feature-batch
Batches features and runs them. Tell it which features to batch and it writes features/QUEUE.md; pass it to /goal and it works that queue overnight, one feature at a time, building and verifying each one and opening a pull request when it passes. Use when the user says "batch these features", "queue these", "run these in a loop", "build these overnight", or names more than one feature to build.
arcasilesgroup · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add arcasilesgroup/ai-engineering --skill feature-batch技能说明
name: feature-batch description: Batches features and runs them. Tell it which features to batch and it writes features/QUEUE.md; pass it to /goal and it works that queue overnight, one feature at a time, building and verifying each one and opening a pull request when it passes. Use when the user says "batch these features", "queue these", "run these in a loop", "build these overnight", or names more than one feature to build.
feature-batch
Two jobs, same skill, because they're the same list.
- Batching — the user says which features to run. You write
features/QUEUE.md. - Running — the user puts this skill in a
/goal. You work the queue.
Which one you're doing is obvious from what they said. "Batch the availability and
the day view" is the first. A /goal firing on its own is the second.
The whole design rests on one thing: every pass starts from nothing. The agent does not remember the last feature, the last failure, or what it already tried. So the state lives in files, and this skill is the instruction for reading them, doing one unit of work, and writing the state back.
The two files
features/QUEUE.md — what's left, what's in flight, what's done.
# Feature batch
| # | Feature | Spec | Status | Failed passes | PR |
|---|---|---|---|---|---|
| 1 | Stylist availability | features/stylist-availability/spec.md | done | 0 | #12 |
| 2 | Admin day view | features/admin-day-view/spec.md | building | 1 | |
| 3 | Booking reminders | features/booking-reminders/spec.md | todo | 0 | |
Status is one of todo, building, done, stopped.
features/NOTES.md — what the runs worked out. Append, never rewrite. Told
to "update" this file the agent replaces it and last night disappears. One dated line
per finding, and only findings that would change what a later pass does.
Batching
When the user names the features they want run:
1. Check what's actually ready. A feature is ready when features/<name>/spec.md
exists. Anything without one hasn't been through new-feature yet. Name those out
loud rather than quietly leaving them out, or they'll assume it's running tonight.
2. Get the order right, and ask if you can't tell. The queue runs top to bottom, so if one feature builds on another the one underneath goes first. This is the last moment a person is in the room; after this nobody is watching and the order is fixed.
3. Write the file. Everything new goes in as todo, zero failed passes, no PR.
If a queue already exists, leave its rows exactly as they are and append underneath.
A done row reset to todo means rebuilding and re-opening a PR for something already
finished and merged.
4. Create features/NOTES.md if it's missing, empty with a one-line header. The
loop appends to it and won't create it mid-run.
5. Hand over the goal condition and stop. Starting the loop is theirs.
/goal Use feature-batch on features/QUEUE.md. Post the queue table each pass. Met when no row is todo or building.
Keep it that short. The condition only needs a pointer, one thing to post, and
what met looks like. Everything else is already in this file, and the queue table
carries the status and the PR number in its own columns, so posting the table is the
evidence. The give-up rule doesn't belong in there either, because a given-up feature
is marked stopped and stopped already counts as met. A long condition is a sign
you're duplicating this file into a place the evaluator can't use anyway.
If the queue is long, say something. Every row is a feature built while nobody is looking, paid for in real session limit. Eight rows is worth asking whether the last three are wanted tomorrow or just on a list.
Running: one pass
Do one unit of work per pass. Not one feature, one unit. A pass that changes nothing still counts as a pass.
- Read
QUEUE.mdandNOTES.md. Nothing else is known. - Pick the work. The topmost row that isn't
doneorstopped. If every row is one of those, the batch is finished. Say so and stop. - If it's
todo, create its branch off main, named for the feature, and set the row tobuilding. - Build. Spawn a subagent using the default task agent type, and give it the feature's spec path and nothing else. Never build in this session. It reads the spec, finds the highest thing not yet green, and fixes that one thing. Tell it not to declare the feature finished, because that isn't its call.
- Verify. Spawn the
adversaryagent. See below, this is the part that gets skipped and it's the part that matters. - Record. Post the verdict into this conversation in full. Append anything worth
carrying to
NOTES.md. Update the row. - If the verifier passed everything, run the close-out in "Finishing a feature".
Otherwise increment
Failed passesand stop the pass.
Only one feature is building at any time. Running several at once means several
session limits burning, work colliding in the same files, and no shared memory of what
any of them worked out. It's slower in wall-clock, not faster.
Both are subagents. Only one of them is custom
Building and verifying both run in subagents, never in this session. This session is the loop, and its job is to read the queue, hand work out, and write state back. The moment it starts building, its context fills with one feature's work and every later pass gets worse.
The builder is the default task agent type, with no custom definition, because everything that steers it is already in the spec: what to build, what done means, what to run. An agent file for it would be the spec copied into a second place, and two copies drift.
The verifier is the custom adversary agent in .claude/agents/, and it has to be
custom for one reason: what makes it work is what it's denied, and a spec can't deny
anything. It has no edit