adapting-skill-creator-to-harnesses

name	adapting-skill-creator-to-harnesses
description	Use when using, creating, editing, testing, benchmarking, optimizing, or packaging skills with `skill-creator` outside Claude Code, or when `skill-creator` mentions Claude, claude-with-access-to-the-skill, `claude -p`, Claude.ai, Cowork, subagents, browser/viewer, trigger evals, description optimization, or packaging in a different agent harness.

Adapting Skill Creator To Harnesses

This skill extends skill-creator.

Read skill-creator first. Then apply these local modifications: treat Claude-specific operations as harness-specific examples, not universal instructions. Use the equivalent operation in the current harness. If no equivalent exists, tell the user what cannot be run and do not fall back to Claude-specific tooling unless the user clearly asked to test Claude behavior.

Harness Check

Before running evals or description optimization, identify:

• whether the current harness can spawn isolated agents
• whether spawned agents inherit the current model and reasoning settings
• whether timing and token metadata are available from spawned runs
• whether a browser, static HTML review file, or file presentation path is available
• whether the harness exposes a real skill-trigger test mechanism
• whether the existing skill-creator scripts test the current harness or a different one

Report any missing capability before substituting a weaker workflow.

Replacement Rules

Claude-specific wording	Current-harness meaning
Claude, Claude Code, or Claude.ai as the actor	The active agent or agent harness, unless the user specifically asks about Claude.
claude-with-access-to-the-skill	A spawned/isolated agent run that is explicitly given the skill path and task.
Baseline run "without skill"	A spawned/isolated agent run with the same prompt and files, without loading the skill.
claude -p, .claude/commands, and available_skills trigger tests	Claude Code trigger tests only. Replace with the current harness's native trigger test if one exists.
open <file> or webbrowser.open()	Use the harness browser, a static HTML file, or a clickable local path supported by the environment.
present_files	Use the current harness's file presentation mechanism if available; otherwise report the generated path.
Cowork or Claude.ai sections	Example capability profiles. Apply the same adaptation logic to the actual environment.

Behavioral Evals

For skill behavior tests, preserve the skill-creator structure: evals/evals.json, per-eval metadata, with_skill and baseline run directories, grading files, benchmark aggregation, and eval-viewer/generate_review.py.

Adapt execution to the current harness:

• In Codex, use subagents only when explicitly authorized by the user or active instructions. Omit model overrides so subagents inherit the current model; override reasoning effort only when requested.
• If the harness lacks subagents, run with-skill cases yourself and explain that baseline benchmarking is unavailable.
• If the harness lacks completion timing or token metadata, leave timing absent or record the limitation instead of fabricating it.
• If the harness lacks an interactive browser, generate a static review HTML file.
• Keep baseline and with-skill prompts otherwise identical so differences measure the skill, not the setup.
• To scaffold the standard workspace from evals/evals.json, run scripts/prepare_behavior_evals.py --skill-dir <skill-dir> --workspace <workspace>.

Trigger Evals

Trigger tests are harness-specific because each harness exposes skills differently.

• Use evals/trigger-evals.json as the shared input format for should-trigger and should-not-trigger queries.
• Use this skill's runner scripts for supported harnesses:
  • scripts/codex_trigger_eval.py --skill-dir <skill-dir> --workspace <workspace>
  • scripts/acpx_trigger_eval.py --skill-dir <skill-dir> --workspace <workspace>
  • scripts/claude_trigger_eval.py --skill-dir <skill-dir> --workspace <workspace>
  • scripts/cursor_trigger_eval.py --skill-dir <skill-dir> --workspace <workspace>
• Each trigger runner installs a temporary probe skill with the target skill's name and description. The probe body contains a unique sentinel and no real task instructions.
• Count a positive only when the harness provides evidence that the temporary skill body was loaded and the sentinel was emitted. Count a negative only when the sentinel is absent and the harness did not cross the runner's documented tool boundary.
• Use the strictest contract the harness can support. When a harness lacks native skill invocation events, use the runner's documented surrogate signal, such as body-read evidence from a project rule shim, and report the weaker contract in the summary.
• Run scripts.run_eval or scripts.run_loop only when testing Claude Code trigger behavior through the original skill-creator Claude-specific path or when the user explicitly asks to use Claude's CLI.
• For an unsupported harness, define the capability gap before substituting a weaker workflow. Do not count a trigger eval unless the runner can observe a credible skill-selection or body-load signal for that harness.
• Do not use Claude trigger results as evidence for Codex or another harness unless the user explicitly asked for a Claude-specific proxy result.

Harness Runner Contracts

Runner	Harness signal	Counted positive	Counted negative
codex_trigger_eval.py	Codex JSONL final message text from an isolated probe skill.	Sentinel appears in Codex output.	Sentinel is absent and the run exits cleanly.
acpx_trigger_eval.py	ACPX/Codex JSON stream plus recorded read of the temporary SKILL.md.	Temporary skill body is read and its sentinel appears.	Sentinel is absent and no disallowed tool calls occur.
claude_trigger_eval.py	Claude Code Skill tool call plus synthetic skill body delivery.	Matching Skill tool call occurs, the synthetic body is delivered, and the sentinel appears.	Sentinel is absent and no tool calls occur.
cursor_trigger_eval.py	Cursor Agent project AGENTS.md shim plus temporary .cursor/skills body read.	Temporary skill body is read and its sentinel appears.	Sentinel is absent and the temporary skill body is not read.

Description Optimization

Only run automated description optimization when both halves target the current harness:

• the trigger detector measures the current harness's actual skill-loading behavior
• the optimizer uses a model/harness the user accepts for rewriting the description

If the provided optimizer calls a different assistant, present it as optional cross-harness analysis, not as authoritative validation for the current harness.

Common Mistakes

Mistake	Fix
Editing the plugin-managed skill-creator copy	Keep the base immutable; apply this extension.
Running claude -p from Codex and calling it a Codex trigger eval	Say it is Claude Code-specific unless the user requested that proxy.
Forcing baseline benchmarks when no isolated baseline runner exists	Run qualitative with-skill checks and report the missing baseline capability.
Dropping the review viewer because the browser differs	Use --static or the harness's browser/file affordance.
Overriding subagent model or reasoning by habit	Inherit current settings unless the user requests a different model or effort.
Keeping reusable eval runners inside one evaluated skill	Put harness-generic runners in this extension skill and leave skill-specific eval data under the target skill.