Content
86%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A well-engineered meta-skill body: dense with tool-specific facts, concrete commands and exact paths, a conditional reading workflow, real guardrails against destructive overwrites, and a references section that maps 1:1 onto the actual bundle with one-level-deep navigation. The weaknesses are minor: some duplication between the intro and Do Not sections, and validation guidance that could be consolidated into an explicit pre-edit checklist.
Suggestions
State the lockstep dual-package rule once (Do Not section) and reference it from the intro instead of repeating it nearly verbatim, saving tokens.
Consolidate the pre-edit safety checks (template-hashes, .new sidecars, trellis --help cross-check) into one explicit ordered checklist in How To Use to add a clear validation checkpoint.
Trim the inline platform-directory enumeration in the scope list to the top few examples and defer the full list to references/platform-files/platform-map.md, which already covers it.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with Trellis-specific facts Claude could not know (config keys like 'channel.worker_guard.idle_timeout', archive-slug collision rules, lockstep dual-package releases) — no padding explaining generic concepts, so it is well above anchors 2-3. It falls short of anchor 5 because of minor repetition and bloat: the lockstep-package rule appears twice ('both @mindfoldhq/trellis and @mindfoldhq/trellis-core ship as published packages sharing one version' at the top and again nearly verbatim in Do Not), and the platform-directory enumeration in the scope list is a long inline run-on that could be delegated to platform-map.md. | 4 / 5 |
Actionability | Fully concrete, copy-paste-ready guidance throughout: exact commands ('task.py create --parent <slug>', 'task.py add-subtask <parent> <child>', 'task.py list-context <task>', 'trellis channel spawn --agent <name>', 'trellis workflow --template <id>', 'trellis mem list | search | context | extract | projects'), exact file paths ('~/.trellis/channels/<project>/<channel>/events.jsonl', '.trellis/.template-hashes.json'), and exact config keys ('hooks.after_create / after_start / after_finish / after_archive', 'codex.dispatch_mode: inline | sub-agent'). This is an instruction/config skill and the guidance is fully actionable, matching anchor 5. | 5 / 5 |
Workflow Clarity | The 'How To Use' section gives a clear conditional 5-step sequence (read overview first, branch by request type, then 'Before editing, read the actual files in the user project and treat local content as authoritative'), and the Do Not section supplies real guardrails: 'check .trellis/.template-hashes.json first and prefer .new sidecar files over destructive overwrites' and 'cross-check against the local .trellis/config.yaml and the installed CLI's trellis --help before claiming a knob exists'. It stops short of anchor 5 because there is no explicit validate-then-retry feedback loop and the pre-edit checklist is spread across sections rather than consolidated as a checklist. | 4 / 5 |
Progressive Disclosure | Exemplary structure: the SKILL.md is an overview with a guided reading order ('Read references/local-architecture/overview.md first'), and every one of the 23 cited reference files exists in the bundle and is one level deep — no nested references. Each reference carries a one-line description ('references/local-architecture/multi-agent-channel.md: trellis channel subcommands, project-scoped event store, forum/thread channels...'), grouped under three clear headings. This matches anchor 5 (clear overview with well-signaled one-level-deep references, easy navigation). | 5 / 5 |
Total | 18 / 20 Passed |