research-swarm
Turn any hypothesis, however fringe, into an argument map rather than a verdict: five independent lenses (consensus, skeptic, frontier, historian, experimental design), then a synthesis of for and against, evidence quality, a confidence tier and the cheapest decisive experiment.
Install
npx skills add https://github.com/tonydzi/second-brain-starter-kit/tree/master/skills/research-swarm
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install tonydzi-second-brain-starter-kit@llmmart
git clone https://github.com/tonydzi/second-brain-starter-kit.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole tonydzi/second-brain-starter-kit collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
/research-swarm — an argument map for any hypothesis (a swarm of research agents)
🧒 When reporting to a non-technical operator: end with a child-simple "In plain words" recap in their language. ⚖️ Doctrine: this skill IS the operational arm of
protocol-epistemic-neutrality-fringe-research. NEVER judge or mock the hypothesis. Consensus ≠ truth. Build a map of arguments + confidence — an investigation, not a verdict. Bounded byoperating-agreementsafety invariants (no illegal/harm/bypass/fraud).
When to use
The operator hands a hypothesis — frontier science (antigravity, perpetual motion, anomalous materials, radical longevity, models of consciousness/time), an unusual business model (DAO/network states, no-employee AI co, token labor markets), or a controversial/stigmatized vertical (adult, gaming, gambling). Goal: a balanced, grounded argument map, not a yes/no.
The 5 lenses (run as a parallel swarm)
- Consensus — steelman the current mainstream/established view. What does well-replicated science / the market actually hold, and why.
- Skeptic — strongest arguments AGAINST. Where the idea breaks, hidden assumptions, falsification points, known failure modes.
- Frontier — strongest arguments FOR. Steelman the believer: what would have to be true, adjacent emerging evidence, why it's not obviously impossible.
- Historian — precedents and base rates. Ideas once dismissed then vindicated (and once hyped then debunked); the historical analog for THIS idea.
- Experimental-Design — what measurement/experiment would confirm or refute, the cheapest decisive test, what data to gather first.
Each lens does light, targeted web grounding (WebSearch/WebFetch) — claims should be evidenced, not just model priors — but stays token-aware. The synthesis NEVER returns a verdict; it returns a confidence tier + open questions.
Control node — Verifier-Calibrator (added 2026-06-15, via Alpha R+DR)
After synthesis, one Verifier-Calibrator node runs (the swarm's acceptance function, not another opinion). It (1) audits the strongest for/against claims — supported / partial / unsupported, (2) re-derives the confidence tier from evidence (support density + cross-lens agreement), not vibe, lowering tiers that rest on unsupported or single-lens claims, and (3) may abstain: the new tier insufficient is a first-class honest answer when evidence is too thin to place the idea on the established→fringe axis. The calibrated tier + claim audit + flags surface in the note and dashboard. This is the DR-recommended control plane over the 5-lens content plane (gaps: no verifier / qualitative-not-calibrated tiers / no abstain lane). Decision: vault 02-Decisions\decision-research-swarm-v2-upgrade.
How to run
Step 0 — parse the hypothesis from the operator's message (everything after the trigger). If it's a vague one-liner, ask ONE clarifying question first (scope/domain), else proceed.
Step 1 — RECALL first (token law — cheap before LLM/web). Pull what the operator already thinks on the topic:
python "$IMPORTS_ROOT/brain_ask.py" "<hypothesis>" --k 8
Keep the synthesized recall as recall_context (the operator's prior notes/leanings on the topic — feeds every lens so the swarm builds on their own thinking, not from zero).
Step 2 — run the swarm via the Workflow tool (this skill IS your explicit opt-in to orchestrate):
Workflow({ scriptPath: "%USERPROFILE%\.claude\\skills\\research-swarm\\workflow.js",
args: { hypothesis: "<hypothesis>", recall_context: "<from step 1>" } })
The script fans out the 5 lenses (parallel barrier — synthesis needs all), a synthesis agent merges them, then the Verifier-Calibrator audits the strongest claims and calibrates the tier (or abstains → insufficient). Returns the structured argument map object with a calibration block.
Step 3 — persist (backup first). python "$IMPORTS_ROOT/vault_backup.py" then:
- Write the map as a vault note in
03-Insights\Research-Swarm\<YYYY-MM-DD>-<slug>.md(frontmattertype: insight,origin: anton,tags: [research-swarm, argument-map]; body = for/against tables, confidence tier, historical analogs, open questions, the decisive experiment). Link it to[[protocol-epistemic-neutrality-fringe-research]]+[[insight-glavnaya-tsel-tsifrovoj-dvojnik]]and any concept the topic touches (no-orphan: ≥1 inbound). - Dump the map object to
<skill>\last_map.jsonand build the visual:
python "$USERPROFILE/.claude/skills/research-swarm/build_argmap_dashboard.py" <map.json> <out.html>
Dashboard → $OBSIDIAN_VAULT/_Dashboards/Research-Swarm/<slug>.html. Offer to open it (the operator works by eye).
Step 4 — reindex so the new note is searchable: python "$IMPORTS_ROOT/brain_embed_update.py".
Step 5 — report to the operator: the confidence tier, the 2-3 strongest for/against, the cheapest decisive experiment, link to the dashboard — and the ELI5 recap. NEVER editorialize a verdict; present the map.
Scaling (AK-47)
Default = 5 lenses, single pass. For a big/important hypothesis the operator can say "go deeper" → add a second round per lens or 3-vote adversarial verification of the strongest claims (loop-until-dry). Don't over-build by default.
Relation
Operational arm of vault protocol-epistemic-neutrality-fringe-research · memory [[epistemic-neutrality]]. Pairs with /alfa-search-recall-deepresearch (R+DR — recall→gap→deep-research→synthesis; the swarm is the multi-lens synthesis engine). Distinct from /ask (single recall) and /deep-research (web fan-out, single perspective). Serves [[main-goals]] goal #1 — the twin reasons like the operator-the-researcher.
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.
Files (second-brain-starter-kit)
-
build_argmap_dashboard.py 10.4 KB
Loading… -
last_map.json 73.9 KB
Loading… -
SKILL.md 7.2 KB
--- name: research-swarm description: >- Turn any hypothesis, however fringe, into an argument map rather than a verdict: five independent lenses (consensus, skeptic, frontier, historian, experimental design), then a synthesis of for and against, evidence quality, a confidence tier and the cheapest decisive experiment. Triggers: "/research-swarm <hypothesis>", "map the arguments on X". license: MIT --- # /research-swarm — an argument map for any hypothesis (a swarm of research agents) > 🧒 **When reporting to a non-technical operator:** end with a child-simple "In plain words" recap in their language. > ⚖️ **Doctrine:** this skill IS the operational arm of `protocol-epistemic-neutrality-fringe-research`. NEVER judge or mock the hypothesis. Consensus ≠ truth. Build a map of arguments + confidence — **an investigation, not a verdict**. Bounded by `operating-agreement` safety invariants (no illegal/harm/bypass/fraud). ## When to use The operator hands a hypothesis — frontier science (antigravity, perpetual motion, anomalous materials, radical longevity, models of consciousness/time), an unusual business model (DAO/network states, no-employee AI co, token labor markets), or a controversial/stigmatized vertical (adult, gaming, gambling). Goal: a balanced, grounded **argument map**, not a yes/no. ## The 5 lenses (run as a parallel swarm) 1. **Consensus** — steelman the current mainstream/established view. What does well-replicated science / the market actually hold, and why. 2. **Skeptic** — strongest arguments AGAINST. Where the idea breaks, hidden assumptions, falsification points, known failure modes. 3. **Frontier** — strongest arguments FOR. Steelman the believer: what would have to be true, adjacent emerging evidence, why it's not obviously impossible. 4. **Historian** — precedents and base rates. Ideas once dismissed then vindicated (and once hyped then debunked); the historical analog for THIS idea. 5. **Experimental-Design** — what measurement/experiment would confirm or refute, the **cheapest decisive test**, what data to gather first. Each lens does light, targeted web grounding (WebSearch/WebFetch) — claims should be evidenced, not just model priors — but stays token-aware. The synthesis NEVER returns a verdict; it returns a confidence tier + open questions. ## Control node — Verifier-Calibrator (added 2026-06-15, via Alpha R+DR) After synthesis, one **Verifier-Calibrator** node runs (the swarm's *acceptance function*, not another opinion). It (1) audits the strongest for/against claims — supported / partial / unsupported, (2) re-derives the confidence tier from **evidence** (support density + cross-lens agreement), not vibe, lowering tiers that rest on unsupported or single-lens claims, and (3) may **abstain**: the new tier `insufficient` is a first-class honest answer when evidence is too thin to place the idea on the established→fringe axis. The calibrated tier + claim audit + flags surface in the note and dashboard. This is the DR-recommended **control plane** over the 5-lens **content plane** (gaps: no verifier / qualitative-not-calibrated tiers / no abstain lane). Decision: vault `02-Decisions\decision-research-swarm-v2-upgrade`. ## How to run **Step 0 — parse** the hypothesis from the operator's message (everything after the trigger). If it's a vague one-liner, ask ONE clarifying question first (scope/domain), else proceed. **Step 1 — RECALL first** (token law — cheap before LLM/web). Pull what the operator already thinks on the topic: ``` python "$IMPORTS_ROOT/brain_ask.py" "<hypothesis>" --k 8 ``` Keep the synthesized recall as `recall_context` (the operator's prior notes/leanings on the topic — feeds every lens so the swarm builds on their own thinking, not from zero). **Step 2 — run the swarm** via the Workflow tool (this skill IS your explicit opt-in to orchestrate): ``` Workflow({ scriptPath: "%USERPROFILE%\.claude\\skills\\research-swarm\\workflow.js", args: { hypothesis: "<hypothesis>", recall_context: "<from step 1>" } }) ``` The script fans out the 5 lenses (parallel barrier — synthesis needs all), a synthesis agent merges them, then the **Verifier-Calibrator** audits the strongest claims and calibrates the tier (or abstains → `insufficient`). Returns the structured **argument map** object with a `calibration` block. **Step 3 — persist (backup first).** `python "$IMPORTS_ROOT/vault_backup.py"` then: - Write the map as a vault note in `03-Insights\Research-Swarm\<YYYY-MM-DD>-<slug>.md` (frontmatter `type: insight`, `origin: anton`, `tags: [research-swarm, argument-map]`; body = for/against tables, confidence tier, historical analogs, open questions, the decisive experiment). Link it to `[[protocol-epistemic-neutrality-fringe-research]]` + `[[insight-glavnaya-tsel-tsifrovoj-dvojnik]]` and any concept the topic touches (no-orphan: ≥1 inbound). - Dump the map object to `<skill>\last_map.json` and build the visual: ``` python "$USERPROFILE/.claude/skills/research-swarm/build_argmap_dashboard.py" <map.json> <out.html> ``` Dashboard → `$OBSIDIAN_VAULT/_Dashboards/Research-Swarm/<slug>.html`. Offer to open it (the operator works by eye). **Step 4 — reindex** so the new note is searchable: `python "$IMPORTS_ROOT/brain_embed_update.py"`. **Step 5 — report** to the operator: the confidence tier, the 2-3 strongest for/against, the cheapest decisive experiment, link to the dashboard — and the ELI5 recap. NEVER editorialize a verdict; present the map. ## Scaling (AK-47) Default = 5 lenses, single pass. For a big/important hypothesis the operator can say "go deeper" → add a second round per lens or 3-vote adversarial verification of the strongest claims (loop-until-dry). Don't over-build by default. ## Relation Operational arm of vault `protocol-epistemic-neutrality-fringe-research` · memory [[epistemic-neutrality]]. Pairs with `/alfa-search-recall-deepresearch` (R+DR — recall→gap→deep-research→synthesis; the swarm is the multi-lens synthesis engine). Distinct from `/ask` (single recall) and `/deep-research` (web fan-out, single perspective). Serves [[main-goals]] goal #1 — the twin reasons like the operator-the-researcher. <!--kit-footer--> --- **Like this skill?** It is part of [second-brain-starter-kit](https://github.com/tonydzi/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](https://github.com/tonydzi/secondop-panel) (a second opinion from a panel of external models), [claude-memory-tidy](https://github.com/tonydzi/claude-memory-tidy) (stop your agent's memory from rotting), [telegram-mcp-kit](https://github.com/tonydzi/telegram-mcp-kit) (your own Telegram over MCP in about 15 minutes). Author: **Anton Dziatkovskii**, Palo Alto AI Research Lab. Telegram [@tonydzi](https://t.me/tonydzi) - WhatsApp [+1 341 222 9178](https://wa.me/13412229178) - X [@Tony_Stef_](https://x.com/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. -
workflow.js 9.7 KB
Loading… -
_demo.html 8.3 KB · in bundle
-
_demo.json 3.3 KB
Loading…
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.