Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Human-Machine Brainstorm (HMB) — CCB Dispatcher Loop · cnfjlhj | DecimalAI

cnfjlhj/Human-Machine Brainstorm (人机风暴)

SkillScore82Maintained

This skill should be used when the user asks to "人机风暴", "Human-Machine Brainstorm", "human storm", "ccb brainstorm", "需求对齐调度", "spec convergence", or wants a CCB-based multi-model requirement alignment loop with Codex as the dispatcher.

agent-toolingby cnfjlhj

Links this skill to your agents and keeps it current with the author’s updates.

On-demandOn-demand: invoked explicitly — never occupies your agent's context until called.

Product Management pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/cnfjlhj-human-machine-brainstorm/SKILL.md

decimalai skills pull cnfjlhj-human-machine-brainstorm --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/cnfjlhj-human-machine-brainstorm/SKILL.md https://app.decimal.ai/s/cnfjlhj-human-machine-brainstorm/SKILL.md

Raw file: /s/cnfjlhj-human-machine-brainstorm/SKILL.md ↗ · machine-readable summary at /s/cnfjlhj-human-machine-brainstorm.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 38d ago

—+20% tokens0% turns82% of whole cases pass with skill (18/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-02	✗→✓	▲ Improved	57%	0%
case-07	✗→✓	▲ Improved	41%	0%
case-08	✗→✓	▲ Improved	0%	0%
case-09	✗→✓	▲ Improved	-18%	0%
case-10	✗→✓	▲ Improved	107%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/3/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

cnfjlhj / ai-collab-playbook

skills/full/human-machine-brainstorm/SKILL.md

View on GitHub ↗

Imported 8/3/2026Version v1Bundled 5 files · scripts

Skill contents

scripts/

references/

SKILL.md

Human-Machine Brainstorm (HMB) — CCB Dispatcher Loop

Purpose

Run a repeatable multi-model requirement alignment loop in CCB where:

Codex acts as the dispatcher (facilitator + router).
Claude Code acts as the scribe (single source of truth spec author).
OpenCode (Gemini) acts as the divergent thinker (alternatives + ASCII prototypes).

Keep every round auditable by exporting per-provider context into ./.ccb/history/ and keeping the canonical spec in ./.ccb/spec/.

Hard Rules

Treat ./.ccb/spec/overview.md as the single source of truth. Only Claude Code edits it.
Never assume panes “share context”. Always broadcast updates explicitly.
Enforce question IDs in this format:
Claude: C-Q01, C-Q02, ...
OpenCode: O-Q01, O-Q02, ...
Accept user answers only in ID-addressed form (so routing is deterministic).

Quick Start (zero-friction, recommended)

This workflow is optimized for a 2×Codex setup:

Cmd pane runs a dedicated Codex Chair (dispatcher).
The normal Codex provider pane participates as a reviewer/solver (not just idle).

1) Create (or choose) a topic directory (recommended location: $HOME/ccb-startups/...).

Optional helper: bash $HOME/.codex/skills/human-machine-brainstorm/scripts/hmb-init.sh "<topic-slug>"

2) Start CCB.

If CCB global config enables the chair cmd pane, just run: ccb
Otherwise: ccb claude codex opencode cmd (fallback)

3) Talk only to the Codex Chair (cmd pane).

Paste the raw requirement.
The chair broadcasts to claude, opencode, and participant codex via ask.
Use the round prompt template in references/round_prompt_template.md if needed.

Round Loop (R1/R2/R3...)

Step A — Broadcast (dispatcher = Codex Chair)

Send the same “Round prompt” to:

ask claude "<ROUND PROMPT>"
ask opencode "<ROUND PROMPT>"
ask codex "<ROUND PROMPT>" (participant Codex pane)

Require them to respond with:

10–20 numbered questions using C-Q## / O-Q## / P-Q##
1 ASCII diagram (flow/state/component)
1 short “current assumptions” list

Step B — Collect Answers (human)

Ask the human to answer in this format:

C-Q01: ...
C-Q02: ...
O-Q01: ...
P-Q01: ...

Optionally allow a shared block:

SHARED: ... (facts that apply to both)

Step C — Route Answers (dispatcher = Codex)

Send Claude only C-* + SHARED. Send OpenCode only O-* + SHARED. Send participant Codex only P-* + SHARED.

Step D — Export Evidence (end of round)

From the cmd pane (or any shell pane) run:

./.ccb/bin/round-save.sh 20

This writes:

./.ccb/history/claude-<timestamp>.md
./.ccb/history/codex-<timestamp>.md
./.ccb/history/opencode-<timestamp>.md

Step E — Update Spec (scribe = Claude Code)

Ask Claude to update:

./.ccb/spec/overview.md (bump version vN)
./.ccb/spec/open_questions.md (close answered questions)
./.ccb/spec/decisions.md (record non-reversible decisions)
./.ccb/spec/changelog.md (vN → vN+1 diff)

Then re-run another round until all reviewers say “no blocking issues”.

Final Handoff to GPT-5.2 (new session)

Provide a clean handoff pack:

./.ccb/spec/overview.md
./.ccb/spec/decisions.md
./.ccb/spec/open_questions.md (should be empty or non-blocking)
./.ccb/spec/changelog.md

Instruct GPT-5.2 to:

Treat the spec as authoritative
Output an executable plan first
Use multi-agent decomposition for implementation/testing/review

Illustrative Want to see what cnfjlhj/Human-Machine Brainstorm (人机风暴) would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/21/2026, 15:09 UTC · 22 cases

Measured on 20 comparable cases — below the 22 needed to publish a lift, so this result is shown here but not on the registry card or leaderboard.

Lift

+64%

82% vs 18%

Agent turns

1.0 vs 1.0

Tokens

+20%

1,920.35 vs 1,601.9

fail → pass · 14pass → pass · 4fail → fail · 4

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 I want to kick off a multi-model requirement alignment brainstorm for my new payment gateway service using our CCB setup. Please initialize the environment and dispatch my raw product prompt across all active helper panes. Make sure each model returns its initial clarifying questions, an ASCII system layout, and its current list of assumptions so I can answer them.	fail→fail	16,908	8,053	-52%	1	1	0%	2,944	1,533	-48%	0	0	—
case-02 Here are my answers to the clarifying questions from the previous discussion round: C-Q01: We will use PostgreSQL for transaction storage. C-Q02: Maximum latency SLA is 200ms. O-Q01: Use Redis for transient caching. P-Q01: We can support retry attempts up to 3 times. SHARED: All inter-service communication must be encrypted via mTLS. Please deliver these responses to their respective target models, archive this round's execution context, and have our spec maintainer update the core specifications and decision logs.	fail→pass	8,120	9,459	+16%	1	1	0%	1,435	2,254	+57%	0	0	—
case-03 We have reached consensus on our system design and there are no remaining unresolved questions. Please run the final wrap-up procedures: save the final context history, verify that all specification documents and decision records are current, and compile the final handoff bundle so the downstream implementation model can generate an execution plan.	fail→fail	13,362	3,968	-70%	1	1	0%	2,172	1,586	-27%	0	0	—
case-04 We need to create a Git pull request to merge feature branch feat/auth-tokens into main for the API gateway repository. Write the terminal commands to push the branch, open a PR via GitHub CLI, and request review from team leads.	pass→pass	6,853	5,287	-23%	1	1	0%	1,256	2,039	+62%	0	0	—
case-05 Draft a standard C4 model Container diagram specification in PlantUML syntax for a microservice order processing system.	pass→pass	14,026	13,322	-5%	1	1	0%	2,582	3,662	+42%	0	0	—
case-06 Create a 5-day human design sprint workshop schedule for an e-commerce mobile app team, covering map, sketch, decide, prototype, and test phases.	pass→pass	21,781	25,569	+17%	1	1	0%	4,112	5,649	+37%	0	0	—
case-07 We need to revise the project overview document for our notification engine in ./.ccb/spec/overview.md following our recent round. Should I ask OpenCode or Codex to update this spec file directly?	fail→pass	7,618	3,861	-49%	1	1	0%	1,259	1,779	+41%	0	0	—
case-08 When collecting clarification questions from Claude Code during a requirement alignment session, what ID prefix format should be assigned to Claude's questions?	fail→pass	8,957	2,168	-76%	1	1	0%	1,404	1,405	+0%	0	0	—
case-09 When OpenCode generates divergent suggestions during a round, what specific question ID tag structure must be used for its query items?	fail→pass	11,703	2,958	-75%	1	1	0%	1,805	1,480	-18%	0	0	—
case-10 I am initializing a new requirement brainstorm project directory for our search service 'search-v2'. What default directory path structure is recommended for hosting CCB project startups?	fail→pass	27,669	5,616	-80%	1	1	0%	1,103	2,281	+107%	0	0	—
case-11 A human participant submitted a response block tagged SHARED: Database pool size is fixed at 50 connections. Which model panes should receive this statement during routing?	pass→pass	7,400	3,211	-57%	1	1	0%	529	1,594	+201%	0	0	—
case-12 What script command should be executed from the command pane at the end of round 5 to export context history across model providers?	fail→fail	12,820	2,585	-80%	1	1	0%	1,913	1,497	-22%	0	0	—
case-13 Where does the context preservation mechanism store per-provider markdown logs after completing a brainstorm iteration?	fail→pass	8,087	2,967	-63%	1	1	0%	1,276	1,535	+20%	0	0	—
case-14 After completing round 2 answer routing, where should the version diff and increment details from v2 to v3 be recorded?	fail→pass	8,801	3,137	-64%	1	1	0%	1,321	1,554	+18%	0	0	—
case-15 During round 3 review, the team confirmed that SQLite will not be supported in production. In which file should this non-reversible architectural choice be documented?	fail→pass	4,459	2,785	-38%	1	1	0%	774	1,549	+100%	0	0	—
case-16 When answers resolve previous ambiguity around authentication mechanisms, which file in the spec directory must be updated to close answered items?	fail→fail	7,311	3,376	-54%	1	1	0%	1,129	1,735	+54%	0	0	—
case-17 Which model pane in the CCB setup is assigned the role of divergent thinker responsible for producing alternative proposals and ASCII prototypes?	fail→pass	11,188	1,472	-87%	1	1	0%	1,731	1,293	-25%	0	0	—
case-18 Which model instance and pane combination acts as the dedicated facilitator and dispatcher chair for broadcasting requirement prompts?	fail→pass	10,577	2,736	-74%	1	1	0%	1,481	1,534	+4%	0	0	—
case-19 When broadcasting a round prompt to participant models, what visual output artifact must each helper model produce alongside questions and assumptions?	fail→pass	5,501	2,395	-56%	1	1	0%	756	1,364	+80%	0	0	—
case-20 If global configuration for the chair command pane is disabled in CCB, what fallback shell command launches the multi-pane environment with all necessary roles?	fail→pass	14,967	2,502	-83%	1	1	0%	2,083	1,378	-34%	0	0	—
case-21 How does the dispatcher determine when to stop repeating round iterations and proceed to final handoff?	fail→pass	13,757	6,765	-51%	1	1	0%	2,019	2,198	+9%	0	0	—
case-22 Which model is specified as the downstream target for receiving the final specification handoff bundle to generate code execution plans?	fail→pass	14,605	2,577	-82%	1	1	0%	1,001	1,322	+32%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 20 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +64 percentage points is the difference between those two pass rates over the 20 comparable cases.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 18.2%
Pass rate with the skill: 81.8%
Cases attempted: 22
Cases counted toward lift: 20
Cases that regressed: 0
Skill version measured: v1
Run completed: 8/21/2026, 15:09 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-02

Prompt

Here are my answers to the clarifying questions from the previous discussion round:

C-Q01: We will use PostgreSQL for transaction storage.
C-Q02: Maximum latency SLA is 200ms.
O-Q01: Use Redis for transient caching.
P-Q01: We can support retry attempts up to 3 times.
SHARED: All inter-service communication must be encrypted via mTLS.

Please deliver these responses to their respective target models, archive this round's execution context, and have our spec maintainer update the core specifications and decision logs.

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run82

Live evalNeeds ≥10 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+53% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 37 days ago

Other measured skills in the registry, with their headline benchmark lift.