draft · git:20261111.eef3c68 · 2026-11-11 · sha256 1e0bd612f24d8d2d
draft git:20261111.eef3c68A
Immutable. This exact content never changes and is served at /api/v1/blob/1e0bd612f24d8d2d.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_gqhnc5g7xxssyfda/1e0bd612f24d8d2dc9cbc8e068da04205324aa00bd9574240f560708da16fd97.json
Then /plugin install draft@markdownregistry-gqhnc5g7xxssyfda in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
---
name: draft
description: "Draft a new Arbor pack and its campaign config for a person to review: start from the serving template (packs/serving, configs/serving-template.json) or the nearest existing pack; take the objective, its bands and the constraints from the benchmark's study (benchmarks/<name>/); keep the model's facts in its dossier, not the pack; write the skills in the lifecycle form (lifecycle.load); check it with a dry run; and end with a review note of what was copied, chosen and left open. Use when a new model, framework, GPU, benchmark or operating point needs a pack, or when asked to derive one pack from another."
metadata:
arbor:
modes: [use]
---
# Draft a pack
A pack is the contract every campaign of one kind shares: the objective and its bands, the protected validators, the
workload, the policy, the constraints the agent reads, and where skills take part (`docs/packs.md`). A new kind of
campaign gets a new pack. Today that means copying the nearest one by hand, and the copies drift apart. This skill drafts
one instead.
**It drafts; a person decides.** It writes the draft, checks it and explains it. It does not run a campaign, commit, or
change another pack.
**By hand:** build the family with `python3 -m arbor.skills plugin pack`, load it with the `claude --plugin-dir` line it
prints, then ask for a pack, or run `/pack:draft <what the campaign is>`.
## 1. Settle what the campaign is
Ask for what is missing, in one message, then go on with what you have:
- **the model:** its path or Hugging Face id, and its precision;
- **the framework:** vLLM or SGLang, and the image;
- **the GPU:** MI355X, MI350X, ..., and the TP;
- **the benchmark and workload:** InferenceX random 1k/1k, an AgentX trace replay, a document campaign;
- **the operating point:** concurrency, ISL/OSL, seed;
- **standalone, or under a host:** Hyperloom's MB, for example (docs/hosts.md).
The pack's name says what it is for, not a version: `agentx-<model>`, `serving-sglang-1k1k`. A variant of an existing
pack takes that pack's name and a suffix that says what differs (`-c8`, `-tp4`), never a bare `-2`.
## 2. Pick the parent
Start from the nearest pack, not from nothing: [parents.md](parents.md) says which, by workload. The default is the
serving template, `packs/serving`: `balanced`'s method with nothing model-specific in it.
**Copy the parent whole, then change it:**
```bash
cp -r packs/<parent> packs/<name> # pack.yaml, knowledge/, validator/, eval/
rm -rf packs/<name>/results # another campaign's numbers are not this one's evidence
```
**What stays with the parent's text:**
- the validators, unless the benchmark's study says the bench or the correctness check differs;
- the standing directives and the harness rules.
Say which parent, and why, in the review note.
## 3. Take what the benchmark decides from its study
Read `benchmarks/<benchmark>/` (the `benchmark` plugin's `study` writes it). If it is missing, say so, and either run
`/benchmark:study` first or leave the objective as the parent's, marked for review.
| The pack's | From the study |
| --- | --- |
| `objective.metric`, `direction`, `unit` | `metric.md`: the official metric, what the chart plots |
| `objective.measurement`: `threshold_rel`, `noise_band_pct`, `secondary_metric` and its band | `metric.md`: the noise, the bands, the second axis |
| `constraints` on what is fixed, forbidden or required | `rules.yaml`: one line per rule a change could break, with its source and who checks it |
| the config's operating point (`concurrency`, `num_prompts`, `ISL`, `OSL`, `BENCH_SEED`) | `metric.md` and `rules.yaml`'s `fixes` |
**Never invent a number.**
- A band comes from a measured noise floor, or stays the parent's, marked for review.
- An accuracy floor (`GSM8K_MIN`) comes from the model's measured reference score, less about three standard errors
of the subset. Without one, leave the template's 0.30 sanity floor and say so.
## 4. Keep the model in its dossier
The pack should hold nothing a different model of the same kind would not share. The model's layers, quantization,
code paths and directions belong to its dossier (`docs/dossier/<name>/`; `/dossier:read` finds it).
- **Knowledge the parent carries about another model:** move it out or drop it; never keep it.
- **A fact the campaign needs on its first iteration:** cite the dossier part that holds it, instead of copying it.
## 5. Write where its skills take part
Use the lifecycle form (`docs/packs.md`, "Where skills take part"; `docs/skills.md`), not `skills:`. The pack only
loads skills; each skill's own metadata says how and where it takes part:
```yaml
lifecycle:
load: ["benchmark:study", "dossier:discover", "dossier:measure", "dossier:directions", "dossier:update",
"dossier:read", profiler]
```
- **Extra skills for the iteration:** load `hardware`, `serving-playbook` and `kernels` too when the campaign may write
kernels.
- **A contributed skill:** load it only when the person asks.
- **Under a host:** add `component.lifecycle.load`, the list to load when a host owns the node. Without
`benchmark:study`, for example, when the host's rules stand in for the study.
## 6. Write the config
Start from `configs/serving-template.json`, or the parent's config:
- **The campaign:** `identity` (workload:shape:gpu:framework:precision:model:conc:tp), `goal` (one sentence: what
rises, at which operating point, without losing what), `pack`, `dossier` and `benchmark` (the names under
`docs/dossier/` and `benchmarks/`).
- **The serving:** `model_path`, `model_name`, `framework`, `tp`, `gpu_ids`, `server_flags` and `server_env`. Copy
them from the benchmark's reference recipe (`reference.md`): the arms must start from what the benchmark measures.
- **Under a host:** the host writes `docker_container`, `session_root`, the scripts and the component block. Draft
only what the host takes from the pack.
## 7. Check it
```bash
python3 -m arbor.skillcheck # the skills the pack names exist and check
python3 -m arbor run --config configs/<...>/<name>.json --dry-run # validate_pack, the config, the program it renders
```
Fix what fails. A dry run that passes says the pack loads. It does not say the pack is right: that is the person's
review.
## 8. End with the review note
In your reply:
- **What:** the pack's name and path, the config's path, and the parent.
- **The diff against the parent:** `diff -ru packs/<parent> packs/<name>`, summarized by section.
- **What you chose,** each with its source: the study, the dossier, the person, or the parent's text kept.
- **What is left for the person:** every `<...>` value, every band or floor marked for review, the open asks.
- **What you did not do:** no campaign run, no commit, no other pack changed.