Harness comparison¶

A single canonical skill produces correct output for every harness — proper frontmatter schema, proper activation, proper aggregation, proper size budgets. This page shows what gets written and why.

Canonical input¶

# skill.yaml
id: pr-review
description: Apply before opening a PR. Tests, security, performance.
activation: agent
allowed-tools: [Read, Grep]

# SKILL.md
## Tests
- [ ] Unit tests added
## Security
- [ ] No secrets in diff

Rendered output per harness¶

Claude CodeCursorOpenCodeCodexGitHub Copilot

.claude/skills/pr-review/SKILL.md

---
name: pr-review
description: Apply before opening a PR. Tests, security, performance.
allowed-tools: [Read, Grep]
---

## Tests
- [ ] Unit tests added
## Security
- [ ] No secrets in diff

Why correct: Claude Code's progressive disclosure relies on name + description for activation; allowed-tools enforces the security boundary.

.cursor/rules/pr-review.mdc

---
description: Apply before opening a PR. Tests, security, performance.
---

## Tests
- [ ] Unit tests added
## Security
- [ ] No secrets in diff

Why correct: Cursor's "Agent Requested" mode activates rules whose description matches the current task — description-only frontmatter, no alwaysApply. (always skills get alwaysApply: true; globs skills get globs: + alwaysApply: false.) Verbatim Claude frontmatter would leave description valid but add unrecognized name/allowed-tools keys.

.opencode/skill/pr-review/SKILL.md

# pr-review

Apply before opening a PR. Tests, security, performance.

## Tests
- [ ] Unit tests added
## Security
- [ ] No secrets in diff

Why correct: OpenCode doesn't require frontmatter and prefers narrative markdown.

AGENTS.md (aggregated)

<!-- adeptability:begin id=pr-review hash=a1b2c3d4 -->
## Apply before opening a PR. Tests, security, performance.

## Tests
- [ ] Unit tests added
## Security
- [ ] No secrets in diff
<!-- adeptability:end id=pr-review -->

When multiple skills are enabled, they are packed by priority descending then byte size ascending, emitted in skill-id order. If the total exceeds 32 KiB (Codex's project_doc_max_bytes), the lowest-priority skills drop and a truncation manifest appears at the top:

<!-- adeptability: omitted 2 skill(s) due to 32KiB budget. Trim or split skills to fit. Dropped: legacy-style,old-runbook -->

Why correct: Codex has no per-skill concept. It reads one or more AGENTS.md files walking up from cwd. Naive copy-as-SKILL.md is invisible to Codex; the 32 KiB cap means quiet truncation if you don't manage it.

Copilot has no agent-requested mode, so this activation: agent skill is skipped (see the activation table below). A skill with activation: always renders into .github/instructions/always.instructions.md:

---
applyTo: "**"
---

<!-- adeptability:begin id=pr-review hash=a1b2c3d4 -->
## Apply before opening a PR. Tests, security, performance.

## Tests
- [ ] Unit tests added
## Security
- [ ] No secrets in diff
<!-- adeptability:end id=pr-review -->

Skills with activation: globs bucket into a per-glob file like .github/instructions/bucket-a63d8819677675ea.instructions.md with applyTo: "**/*.ts,**/*.tsx".

Why correct: Copilot uses applyTo glob match to decide when to inject the instructions. Without it, the file is loaded for the wrong files or ignored entirely.

Activation translation table¶

Canonical activation	Claude	Cursor	OpenCode	Codex	Copilot
`always`	description-only	`alwaysApply: true`	`# id` + desc + body	aggregated body	`applyTo: "**"`
`globs: [...]`	desc + glob hint	`globs:` + `alwaysApply:false`	`# id` + desc + body	aggregated body	bucketed by globs
`agent`	description-only	description-only	`# id` + desc + body	aggregated body	skipped (no agent mode)
`manual`	`disable-model-invocation:true`	desc + `@id` hint	`# id` + desc + body	aggregated body	skipped

Sidecars¶

Skills can bundle scripts/, references/, assets/. Where supported:

Claude Code: full sidecar tree preserved.
OpenCode: full sidecar tree preserved.
Cursor: dropped (single-file); a warning is reported on sync.
Codex: dropped (aggregator file model); a warning is reported on sync.
Copilot: dropped (aggregator file model); a warning is reported on sync.

Beyond the specialized five¶

Everything above describes the five specialized renderers. 50+ more agents (Windsurf, Gemini CLI, Cline, Continue, Roo, Goose, Junie, …) render through generic per-skill adapters that write SKILL.md into that agent's convention. Run adept harness list for the live registry, or add your own with a config-driven adapter.

Why this matters¶

Prior art treated multi-harness sync as a path-multiplexing problem (write one file into N paths). That breaks activation in Cursor and Copilot, produces nothing usable in Codex, and silently loses size budget. adept makes content fidelity to each harness's actual loader the headline guarantee — and proves it with golden files any contributor can inspect.