tonydzi INTENTION_LANE_RUN
/intention — the day's intentions → intention posts (draft-first)
GOAL: turn the owner's INTENTIONS of today (every pain / "how do I do X" question / "I need someone who…" from the day's sessions) into 2-3+ honest intention posts with an explicit ASK, so the audience returns ALPHA (a solution, a warm intro, a warning, a co-author). Draft-first: nothing is published, the drafts go to the owner over Telegram.
Paths and engine
- Miner + store:
python $IMPORTS_ROOT/content-factory/intention/intention_mine.py <cmd>(0 tokens). - Store:
intentions.db(never loses anything; deduped by content hash). - Candidates:
candidates\cand-<DAY>.md. Drafts:drafts\intentions-<DAY>.md. - Decision/canon:
$OBSIDIAN_VAULT/02-Decisions/decision-intention-lane-content-factory-2026-07-02.md. - Telegram Saved Messages (the primary work account): numeric chat_id
<YOUR_CHAT_ID>(NOT "me", no parse_mode). - ⚠️ DELIVERY FALLBACK (when the Telegram MCP is unavailable/connecting — it drops out routinely): the Telethon rail, MCP-independent. Text:
TG_BUS_GROUP=<YOUR_CHAT_ID> python $USERPROFILE/.claude/scripts/tg_bus_send.py --raw "…"; file:… --raw --file "<path>" "<caption>". Verified 2026-07-02.
Channel rules (from Deep Research 2026-07-02) — IMPORTANT
- Channels for alpha-seeking ask posts: Telegram (RU) · X (EN, #buildinpublic) · Indie Hackers (EN) · Ask HN (EN, "Ask HN:"). Optional: LinkedIn (business tone).
- ⛔ Do NOT post asks on long-form article platforms or in the Facebook body — the ask format does not work there (an article platform is for articles, not quick answers). Those platforms are separately the channel for longreads/dev-logs, not intentions.
- The platform decides the language: Telegram=RU, X/IH/HN=EN. No forced bilingual output.
The intention-post format (established)
Context (what I'm doing) → what failed / what I realised → an explicit, concrete ask at the end. Honest, vulnerable tone, no marketing gloss. A concrete request ("do I pick A or B?", "I need an expert on X, who has done it?") beats a vague one. Describe IN DETAIL how the owner got here and why (his own rule). Length: X = teaser (240-370), Telegram/IH = medium, Ask HN = "Ask HN: " + a paragraph. 5 templates (need an expert / A-vs-B choice / hit a wall / realised I was wrong / general poll) — in the decision memo §A4 / the DR.
PRIVACY (hard)
No real names / @handles of third parties, no exact amounts (revenue/rounds/other people's deals), no secrets or paths with personal ids. Generalise ("a large lead", "$X"). Privacy beats completeness.
MODEL
The post is the owner's authorial voice → written by the top-tier model. If the current run is NOT on it — delegate THE WRITING ITSELF to a subagent with model:'opus' (Agent tool), passing it the candidates + the format + the privacy rules. Clustering / candidate selection can run on a cheap model.
STEPS
- DAY = today's local date YYYY-MM-DD (print ASCII only).
- Mining:
intention_mine.py mine <DAY>. Read the summary (CANDIDATES) and the filecandidates\cand-<DAY>.md. - If CANDIDATES == 0: do not invent anything. Tell the owner over Telegram "📭 : no intentions mined today" and finish.
- CLUSTERING (the judge): assemble DISTINCT intentions out of the raw candidates (one pain = one intention; merge repeats of the same topic). For each: a short title · pain (the pain/question) · journey (how he got here + why, in detail) · ask (the explicit request). Discard pure noise (non-intentions, housekeeping).
- Store: for every distinct intention —
intention_mine.py add --day <DAY> --title "…" --pain "…" --journey "…" --ask "…" --sessions "sid1,sid2". Dedup will drop the ones already stored (that is fine — we don't post twice). - SELECTION for posting: take the 2-3 most alpha-worthy intentions OF TODAY (a strong concrete ask, freshness). More if the day was rich.
- WRITING (top model): for each selected intention, write posts for the suitable channels (Telegram-RU + X-EN at minimum; add IH/Ask HN where the ask fits). Format + privacy as above. Vary the type (request / choice / realised-I-was-wrong).
- Save every draft into
drafts\intentions-<DAY>.md(UTF-8, no BOM; APPEND ONLY, never overwrite someone else's edits — if the file exists, append a section stamped with the run time). Mark the intentions:intention_mine.py mark --id N --status drafted. - Send to Telegram (chat_id
<YOUR_CHAT_ID>, no parse_mode) a header "🎯 Intention drafts for (edit and post them yourself):" + per intention a block "— —" and the per-channel versions. Longer than ~4000 → split on paragraph boundaries into "Part N/M". - Report to the owner: how many intentions were mined / stored / drafted, which channels, what was deferred. Finish with "what's next".
The bridge into an episode (intention → longread/dev-log on GitHub)
A strong intention is worth expanding into a "reality show episode": intention_mine.py episode --id <N> creates an episode bundle (9 drafts across the tiers of §7.2, including the longread tier — the article platform here is the LONGREAD channel, not an ask channel) + intention-seed.md carrying the context (pain/journey/ask) for the writer. Then write the drafts (voice = top model) in episodes/<slug>/, then python episode_adapter.py check --slug <slug>. The intention is marked status=episode. This is the seam with /episode — don't duplicate it, call it.
The alpha loop (as responses arrive)
The owner says " replied to the post about X" → intention_mine.py respond --id <N> --who "<name/@>" --channel <channel> --note "<gist>". Check the queue: responders --pending. Promotion into the CRM (leads.db) is a separate verifiable step through the normal lead flow (never write into the production CRM blind).
Boundaries
Draft-first is HARD — nothing goes outbound without the owner's explicit "publish". Don't post asks on article platforms. AK-47: don't spawn platforms beyond the decided set. Tier-2 (outbound/money) is not waived.
Like this skill? It is one of 100 in second-brain-starter-kit: the second brain we built for ourselves and run every day at Palo Alto AI Research Lab. Install the whole set with npx skills add tonydzi/second-brain-starter-kit. Everything is open source and free, so take what you need.
Flagships worth a look on their own: secondop-panel (a second opinion from a panel of external models), claude-memory-tidy (stop your agent's memory from rotting), telegram-mcp-kit (your own Telegram over MCP in about 15 minutes).
Author: Anton Dziatkovskii, Palo Alto AI Research Lab. Telegram @tonydzi - WhatsApp +1 341 222 9178 - X @Tony_Stef_
Engineers: want to test-drive this setup? Message me. I hand out free starter seeds to engineers who test and report back, and custom skill requests are welcome.