tonydzi/second-brain-starter-kit/alfa-search-recall-deepresearch

🅰️ Alpha Protocol — Recall + Deep Research before deciding

🧒 When reporting to the operator: end with a child-simple "In plain words" recap (his standing request; reports TO the operator only).

Binding rule (must): For ANY new strategic function, product, business model, tokenomics, AI-functionality, GTM, market, investment hypothesis, or architecture decision — it is FORBIDDEN to go to implementation on recall alone. Do Recall → Gap → Deep Research → Synthesis → Decision Memo first. Recall is necessary but NOT sufficient for strategic work. Canon: vault note protocol-alpha-protocol-recall-plus-deep-research.

When this fires

• The operator writes the trigger: `R+DR` (= `RDR` — +/space & case don't matter), `alpha protocol` (or /alpha).
• OR you are about to start a Level-2 task (below) — invoke this protocol proactively, don't wait for the trigger.

Levels (size the response to the task)

• L0 — Quick (answer a question, fix a bug, tiny tweak): plain recall is enough. No DR.
• L1 — Recall (new module / hypothesis / feature / market): recall memory + internal docs + past research + form hypotheses. Do NOT decide immediately. Usually no external DR unless it turns strategic.
• L2 — Recall + DR (strategic — see binding rule): run the full flow below.

★ Proactive multi-agent reflex (EVERY task, not just strategic) — set 2026-06-25

The user FORGETS whether they need agents — so YOU remember and propose, reflexively, after RECALL on any task. Canon: vault reglament-proaktivno-predlagay-agentov + memory multi-agent-offer-reflex; tool-choice canon = decision-adopt-agent-teams-scoped.

1. Cheap-first: did SQL/grep/RAG already answer it? → done, no agents.
2. Type = Decision · Comparison · Analysis · Research-synthesis where several INDEPENDENT lenses materially improve the answer (inclusion test: will one lens's finding redirect another before both finish?)? NO (import/fix/ops/mechanical/trivial) → single agent, stay silent about agents. YES → multi-agent fits → fork:

• AUTO-RUN (announce, don't ask): read-only · no vault write · no outbound · ~2 Sonnet agents · not huge → spawn advocate↔skeptic (or champion-X↔champion-Y, or N orthogonal lenses) then synthesize. First line: 🤝 Spawning 2 Sonnet agents (read-only)….
• OFFER + ASK (`+`): vault-write/outbound · ≥3 agents or long/expensive · strategic-irreversible (→ full R+DR L2 below) · money/secrets/Tier-2. One line: 🤝 Agents: I recommend (…); shall I spawn them? (+).

1. Teammates on Sonnet; don't auto-keep an Opus lead; never ask for "consensus" — preserve dissent. Mechanism = Agent-tool subagents (cheap, ~80% of the value) or native Agent Teams when enabled.

This is L1.5 — broader than the L2 strategic flow: it fires on everyday comparisons/analysis, not only big decisions. The advocate/skeptic/frontier panel below is the same machinery.

The flow (what YOU, Claude Code, do)

Step 1 — RECALL. Pull EVERYTHING we already have on the topic (per the RECALL-before-activity rule):

• memory: MEMORY.md + grep the memory dir
• vault meaning: python "$IMPORTS_ROOT/brain_ask.py" "<topic>" (or /ask)
• vault exact: grep concepts/insights/protocols/people/leads
• SQLite facts where relevant (Platinum CRM, browser_history, etc.)

Report concisely: what we know · what we already researched · prior conclusions · prior mistakes.

Step 2 — GAP ANALYSIS. State plainly: what we do NOT know, which questions stay open. This sharpens the DR prompt so the external tool digs where we're blind, not where we're already strong.

Step 3-pre — DR-DEDUP: first check whether this DR already exists (operator's order, 2026-07-14). Plenty of sessions ordered a DR and never carried it to an external LLM; meanwhile the exporters (ChatGPT/Claude/Downloads) may have ALREADY pulled the report into the vault, the nightly dr_collect.py (hub, 05:05) flips the registry issued→collected on its own and dr_synthesize writes the digest — the registry status can be fresher than your memory. Before minting a NEW number, two cheap checks (0 tokens):

1. The registry: grep -i "<topic keywords>" "$OBSIDIAN_VAULT/_DR-Registry.md" (and/or dr_registry.py list) — is there already a DR on this topic, and in what state.
2. The vault: digests 03-Insights\insight-DR-*.md + originals _originals\deep-research\*<DR-ID>* (grep by topic/ID).

The fork: synthesized/collected → do NOT re-order — read the finished report/digest and go straight to Step 4 (Synthesis); issued on the same topic → reuse THAT SAME ID (don't mint a new one) and emit the outbound prompt with it; nothing → mint a new one below. Canon: reglament-numeratsiya-dr-i-reestr, memory dr-pipeline-endtoend.

Step 3 — DEEP RESEARCH PROMPT (hand-off). ⭐ FIRST allocate the DR number (operator's order, 2026-07-03): python $IMPORTS_ROOT/dr_registry.py new "<topic>" --tool <chatgpt|gemini|grok> → prints DRYY-MM-DD-MACHINE-NN (machine code auto-detected — per-machine sequences, no sync collisions) and registers it in _DR-Registry.md. Put the ID as the FIRST line of the emitted prompt (# DR26-07-03-ZB-01 — <topic>) so the external chat inherits it as its title and exports match the registry. Counter unsure → --gap (+10). When the report comes back: dr_registry.py update <ID> --status collected --file "<path>" — OR do nothing: the nightly dr_collect.py will find the DR-ID in the imported chats/Downloads and flip the status itself (the ownership loop closes automatically; a manual update is only needed if you want the report RIGHT NOW). Canon: the house rule reglament-numeratsiya-dr-i-reestr. You do NOT run the deep research yourself. Hybrid mode: you MAY do a light in-session web pass (WebSearch/WebFetch) to sharpen the prompt and catch the obvious — but the deep, exhaustive DR is done in an EXTERNAL tool (ChatGPT/Gemini/Perplexity Deep Research). Emit the prompt for the operator to copy:

• Read the canonical template: $OBSIDIAN_VAULT/08-Templates/deep-research-prompt-template.md and emit it filling `{topic}`, prepending a CONTEXT: block with the recall findings + open questions from Step 2 (so the external DR doesn't redo what we know).
• ⭐ Always append the `§1 UNIVERSAL ADD-ON` from `08-Templates\dr-platform-playbook.md` (mandatory citations w/ URLs, confidence tiers, epistemic neutrality, decision-oriented ending). If the operator names the target vendor, or you know it, also append that vendor's block (§2 Grok / §3 Gemini / §4 ChatGPT) so the platform does DR better. Playbook = tuning layer over the universal template.
• Present it as one clean, clearly-marked copy-paste block. ⭐ Standing mandate (operator, 2026-07-14): don't wait for him to carry the prompt anywhere — fan it out YOURSELF right away via `/dr-fanout` (ChatGPT+Gemini+Grok), any number of DRs, without asking (we don't ration quotas; degradation is soft; duplicates are caught by Step 3-pre). The copy-paste block stays in the chat as a fallback in case he wants to run it himself.
• ⭐ Two things into the chat, always (operator, 2026-07-17): ① the full prompt text in a ``text` block (paste-ready) + ② **a link to the actual research chat, the moment it exists**, without waiting for the report: 🔗 <vendor>: <url> for every fan-out chat that started. The link = the private URL (chatgpt.com/c/… · gemini.google.com/app/… · grok.com/chat/…); a public share link is ⛔ by default (that is publishing outward). Couldn't capture the URL → say so honestly with the reason. Broader than DR: if you couldn't pull something out of the browser, give the operator the address to pull it from and he will. Canon: reglament-vneshniy-resech-vsegda-promt-i-ssylka-v-chat`, memory [[dr-prompt-paste-in-chat]].

Step 4 — SYNTHESIS + DECISION MEMO (after the DR results come back). Merge memory + DR into a Decision Memo: the problem · what we already know · DR findings · options · risks · recommendation. Only AFTER the Decision Memo do we start implementation. ⭐ Close the DR in the same pass (operator's "+", 2026-07-28): the Decision Memo cites a DR-ID → in THAT SAME pass run dr_registry.py update <ID> --status applied --note "<what closed it>" (not adopting it → --status parked --note "<why>"). A memo that leaves the registry open = a half-built bridge. When ordering a DR in Step 3, name its consumer: new --for "<the decision this report feeds>" — with no consumer the report auto-parks after 30 days. Canon: the house rule reglament-numeratsiya-dr-i-reestr §amendment 2026-07-28, memory [[dr-finish-applied-or-parked]].

Notes

• If a /schedule-style recurring research cadence emerges for a topic, evaluate per evaluate-recurring-into-routine.
• The DR prompt template is the single source — if it changes, edit the vault template file, not a copy.
• This skill is SEARCH+STRUCTURE; the deep-research harness skill (deep-research) is the in-session web variant the operator can opt into, but his default flow is the external hand-off above.

<!--kit-footer-->

Like this skill? It is part of 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.