AGENT SKILL

goal-create

Create a session goal prompt or draft repository purpose. This skill never fires on its own; it must be invoked by name. Can synthesize a prompt from confirmed intent or spec and activate it through an available native host goal tool; does not operate the Loom workflow.

License 许可
MIT
所属仓库
kouko/loom-plugins
最近更新
2026-11-15
Stars
★ 0
Forks
0

01 前言 Front Matter

name
goal-create
version
0.2.0
description
Create a session goal prompt or draft repository purpose. This skill never fires on its own; it must be invoked by name. Can synthesize a prompt from confirmed intent or spec and activate it through an available native host goal tool; does not operate the Loom workflow.
user-invocable
true

02 资源文件 6

03 SKILL.md 正文

Goal Create

One skill, two named modes: SESSION and ARC. Which mode runs is chosen by what the user asks for — a goal for this run, or a purpose for the repository — never by the agent guessing from context.

<skill-dir> is this skill's folder: ${CLAUDE_SKILL_DIR} on Claude Code; on any other host, the directory that holds this SKILL.md.

SESSION mode

SESSION emits the four-field goal condition defined in references/goal-shape.md. Read that file for the field order and each field's definition before drafting; do not reconstruct the shape from memory.

Before drafting, read references/input-floor.md for the two input slots a goal is drafted from, the refusal rule when a slot is empty, the provenance tag every field must carry, and the citation boundary between quoting a recorded purpose and a user's own decision.

Once a draft is ready, run the mechanical floor:

python3 <skill-dir>/scripts/goal_lint.py <goal-file>

(or pipe the goal text on stdin with no argument). This checker enforces structure only — a field label present with content, a backticked command inside Verification, the character limit — never whether the prose actually reads as decidable, already-false, or free of a person. The bar in references/input-floor.md §4 stays a judgement call the agent makes each time; no script can pass or fail it on the agent's behalf.

The script exits 1 on any hard failure and 0 otherwise. On exit 1, the draft is rewritten to fix what it flagged and the checker is re-run — a draft is never shown to the user until it exits 0.

<!-- gate: goal-create.session-activation --> This skill ends after presenting the prompt and reporting host activation or recovery below. It does not invoke Loom stations or change their workflow rules. The prompt may refer its executor to an existing workflow; executing that prompt is outside this skill.

After exit 0, show the complete four-field condition in one copyable code block, then use the capability actually exposed by the current host. Invoking the named goal-create SESSION mode is explicit authorization to replace any active Goal with this condition, without a separate replacement confirmation. Confirmation is only for content: when the Outcome was inferred or materially changed from the user's words, use the host's confirmation path.

Codex activation

When the current session exposes create_goal, call it with the complete four-field condition as its objective. If create_goal is unavailable, say the Goal is not active and keep the complete prompt available for manual use; do not emulate activation through another process or tool. The authoritative schema and unfinished-Goal refusal are in OpenAI Codex's spec.rs and tool.rs. Call the Goal active only after host-provided success. On any failure, report the host's reason and say the new Goal is not active. In particular, when an unfinished Goal makes create_goal refuse replacement, preserve it and tell the user to run /goal clear, then run goal-create SESSION again.

Claude Code activation

/goal availability does not imply that ProposeGoal is exposed to the agent. Use the native proposal path only when the current session's tools actually include it. Otherwise go directly to the manual fallback below. Do not ask the user to change internal settings or feature flags to expose the tool, or infer availability from SDK types or installed package code.

When the interactive session exposes ProposeGoal, derive a proposal of at most 500 characters from the complete condition just shown. It preserves the Outcome, every behavior-changing Constraint, the decisive Verification, and the Stop-when bound, either inline or through an exact reference to that immediately preceding condition. A generic phrase such as "the goal above" is not an exact reference. If the faithful proposal cannot fit, skip ProposeGoal and use the manual fallback below.

Anthropic's published Claude Code package defines the condition, its 500-character maximum, ask_user default and direct-set path, and the result shape in sdk-tools.d.ts.

Call ProposeGoal with ask_user: false only when the user's own words state the exact Outcome. Otherwise use its confirmation path and report the Goal as pending, not active, until the host confirms kickoff. An existing Goal does not change this content-confirmation rule: the SESSION invocation already authorizes replacing it without a separate replacement confirmation.

When the current session does not expose ProposeGoal, or native activation returns a non-success result, say the Goal is not active and user action is required. Emit one copyable /goal <condition> command in a code block containing the complete condition, not the compact proposal, and disclose that submitting it replaces any active Goal. Do not invoke claude -p, inject keystrokes, launch another process, or install a custom Stop hook.

For either host, an attempted call, displayed prose, pending confirmation, or manual command is not activation evidence. Report active only from the host's successful result, and keep its recoverable reason visible otherwise. <!-- /gate -->

ARC mode

ARC drafts a Why and a Done when for the repository's purpose artifact at docs/loom/PURPOSE.md — see that file for its exact two-field format; this skill does not restate it. ARC never writes that file itself; the draft is only ever landed by the user's own confirmation.

ARC is conditional. When the repository has no docs/loom/ store and no docs/loom/PURPOSE.md file — nothing yet scaffolded to hold one — ARC reports itself not applicable, names the reason, and scaffolds nothing — creating the store is loom-init's job, not this skill's.

Invocation

This skill never fires on its own — the description above makes no auto-fire claim. It is named as an available option at exactly one point where the need for a goal is already visible: loom-workflow:handoff's Prepare mode, when a user closes a session without capturing an explicit goal. That surface names this skill as an option the user can invoke; it never invokes it. loom 1.0 deleted the other two offer sites with the skills that carried them.

When loom-design:capture-intent is already running for the same work, that station keeps discovery and this skill runs only after its intent exists, rather than competing for the same turn.

A user may also invoke this skill by name later in a run — after a confirmed intent or spec — to synthesize the prompt from those artifacts and attempt native host activation. That is the same manual invocation, not a new offer site: this skill is still never invoked by anything other than a user.

See also

  • references/goal-shape.md — the four-field goal shape, SESSION's SSOT.
  • references/input-floor.md — the two input slots, the refusal rule, the bar, and provenance tags.
  • scripts/goal_lint.py — the mechanical floor SESSION's draft is run through before it is presented.

04 同仓库更多 Skills