Give every skill description a one-line lead the skill list can show in full #76

New issue

Closed

opened 2026-09-30 23:43:29 +01:00 by vnprc-agent · 0 comments

vnprc-agent commented

2026-09-30 23:43:29 +01:00

Contributor

Reformat the six public skills' frontmatter descriptions so the first sentence stands alone in at most 100 characters — the shape allod skill's list line can render without cutting off what a skill is.

Part of allod/tools#259 (memory-skills arc).

Primary goals:

One-line lead — every description opens with a sentence of at most 100 characters that says what the skill is; the allod skill list line then shows a complete identity for every skill, with no ellipsis mid-identity.
Triggers stay in frontmatter — the "Use when …" tail stays after the lead, so agent skill discovery (which reads the full frontmatter description) loses nothing; the text is reordered, not deleted.
Convention written down — the skills index in README.md states the shape so new skills inherit it.

Current state and specifics

The list line truncates at 100 characters with a word-boundary cut (briefLimit, cmd/allod/skill.go:159-175 on allod/tools origin/master). Every public description exceeds it, so every line ends in "..." and most lose the "Use when" clause that names when to reach for the skill:

Skill	Description length (chars)
delegate-implementation	709
fleet-diff	358
forge-groom	317
list-models	564
memory-backpass	129
pick-next-issue	295

Draft leads (implementation may refine; each must stay under 100 characters), with the rest of the current description preserved as the tail:

delegate-implementation: "Split an implementation arc across agents: one overseer, workers in worktrees, cold reviewers."
fleet-diff: "Gate a merge against which machines it actually rebuilds."
forge-groom: "Triage every open issue and pull request in the allod repos against what has actually landed."
list-models: "Find out which models each installed agent CLI declares and whether a named one answers."
memory-backpass: "Read recent session traces against the memory indexes and open one evidence-backed memory PR."
pick-next-issue: "Sweep every open issue in the repos this agent can push to and pick the next one to implement."

The smaller change deliberately chosen over alternatives: lengthening briefLimit only moves the cut point, and a CLI-side hard cap would reject vendored third-party skills passed via --from, so the fix is the content shape plus a written convention. If the shape drifts, the follow-up is a .hooks description check mirroring memory-budget — not part of this issue.

Scope

In scope: the six skills/*/SKILL.md frontmatter descriptions and the convention paragraph in README.md's skills index. Out of scope: any CLI change (allod/tools#259 owns behavior); the private fork's four descriptions (vnprc/agent-memory child of allod/tools#259); an enforcement hook.

Validation

allod skill (from ~/work/allod) prints every public skill's lead sentence in full — no line ends in "..." before the lead's final word.
Each rewritten description still carries its original "Use when" material in the tail; diff review confirms nothing was dropped, only reordered.
allod skill <name> output is unchanged apart from the reordered description.

Reformat the six public skills' frontmatter descriptions so the first sentence stands alone in at most 100 characters — the shape `allod skill`'s list line can render without cutting off what a skill is. Part of allod/tools#259 (memory-skills arc). Primary goals: - **One-line lead** — every description opens with a sentence of at most 100 characters that says what the skill is; the `allod skill` list line then shows a complete identity for every skill, with no ellipsis mid-identity. - **Triggers stay in frontmatter** — the "Use when …" tail stays after the lead, so agent skill discovery (which reads the full frontmatter description) loses nothing; the text is reordered, not deleted. - **Convention written down** — the skills index in `README.md` states the shape so new skills inherit it. ### Current state and specifics The list line truncates at 100 characters with a word-boundary cut (`briefLimit`, `cmd/allod/skill.go:159-175` on allod/tools origin/master). Every public description exceeds it, so every line ends in "..." and most lose the "Use when" clause that names when to reach for the skill: | Skill | Description length (chars) | |---|---| | delegate-implementation | 709 | | fleet-diff | 358 | | forge-groom | 317 | | list-models | 564 | | memory-backpass | 129 | | pick-next-issue | 295 | Draft leads (implementation may refine; each must stay under 100 characters), with the rest of the current description preserved as the tail: - delegate-implementation: "Split an implementation arc across agents: one overseer, workers in worktrees, cold reviewers." - fleet-diff: "Gate a merge against which machines it actually rebuilds." - forge-groom: "Triage every open issue and pull request in the allod repos against what has actually landed." - list-models: "Find out which models each installed agent CLI declares and whether a named one answers." - memory-backpass: "Read recent session traces against the memory indexes and open one evidence-backed memory PR." - pick-next-issue: "Sweep every open issue in the repos this agent can push to and pick the next one to implement." The smaller change deliberately chosen over alternatives: lengthening `briefLimit` only moves the cut point, and a CLI-side hard cap would reject vendored third-party skills passed via `--from`, so the fix is the content shape plus a written convention. If the shape drifts, the follow-up is a `.hooks` description check mirroring `memory-budget` — not part of this issue. ### Scope In scope: the six `skills/*/SKILL.md` frontmatter descriptions and the convention paragraph in `README.md`'s skills index. Out of scope: any CLI change (allod/tools#259 owns behavior); the private fork's four descriptions (vnprc/agent-memory child of allod/tools#259); an enforcement hook. ### Validation - `allod skill` (from `~/work/allod`) prints every public skill's lead sentence in full — no line ends in "..." before the lead's final word. - Each rewritten description still carries its original "Use when" material in the tail; diff review confirms nothing was dropped, only reordered. - `allod skill <name>` output is unchanged apart from the reordered description.

vnprc-agent referenced this issue from allod/tools

2026-09-30 23:44:07 +01:00

allod skill: derive the default skill sources from the registry's memory flag #259