sushiai-task
Skill2 starsChanged 26 days agoMITView on GitHub
Carry an owner's sushiAI task to a verified local commit without supervision: scope it, delegate implementer, functional-qa, reviewer and design-critic lanes, loop on findings, and report back with only the questions the owner must answer. Use for any change bigger than one obvious edit; not for read-only questions or a one-line fix.
In this repository11 other files it gives its agents.
- AGENTS.mdAGENTS.md
- author-and-verify-extensionSkill
- autoreviewSkill
- cut-releaseSkill
- deslopSkill
- extension-manifest-changeSkill
- figma-to-codeSkill
- ship-prSkill
- sushiai-testingSkill
- sushiai-uiSkill
- ui-evidenceSkill
The file
We split this skill into the rules it gives an agent
11 rules. About 0 are common, 5 are distinctive, and 6 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 29
Scope: Read AGENTS.md, git status -sb, the files the task touches with their callers and tests, and docs/LESSONS.md.
- Distinctive: line 34
Done when: every criterion is observable and every touched file is named.
- Very specific: line 35
Separate what needs the owner: Ask only about: the extension manifest contract or other public shapes, the security model, …
- Very specific: line 51
Clean: Run $deslop over the whole diff.
- Very specific: line 59
Review: Run $autoreview on the final diff.
- Distinctive: line 60
Done when: every P0/P1 finding is verified against the code and either fixed or rejected with a reason.
- Very specific: line 62
Judge the look: Only when a screen changed: brief design-critic with the screenshot paths and the owner's request, nothing else.
- Distinctive: line 69
Done when: no blocking finding is open.
- Very specific: line 83
Cutline: After step 1 write the done list (about eight items) and a time budget, and tell the owner the ETA.
- Distinctive: line 87
Ask consent and security UX before building: Default to the lowest-friction choice for owner-owned resources (a host the owner added is …
- Distinctive: line 108
Never stop, restart or reconfigure a process, dev server or app you did not start.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
sushiAI task loop
The owner hands over a task and leaves. Carry it to a verified local commit. The owner hears back once: the result, the evidence, and only the questions nobody else can answer. A plan or a progress note is a checkpoint, not done.
Roles and models
| Lane | Who | Model | Owns |
| lead | the main session | session | scope, briefs, integration, commit, report |
implementer | subagent | sonnet | one bounded code lane and its tests |
functional-qa | subagent | sonnet | proof against the acceptance criteria |
reviewer | subagent | opus | fresh correctness and root-cause review of the diff |
design-critic | subagent | opus | judgement of the screenshots, only when a screen changes |
| wide read-only search | built-in Explore | default | locating code across many files |
Sonnet does work that something else checks afterwards (tests, review). Opus does judgement nothing checks after it. The lead stays hands-on: it verifies every consequential conclusion a lane reports instead of forwarding it.
Loop
- Scope. Read
AGENTS.md,git status -sb, the files the task touches with their callers and tests, anddocs/LESSONS.md. Reproduce a defect through the real entry point before editing. Check why a path is missing before restoring it (git log -p -S <symbol>). Write acceptance criteria a person could check from outside the app. Done when: every criterion is observable and every touched file is named. - Separate what needs the owner. Ask only about: the extension manifest
contract or other public shapes, the security model,
src/app/electron/main.cjsstructural changes (AGENTS.md"Vision & boundaries"), destructive or irreversible actions, paid services, pushing, merging, packaging or releasing. Resolve everything else with a reasonable assumption and record it. When the owner is away, park the question in the report and keep every lane that does not depend on it moving. Done when: each open decision is either assumed and recorded or parked. - Build. Small or tightly coupled work: the lead, or one
implementerin the main checkout. Independent lanes: oneimplementereach, run with worktree isolation. A worktree branches frommain, not from the lead's branch - give the lane the lead's commit SHA and have it start withgit checkout -B <lane> <sha>and linknode_modulesfrom the main checkout. Brief every lane with: the criteria it serves, the files it owns, the tests to add, and what it must not touch. Two lanes never own one file. Done when:npm run buildexits 0 and each lane's tests pass. - Clean. Run
$deslopover the whole diff. Done when: the diff carries no narration comments, imagined-state guards or type laundering. - Prove. Brief
functional-qawith the criteria. It follows$sushiai-testing, and$ui-evidencewhensrc/app,src/extensions,electron/or any screen changed. Done when:npm run ciexits 0, every criterion maps to a test or a reproducible step, and gaps are listed. - Review. Run
$autoreviewon the final diff. Done when: every P0/P1 finding is verified against the code and either fixed or rejected with a reason. - Judge the look. Only when a screen changed: brief
design-criticwith the screenshot paths and the owner's request, nothing else. Done when: the verdict is PASS or PASS_WITH_NOTES, or its FAIL is fixed. - Loop. Each blocking finding goes back to step 3 as a brief carrying its evidence; rerun only the proof the fix affects. The same finding failing a third time means the approach is wrong: stop patching, write the root cause and a different approach into the report. Done when: no blocking finding is open.
- Land locally. Stage only intended files. One Conventional Commit in
English with no attribution trailers. A
docs/releases/unreleased/<slug>.mdfragment for anything user-visible. Adocs/LESSONS.mdentry, or the report says there was none. Never push, merge, package or runship-pr/cut-releaseunless the owner asked in this task. Done when:git status -sbshows only unrelated work left. - Report in the shape of references/report.md.
Pace
These rules keep a task from growing into days. Source: the project-environments
retro (artifacts/RETRO-project-environments.md).
- Cutline. After step 1 write the done list (about eight items) and a time
budget, and tell the owner the ETA. A new ask goes to a
Nextlist unless the owner marks it blocking. Update the ETA after every pass; never answer "how is it going" with only "still running". - Ask consent and security UX before building. Default to the lowest-friction choice for owner-owned resources (a host the owner added is trusted). Never build a second mechanism for a decision the owner already made.
- Two gate tiers. Inner loop: targeted tests and only the affected screens.
Final gate once:
prettier, build,npm run ci,npm run test:desktop, the full screen pass.npm run cialready selects the checks the diff needs; do not add whole suites by hand. Revert-red checks run once at the end, only for tests that guard security or data. - Review budget. One full
reviewerpass and one delta pass. The second reports P0/P1 only; P2 goes to the report's "Left open" unless it is a security issue.design-critic: one full pass, one delta pass. - Fixed fidelity tolerances, set by the lead before the run (text width 3%, shifts 2 px, no data-list excuses for layout). The critic, not a diff tool the implementer wrote, decides.
- Parallel lanes for independent surfaces (electron/orchestrator module, renderer, docs and Figma), each in its own worktree.
- Preflight (two minutes): Figma bridge connected, ssh alias answers, the build runs, the ports you need are free. Fix this before briefing a lane.
Boundaries that hold in every lane
- Never stop, restart or reconfigure a process, dev server or app you did not start; never free a port by force.
- Tests and examples use synthetic data:
192.0.2.0/24,user@devbox, invented names. Real hosts, IPs, paths and credentials stay out of files, commits and screenshots. - A lane that changes how a screen looks runs the matching
$ui-evidencedriver, opens the PNGs with the Read tool and lists what it saw (crooked, cramped, clipped, low contrast) before it reports done; a new screen also goes todesign-critic. Green tests do not count as having looked. - One owner, complete cutover (
AGENTS.md): remove the old path in the same change. - The exit code is the verdict, not a summary line. Use
npm run cior./node_modules/.bin/<tool>, nevernpx.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

