English. - Do not turn the brief into a technical digest. - Do not fill the brief with papers, benchmarks, parameter counts, or low-level infra unless they have already become broad public stories. - Cross-run repeats are allowed. Only dedupe inside the current run. - `简介` must explain what the news actually says in 1 to 2 short sentences, not generic trend commentary. ## Source And Style References - Read [references/source-policy.md](references/source-policy.md) for source priority, collection rules, and final ranking rules. - Read [references/output-style.md](references/output-style.md) for the exact Chinese output format.