원클릭으로 Manus에서 모든 스킬 실행

cmax

스타1

포크2

업데이트2026년 6월 9일 18:32

The default FAT umbrella. Auto-runs deepresearch + multispec decomposition + parallel /goal + verify. Opus for planning/judgment/verification; Sonnet for sub-Spec execution (escalates per router). Multispec engine selects Mode A (SDK subagents) or Mode B (Claude Code Agent Teams) automatically based on work size. Your default daily-driver.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

Manus에서 실행

출처

waitdeadai

waitdeadai/claudemax

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

/cmax — default daily-driver

The umbrella. You describe what you want; the pipeline produces it — 0 → production-ready, autonomously, on a new or an existing project (it runs in the current directory).

One simple entry: cmax ask "<your goal>" (= /cmax). Hardened by default — production-ready bar (PRC) + spec self-correction (SSC) + adversarial, decomposed, evidence-gated verify. No flags to remember. Lighten only if you want: --mvp (ship an MVP), --no-ssc, --no-adversarial. Granular control lives on cmax run.

Ask. Achieve. No 10-question wizards, no model picker, no decomposition-by-hand. The SOTA-2026 spec-driven + deepresearch-backed pipeline runs automatically.

Pipeline (default behavior, no flags)

Taste check — read taste.md + taste.vision if present. If absent and the repo has signal, suggest cmax taste init. Don't violate taste.
/deepresearch — web-current research with source ledger. Persisted to memory.research_sources. 2-bis. verify-research (default ON, added 2026-06-09) — decomposed per-claim verification of the brief: parallel Haiku citation-support checks on the brief's own cached excerpts (no network, 60s timeouts), Sonnet escalation (one spot-fetch + one fresh search, 120s) only for findings the excerpts can't support. Contradicted findings are stripped into openQuestions; unverified ones flagged [unverified]; >30% escalation demand fails the brief and re-researches once. Timeouts count as unverified, never as pass.
multispec decompose — Fable 5 (while included →2026-06-22; Opus after) authors a MultiSpec: N sub-Specs (each with measurable completion conditions + verifyHints) + DAG + rollup completion conditions + per-sub-Spec writeSet.
/specqa (parallel per sub-Spec) — gate spec quality.
/introspect (parallel per sub-Spec) — confidence/assumption check. Block at confidence < 6.
Mode selection — engine picks Mode A (SDK subagents in one query()) or Mode B (Claude Code Agent Teams) based on sub-Spec count, est. duration, cross-spec coordination, write-set overlap. Override with --mode {auto|solo|teams}.
Parallel /goal per DAG leaf — every leaf at the current frontier runs in parallel up to the hardware + credit-aware cap. Sonnet executes sub-Specs by default; Opus on router escalations.
per-sub-Spec /verify (parallel) — blind Opus re-reads repo, re-checks each completion condition.
rollup /verify — blind Opus checks the rollupCompletionConditions across all sub-Spec outputs.
Memory record — run snapshot to memory.runs + state to .claudemax/state/.

Variants

Skill	Routing
`/cmax`	Default — Fable 5 plan/decompose (while included on Max, →2026-06-22; Opus after) + Sonnet/Opus exec per era + Opus verify
`/opusolo`	Opus exec for sub-Specs too (max effectiveness, ~3× cost)

The /workflow + /opussonnet aliases were removed 2026-05-28; the opussonnet/opusolo exec-tier variants remain via cmax run --variant. Fable 5 defaults auto-demote to Opus on 2026-06-23 when Fable bills usage credits (CMAX_FABLE_ACCESS=included|credits flips the configurations); cmax loop run --fable / --tier fable reach Fable explicitly in both modes — see docs/MODEL_ROUTING.md.

Hard rules

Never declare success on the strength of the executor's own claim. The blind /verify is the source of truth.
Verify and spec always run on Opus, never demoted regardless of credit % or --cheap.
If /verify returns partial or failed, iterate /goal against the failing conditions; re-verify. If two iterations don't move it, the SPEC is probably wrong — re-spec, don't re-grind.
Cost-guard kicks in at 70% of monthly Agent SDK credit (demote non-essential Opus → Sonnet); at 90% aggressive demote; at 95% --force required.

When NOT to invoke /cmax

Trivial one-line changes — just edit.
A single SPEC that fits in one /goal loop — invoke /goal directly on the existing SPEC.md.
Pure research with no implementation — invoke /deepresearch directly.

이 저장소의 다른 Skills

같은 저장소

ask

waitdeadai/claudemax

Ask. Achieve. The canonical user-facing entry to claudemax — describe your goal, the SOTA-2026 pipeline (deepresearch → multispec → parallel /goal → verify) handles the rest. Same engine as /cmax with a friendlier verb. CLI equivalent is `cmax ask "<goal>"`.

2026-06-091

opusolo

waitdeadai/claudemax

Max-effectiveness umbrella. Opus for EVERYTHING — planning, decomposition, sub-Spec /goal execution, verification. Use for novel domains, security/auth/payments, or when you want the highest ceiling regardless of cost.

2026-06-091

route

waitdeadai/claudemax

Show the model-routing decision for a task (Fable/Opus/Sonnet/Haiku) with reasoning. Use before launching a packet to sanity-check the tier and to override when you have stronger judgment than the heuristic.

2026-06-091

review

waitdeadai/claudemax

Reviews a proposed diff for correctness, security, style, taste alignment. Distinct from /audit (which reviews existing code, not diffs) and /verify (which checks SPEC completion).

2026-06-091

specqa

waitdeadai/claudemax

Spec quality gate. Every completion condition must have a mechanically-checkable verifyHint (file path, command, test name, behavior). Blocks /goal handoff if the spec wouldn't be verifiable.

2026-06-021

verify

waitdeadai/claudemax

Independent Opus supervisor pass that re-reads the repo blind and re-checks every SPEC completion condition. The source of truth for whether a run actually succeeded. Never skip.

2026-06-021

name	cmax
description	The default FAT umbrella. Auto-runs deepresearch + multispec decomposition + parallel /goal + verify. Opus for planning/judgment/verification; Sonnet for sub-Spec execution (escalates per router). Multispec engine selects Mode A (SDK subagents) or Mode B (Claude Code Agent Teams) automatically based on work size. Your default daily-driver.

/cmax — default daily-driver

The umbrella. You describe what you want; the pipeline produces it — 0 → production-ready, autonomously, on a new or an existing project (it runs in the current directory).

One simple entry: cmax ask "<your goal>" (= /cmax). Hardened by default — production-ready bar (PRC) + spec self-correction (SSC) + adversarial, decomposed, evidence-gated verify. No flags to remember. Lighten only if you want: --mvp (ship an MVP), --no-ssc, --no-adversarial. Granular control lives on cmax run.

Ask. Achieve. No 10-question wizards, no model picker, no decomposition-by-hand. The SOTA-2026 spec-driven + deepresearch-backed pipeline runs automatically.

Pipeline (default behavior, no flags)

Taste check — read taste.md + taste.vision if present. If absent and the repo has signal, suggest cmax taste init. Don't violate taste.
/deepresearch — web-current research with source ledger. Persisted to memory.research_sources. 2-bis. verify-research (default ON, added 2026-06-09) — decomposed per-claim verification of the brief: parallel Haiku citation-support checks on the brief's own cached excerpts (no network, 60s timeouts), Sonnet escalation (one spot-fetch + one fresh search, 120s) only for findings the excerpts can't support. Contradicted findings are stripped into openQuestions; unverified ones flagged [unverified]; >30% escalation demand fails the brief and re-researches once. Timeouts count as unverified, never as pass.
multispec decompose — Fable 5 (while included →2026-06-22; Opus after) authors a MultiSpec: N sub-Specs (each with measurable completion conditions + verifyHints) + DAG + rollup completion conditions + per-sub-Spec writeSet.
/specqa (parallel per sub-Spec) — gate spec quality.
/introspect (parallel per sub-Spec) — confidence/assumption check. Block at confidence < 6.
Mode selection — engine picks Mode A (SDK subagents in one query()) or Mode B (Claude Code Agent Teams) based on sub-Spec count, est. duration, cross-spec coordination, write-set overlap. Override with --mode {auto|solo|teams}.
Parallel /goal per DAG leaf — every leaf at the current frontier runs in parallel up to the hardware + credit-aware cap. Sonnet executes sub-Specs by default; Opus on router escalations.
per-sub-Spec /verify (parallel) — blind Opus re-reads repo, re-checks each completion condition.
rollup /verify — blind Opus checks the rollupCompletionConditions across all sub-Spec outputs.
Memory record — run snapshot to memory.runs + state to .claudemax/state/.

Variants

Skill	Routing
`/cmax`	Default — Fable 5 plan/decompose (while included on Max, →2026-06-22; Opus after) + Sonnet/Opus exec per era + Opus verify
`/opusolo`	Opus exec for sub-Specs too (max effectiveness, ~3× cost)

The /workflow + /opussonnet aliases were removed 2026-05-28; the opussonnet/opusolo exec-tier variants remain via cmax run --variant. Fable 5 defaults auto-demote to Opus on 2026-06-23 when Fable bills usage credits (CMAX_FABLE_ACCESS=included|credits flips the configurations); cmax loop run --fable / --tier fable reach Fable explicitly in both modes — see docs/MODEL_ROUTING.md.

Hard rules

Never declare success on the strength of the executor's own claim. The blind /verify is the source of truth.
Verify and spec always run on Opus, never demoted regardless of credit % or --cheap.
If /verify returns partial or failed, iterate /goal against the failing conditions; re-verify. If two iterations don't move it, the SPEC is probably wrong — re-spec, don't re-grind.
Cost-guard kicks in at 70% of monthly Agent SDK credit (demote non-essential Opus → Sonnet); at 90% aggressive demote; at 95% --force required.

When NOT to invoke /cmax

Trivial one-line changes — just edit.
A single SPEC that fits in one /goal loop — invoke /goal directly on the existing SPEC.md.
Pure research with no implementation — invoke /deepresearch directly.