원클릭으로 Manus에서 모든 스킬 실행

delegate-and-verify

스타4

포크0

업데이트2026년 7월 9일 02:00

Subagent discipline — invoke whenever dispatching subagents (reviews, audits, fix fleets, research fan-outs) or whenever consuming another agent's/session's claimed results (handoff notes, "previous session fixed X", fleet self-reports). The recorded Opus failure modes: review subagents given edit access that collided and broke the build, and subagent self-reports relayed as fact without independent verification.

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

VoMinhKhoii

VoMinhKhoii/Kallo

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Manus에서 실행

Delegate and Verify

The recorded failures this prevents: during a /review-before-pr, 8 subagents ran WITH edit access and "auto-applied several fixes concurrently — multiple agents touched linked-accounts.tsx… in parallel. That's risky" — the collision regressed flutter analyze from clean to failing, a CI blocker the orchestrator then had to fix (47ea3a99 L339→349). Separately, a fix-fleet's self-reported test counts were relayed as fact and pushed without an independent orchestrator run (8c0bed51 L3012→3023).

Rule 1 — Reviewers and auditors are REPORT-ONLY

Every review/audit/verification subagent prompt contains, verbatim: "REPORT ONLY. DO NOT EDIT ANY FILES." (8c0bed51 L2912-18 — the recorded Fable contract.) Fixing is a separate, deliberate phase after the verdict. If two agents may write, they get isolated worktrees or strictly disjoint file sets — never shared write access to one file.

Rule 2 — Fix-fleets get prescribed designs, not open mandates

When dispatching agents to FIX findings, prescribe the fix design for anything tricky "so nothing gets improvised badly" (8c0bed51 L2992): file:line, the intended mechanism, and what must NOT change. Open-ended "fix it" prompts to fleets produce improvisations the orchestrator then has to re-review anyway. Credit justified deviations when an agent explains one (8c0bed51 L3032) — prescription is a floor, not a cage.

Rule 3 — Never relay a subagent's self-report as fact

A subagent saying "all tests pass" or "removed all occurrences" is a CLAIM. Before you repeat it to the user, commit, or push:

re-run the gate the claim rests on yourself (or a targeted subset + the full gate before push), and
grep for what they say they removed/changed ("Verified clean — zero surviving uncertainty surfacing. Pushing" — 8c0bed51 L3013, done right; L3012→3023, the recorded lapse when the suite counts were relayed unverified). This applies equally to handoff notes from previous sessions: verify the claimed state before building on it.

Rule 4 — Delegate to buy wall-clock, not to avoid thinking

Launch research/exploration agents in the background FIRST, then do foreground work while they run ("let me launch those in the background first, then do the visual fixes while they research", ad88fbba L1012). When an agent times out or dies, adapt immediately rather than blocking: "the review agent timed out → I'll do the grilling myself with direct reads" (932ee514 L983).

Rule 5 — Size fan-outs to the budget; make them resumable

A 37-agent sweep tripped the session token limit three times and needed two "please continue" nudges (8c0bed51 L114-151). It was rescued only because the workflow was journal-resumable, and when told to stop, the right move was salvage-from-journal, not rerun (L197-207). Before a large fan-out: estimate agents × tokens vs remaining budget; prefer pipelined batches; make partial results harvestable.

Status: WEAKLY VALIDATED 2-0 (blind A/B on Opus, 2026-07-05) — read the caveat. Both wins exercised Rule 3 (never relay claimed results — the lying-handoff trap and independent ground-truthing). Rule 1 (REPORT-ONLY dispatch) was NOT testable: arms are subagents and cannot spawn subagents, so that rule rests on transcript evidence alone (47ea3a99 L339/349). Details: .claude/skills/_evidence/validation/results-batch2.md. Evidence: .claude/skills/_evidence/findings.md §F7/§D2.5/§D7 (47ea3a99 L339/349; 8c0bed51 L2912-18, L2992, L3013, L3012→3023, L197-207; ad88fbba L1012; 932ee514 L983). Last verified: 2026-07-05. Drift re-check: ls .claude/commands/review-before-pr.md

이 저장소의 다른 Skills

같은 저장소

nham-design

VoMinhKhoii/Kallo

Use this skill to generate well-branded interfaces and assets for Nhẩm — the AI-powered Vietnamese meal tracker — either for production or throwaway prototypes/mocks/etc. Contains essential design guidelines, colors (warm earthy palette anchored on a cream surface and a tan accent), type (Lora serif + Be Vietnam Pro sans), fonts, assets (logo + app mark + favicon), Lucide iconography conventions, and a click-through React UI kit recreating the Landing / Dashboard / Logging surfaces of the product. Covers two platforms — web (Next.js, Be Vietnam Pro + Lora, --nham-* tokens) and the mobile Flutter app (Be Vietnam Pro, a calmer Threads-tuned type scale; see mobile.md) — routed by platform so you only load what you need.

2026-07-184

thermo-nuclear-code-quality-review

VoMinhKhoii/Kallo

Run an extremely strict maintainability review for abstraction quality, oversized files, flat-folder sprawl, and spaghetti-condition growth — tuned to this repo. Use for a thermo-nuclear code quality review, thermonuclear review, deep code quality audit, or especially harsh maintainability review; also consult whenever a change grows any file near the repo's 400/200 LOC limits or adds files to an already crowded folder. This project-local version shadows the global gstack skill of the same name: it replaces the generic 1000-line rule with the repo's strict 400/200 thresholds (mechanically enforced by `bun check:filesize`) and adds binding File & Folder Organization standards (feature-first nesting, separation of concerns, folder size limits, no barrels).

2026-07-114

feature-workflow

VoMinhKhoii/Kallo

The umbrella dev loop for any non-trivial feature, fix, or refactor in this repo — invoke at TASK START, before exploring or editing. It does not add new rules; it sequences the validated discipline skills at the phase boundaries where the recorded Opus sessions put them in the wrong order (gates at the end instead of per phase, review after merge instead of before done, decisions locked before the user saw them). Routes to: decision-gate → probe-state-before-acting → implement with per-phase verification → verify-before-done → grill-your-own-work → /ship.

2026-07-094

root-cause-first

VoMinhKhoii/Kallo

Trigger discipline for bugs, failures, and anomalies: invoke on ANY unexpected behavior — a test failure, a broken feature, a retry, a "still broken" report, OR an anomaly during your own QA (a tap that doesn't land, output that looks off). Forbids a second patch on the same symptom until the mechanism of the failure is stated; forbids rationalizing away anomalies; requires checking memory/docs for known fixes before improvising. Complements superpowers:systematic-debugging (the full protocol — load it for any nontrivial bug); this skill encodes when to trigger it and the repo-specific failure patterns recorded in this project's session history.

2026-07-094

decision-gate

VoMinhKhoii/Kallo

Decision discipline for open-ended briefs — invoke whenever a task says "figure out the best X", "design", "improve", "you decide", or whenever your implementation is about to lock in something expensive to reverse (schema/migration shape, API contract, UX paradigm, dependency choice, data model). The recorded Opus failure is building past decisions that were the user's to make; the recorded low-steering sessions won by surfacing exactly one well-framed decision at the real fork. This skill does NOT teach planning (the evidence shows no planning-quality gap) — it teaches WHERE to stop.

2026-07-094

grill-your-own-work

VoMinhKhoii/Kallo

Adversarial self-review closer — run AFTER a feature is "complete and verified" and BEFORE declaring done, /ship, or opening a PR; also when asked to review another session's/model's/agent's work ("grill this", "audit this", "check Opus's work"). Encodes the exact protocol the user repeatedly switches models to get: independent gates first, suspicion-driven probes, defect-preexists-on-base checks, REPORT-ONLY subagents, self-verified headline findings, severity-ordered graded verdicts. In the one recorded direct A/B, this pass turned a green-gated "complete" feature port into a one-go merge by finding 9 logic bugs and 4 missing parity features first.

2026-07-094

name	delegate-and-verify
description	Subagent discipline — invoke whenever dispatching subagents (reviews, audits, fix fleets, research fan-outs) or whenever consuming another agent's/session's claimed results (handoff notes, "previous session fixed X", fleet self-reports). The recorded Opus failure modes: review subagents given edit access that collided and broke the build, and subagent self-reports relayed as fact without independent verification.
allowed-tools	["Agent","Bash","Read","Grep"]
metadata	{"author":"distilled-from-fable-5-sessions","version":"0.1.0-candidate"}

Delegate and Verify

Rule 1 — Reviewers and auditors are REPORT-ONLY

Rule 2 — Fix-fleets get prescribed designs, not open mandates

Rule 3 — Never relay a subagent's self-report as fact

A subagent saying "all tests pass" or "removed all occurrences" is a CLAIM. Before you repeat it to the user, commit, or push:

re-run the gate the claim rests on yourself (or a targeted subset + the full gate before push), and
grep for what they say they removed/changed ("Verified clean — zero surviving uncertainty surfacing. Pushing" — 8c0bed51 L3013, done right; L3012→3023, the recorded lapse when the suite counts were relayed unverified). This applies equally to handoff notes from previous sessions: verify the claimed state before building on it.