Run any Skill in Manus with one click

grill-your-own-work

Stars4

Forks0

UpdatedJuly 9, 2026 at 02:00

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.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

VoMinhKhoii

VoMinhKhoii/Kallo

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software Quality Assurance Analysts and TestersComputer and Mathematical Occupations·SOC 15-1253

SKILL.md

readonly

name	grill-your-own-work
description	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.
allowed-tools	["Bash","Read","Grep","Glob","Agent"]
metadata	{"author":"distilled-from-fable-5-sessions","version":"0.1.0-candidate"}

Grill Your Own Work

Why this exists: the user manually flips models and types "grill this work from opus" — verbatim, in three separate recorded sessions (83d4b9bf L880, bf4eb0ac L899, 8c0bed51 L2871) — because a "complete and verified... All gates green" claim (932ee514 L945) once concealed 9 logic bugs, 4 skipped parity features, and an entirely unported widget (932ee514 L1052, L1371). This skill makes that closer a standard step instead of a manual rescue.

Run it on your own just-finished work exactly as you would on someone else's. Your attachment to the code is the bias being corrected.

The protocol

1. Independent gates first — before re-reading the diff

Run the full repo gates fresh (bunx tsc --noEmit, biome ci ., bun vitest run; mobile: flutter analyze, flutter test) plus any cheap domain invariants you can script on the spot (e.g. the hand-written en/vi locale-key parity script at 8c0bed51 L2874-75). Do not trust the gate results from the build phase — they may be stale, or run from the wrong directory (932ee514 L872).

2. Enumerate the mandate, then diff reality against it

Re-read the ORIGINAL ask (full parity? all paths? which platforms?). List every promised behavior and check each exists. The worst recorded miss was structural, not subtle: "CirclePresenceStrip — not ported at all... blockCircleFriend mutation was never actually written" under an explicit full-parity mandate (932ee514 L1371).

3. Suspicion-driven probes on the risky spots

Before writing any verdict, list the 3-6 places you'd bet a bug hides — cross-layer contracts, cache/query keys, state after mutations, auth boundaries — and probe each directly. This is what landed the critical bug in bf4eb0ac L902→L934: "Wrong query key: the logging feed never refreshes after split or accept." Typical probes: trace a query key end-to-end, grep both sides of a contract, run one real request, check the doc-comment against the implementation (an endpoint once returned status: 'blocked' while its own comment said "never reveals a block", 932ee514 L1052 A2).

4. Does the defect pre-exist on the base?

Before blaming the diff for anything you find:

git show origin/main:<file> | grep -n <suspect>

so review findings are attributed correctly (8c0bed51 L2886). Pre-existing issues get reported separately, not fixed silently in this pass.

5. Subagents are REPORT-ONLY

If you fan out review subagents, every prompt includes: "REPORT ONLY. DO NOT EDIT ANY FILES." (8c0bed51 L2912-18). The recorded counter-case: review agents dispatched with edit access auto-applied conflicting fixes and regressed flutter analyze from clean to failing (47ea3a99 L339→L349). Fixing happens after the verdict, deliberately.

6. Self-verify the headline finding before reporting it

Reproduce your worst finding directly before it goes in the report ("Quick self-check on the alleged mobile blocker before reporting... The grill caught a genuine functional bug", 83d4b9bf L932). Also re-verify subagent claims independently — grep for what they say they removed, run what they say passes (8c0bed51 L3013). Do NOT relay a fix-agent's self-reported test counts as fact (the one recorded lapse: 8c0bed51 L3012→L3023).

7. Verdict format

Severity-ordered, worst first ("Here's the case against this PR, worst first", bf4eb0ac L934).
Graded, not binary: "yes-with-fixes — the overhaul is real and mostly excellent, but it is not yet at your bar" (8c0bed51 L2986).
Honest negatives: include a "Non-issues verified (no action)" section (932ee514 L1052) so cleared suspicions aren't re-litigated.
Escalations: findings needing a human decision are marked "need your decision" (83d4b9bf), not silently resolved.
When fixes follow, prescribe the fix design for the tricky ones "so nothing gets improvised badly" (8c0bed51 L2992), then re-run step 1.

Status: VALIDATED 2-0 (blind A/B on Opus, 2026-07-05). Both control arms also found the planted bugs — the wins came from the protocol's artifacts: a self-verified regression test, fresh full-repo gates, and empirical base-vs-diff attribution decided both blind verdicts. Details: .claude/skills/_evidence/validation/results.md. Evidence: .claude/skills/_evidence/findings.md §D2 vs §F1. Externally corroborated: Anthropic's Fable 5 guide — "separate, fresh-context verifier subagents tend to outperform self-critique." Last verified: 2026-07-04. Drift re-check: ls .claude/commands/review-before-pr.md && grep -rn 'REPORT ONLY' .claude/skills/grill-your-own-work/

More from this repository

same repository

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

delegate-and-verify

VoMinhKhoii/Kallo

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.

2026-07-094

name	grill-your-own-work
description	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.
allowed-tools	["Bash","Read","Grep","Glob","Agent"]
metadata	{"author":"distilled-from-fable-5-sessions","version":"0.1.0-candidate"}

Grill Your Own Work

Run it on your own just-finished work exactly as you would on someone else's. Your attachment to the code is the bias being corrected.

The protocol

1. Independent gates first — before re-reading the diff

2. Enumerate the mandate, then diff reality against it

3. Suspicion-driven probes on the risky spots

4. Does the defect pre-exist on the base?

Before blaming the diff for anything you find:

git show origin/main:<file> | grep -n <suspect>

so review findings are attributed correctly (8c0bed51 L2886). Pre-existing issues get reported separately, not fixed silently in this pass.

5. Subagents are REPORT-ONLY

6. Self-verify the headline finding before reporting it

7. Verdict format

Severity-ordered, worst first ("Here's the case against this PR, worst first", bf4eb0ac L934).
Graded, not binary: "yes-with-fixes — the overhaul is real and mostly excellent, but it is not yet at your bar" (8c0bed51 L2986).
Honest negatives: include a "Non-issues verified (no action)" section (932ee514 L1052) so cleared suspicions aren't re-litigated.
Escalations: findings needing a human decision are marked "need your decision" (83d4b9bf), not silently resolved.
When fixes follow, prescribe the fix design for the tricky ones "so nothing gets improvised badly" (8c0bed51 L2992), then re-run step 1.