KEEP IT HANDS-ON functional ~ tested 2026-09-26
// sandboxed in macOS (host) Β· aarch64 Β·install log Β· why not fully functional: Marketplace install fidelity, plugin validation, structural contracts, cross-references, and every checkable content claim (pattern counts, mode tables, runtime identifiers in the Claude Code 2.1.152 binary, homepage and v1-repo liveness) were verified on host. The core deliverable, generating an agent team from one sentence, is an in-agent prompt contract that requires a live Claude Code session, which this review did not run; the v1 A/B quality numbers (49.5 to 79.3) are author-run on the deprecated v1 architecture and the README labels them as such. Β·functional log

revfactory/harness

by revfactory (robin) · https://github.com/revfactory/harness · Apache-2.0 · vv2.1.0 (manifest strings only; no git tag) · updated 2026-09-26
Don't install your next skill blind. One email a week: what's worth installing.

The team-architecture factory: one sentence in, an agent team and its skill set out, plus an evolution loop that feeds usage back.

⚙ ⚙ ⚙ ⚙ ⚙ 4 / 5
quality 4/5
documentation 4/5
setup 5/5
value 4/5
ecosystem fit 3/5
// bottom line

The install path is verified byte-exact through the real Claude marketplace, every structural contract we could measure holds (strict YAML, 9 of 9 references, 5-surface version sync), and the methodology content is the most disciplined team-design curriculum GearScope has read. The costs: the entire skill contract is Korean-only, strict plugin validation fails on a missing marketplace description, and there are zero tags or releases at 9,066 stars. Claude Code users who design per-project agent teams should install it; everyone else should read the reference set and move on.

Get the skill digest The best and the broken from 100+ hands-on tests. One email a week.
install via Claude Code marketplace (README commands)
$claude plugin marketplace add revfactory/harness && claude plugin install harness@harness-marketplace
or universal installer, works but undocumented
$npx -y skills add revfactory/harness --agent hermes-agent --copy -y
or manual copy (README)
$cp -r skills/harness ~/.claude/skills/harness && cp -r skills/evolve ~/.claude/skills/harness-evolve

install if

  • Claude Code operators who build per-project subagent teams. The 8-phase factory replaces ad-hoc team prompting with a repeatable, auditable process, and the evolve loop turns usage feedback into next-generation harnesses.
  • Korean-speaking teams. Trigger phrases, failure FAQs, and the entire reference set are first-class in Korean, which is rare in this tier.
  • Skill authors in any language. The skill-writing-guide (259 lines) and skill-testing-guide (270 lines) work as a standalone authoring curriculum, including with-skill vs baseline A/B methodology and trigger-boundary testing.

What It Does

revfactory/harness is a Claude Code plugin that acts as a team-architecture factory: one sentence ("build a harness for this project") sends it through an 8-phase workflow that audits existing setup, analyzes the domain, picks an execution mode, writes agent definition files to .claude/agents/, generates skills to .claude/skills/, wires an orchestrator skill, and records a pointer plus change log in CLAUDE.md. It ships as two doc-only skills: harness (a 405-line Korean SKILL.md routing to 9 reference files, 1,966 lines) and evolve (89 lines), which captures the delta between initial and current harness state and generalizes feedback into the next generation. The target user is a Claude Code operator who sets up domain-specific subagent teams per project. The repo holds 9,066 stars and 1,284 forks after six months, and v2.1.0 was released the same day as this review.

The Good

The documented install path works exactly as written, byte for byte. Running the README's two marketplace commands in an isolated HOME added the marketplace, installed harness v2.1.0 as an enabled user-scope plugin, and landed both SKILL.md files byte-identical to the clone in both the plugin cache and the marketplace mirror. As a bonus, the undocumented universal path (npx skills add, hermes-agent) also landed both skills byte-identical with a skills-lock.json receipt. The always-loaded weight is two descriptions totaling 594 characters, which is light for what the pack does.

Self-consistency discipline that survives measurement. A strict PyYAML parse of both SKILL.md files passes, every frontmatter name matches its directory, all 9 referenced files exist with zero orphans, all 5 local links in the README pair and docs resolve, and version 2.1.0 is synchronized across five surfaces (plugin.json, marketplace.json, both README badges, CHANGELOG head). The factory imposes a sub-500-line limit on the SKILL.md files it generates, and its own 405-line contract respects that limit. For comparison, Unity's official pack shipped 4 of 31 skills with unparseable frontmatter.

The changelog is the most honest document in this tier. Across 150 lines and eight releases since 2026-03-27, every entry is issue-referenced and dated. The 2.0.0 entry states plainly that the /harness:evolve skill v1's README advertised never existed, and that v1 orchestrators built on the removed TeamCreate API quietly degraded to single-agent execution. The README frames the v1 A/B result (mean quality 49.5 to 79.3, 15 of 15 win rate) as author-measured numbers and tells readers to run their own pilot. That framing is rare and worth crediting.

The methodology content earns its references. Mode selection is keyed to whether control flow can be expressed in code rather than to team size, with a decision tree. The quality catalog is concrete: adversarial verification that passes only items a majority confirms, judge panels over N drafts, loop-until-dry with a seen-set to stop repeats, multi-criteria sweeps, and a completeness critic as the final agent. The model policy assigns fable, opus, or sonnet per agent by task complexity, duration, autonomy, and latency, and forbids blanket pins. The error policy distinguishes retryable failures from terminal ones (rate limits get reported, not retried) and forbids the orchestrator from inventing a stalled agent's reasoning when filling gaps.

The Bad

Strict plugin validation fails on a one-field gap. claude plugin validate --strict . exits 1 with "No marketplace description provided"; non-strict validation passes with the same warning. The marketplace.json carries a per-plugin description but no top-level marketplace description. This is the same cosmetic class as the dotnet/skills issue #1087 and is a one-line fix.

Zero tags and zero releases at 9,066 stars, and the repo's own changelog records a tagging plan that never executed. The v1.2.1 entry (2026-04-18) wrote a plan to retro-tag v1.0.0 through v1.2.0. The GitHub API returns an empty tag list today, so consumers cannot pin 2.1.0 and marketplace installs track an unpinned main.

The skill contract is Korean-only. The 405-line harness SKILL.md and all 9 references (1,966 lines) are written in Korean; only the README pair and plugin.json descriptions are bilingual. The skill explicitly instructs the agent to write generated harnesses in the user's conversation language (a rule added in 2.1.0 via issue #28), and frontier models read Korean instructions well, but an operator who does not read Korean cannot audit the 2,371 lines of contract being installed. This is an ecosystem-reach cost, not a quality defect.

A 45MB installed footprint for 184KB of skill content. The plugin cache and the marketplace mirror each land the full repo including the two root demo GIFs (7.4MB and 7.1MB, stored twice). The 2.0.0 rebuild removed 9MB of marketing PNGs and a landing page but kept the reels. Half of the 30MB clone is the two GIFs.

Three names for the evolve skill across surfaces. The README's copy command installs it as harness-evolve, the plugin and npx skills add surfaces expose it as evolve, and the harness SKILL.md prose refers to harness:evolve. The skills.sh registry indexes only the harness entry (204 installs), so name-based discovery misses evolve entirely.

Smoke Test Results

Host-based runs on macOS 27.0 (aarch64), Node 24.13.1, Claude Code 2.1.152 (the sbx microVM daemon is down pending Docker, so the documented host fallback was used). The review ran within hours of the v2.1.0 release commit (2026-09-26 09:01 UTC). Logs are in the sandbox folder.

Structural validation

$ find skills -name SKILL.md | wc -l && find skills -path "*/references/*.md" | wc -l
βœ… 2 SKILL.md + 9 references (1,966 reference lines)
$ strict PyYAML parse of both SKILL.md frontmatter blocks
βœ… both parse; names match directories; descriptions 313 and 281 chars
$ harness SKILL.md body length vs the sub-500-line rule it imposes on generated skills
βœ… 405 lines, under its own limit
$ SKILL.md to references cross-check
βœ… all 9 referenced files exist, no orphan files on disk
$ local link check across README, README_KO, docs, CONTRIBUTING (fence-stripped, relative to linking file)
βœ… all 5 links resolve
$ version consistency: plugin.json, marketplace.json, README EN badge, README KO badge, CHANGELOG head
βœ… 2.1.0 on all five surfaces
$ grep -rn "TeamCreate|TeamDelete|EXPERIMENTAL_AGENT_TEAMS" skills/
βœ… 13 mentions, all intentional (v1 detection and migration mapping; zero v1 syntax in live templates)
$ claude plugin validate --strict .
❌ rc=1, "No marketplace description provided" (warning treated as error under --strict; non-strict passes)
$ LICENSE file present
βœ… Apache-2.0, 10,756 bytes

Pass rate: 8 of 9. The single failure is the real missing marketplace description; an initial broken-link FAIL turned out to be my checker resolving a docs-relative link against the repo root, and the run was re-executed with the fix so the published log matches.

Full structure log β†’

Run A. Install probes, fresh isolated HOME

$ claude plugin marketplace add revfactory/harness
βœ… rc=0, harness-marketplace added from GitHub
$ claude plugin install harness@harness-marketplace
βœ… rc=0, v2.1.0 installed, enabled, scope user
$ cmp landed skills/harness/SKILL.md vs clone (plugin cache and marketplace mirror)
βœ… BYTE-IDENTICAL in both locations
$ cmp landed skills/evolve/SKILL.md vs clone (both locations)
βœ… BYTE-IDENTICAL in both locations
$ du of the installed plugin tree
βœ… 45MB total; demo GIFs land twice (7.4M and 7.1M in cache, again in mirror) for 184KB of skill payload
$ npx -y skills add <clone> --agent hermes-agent --copy -y
βœ… both skills land byte-identical at ./.hermes/skills with skills-lock.json written

Pass rate: 6 of 6. The README's marketplace flow and an undocumented universal-installer path both install cleanly.

Full install log β†’

Run C. Functional verification (documented claims vs reality)

$ six team-architecture patterns documented?
βœ… named sections: Pipeline, Fan-out/Fan-in, Expert Pool, Producer-Reviewer, Supervisor, Hierarchical Delegation
$ quality-pattern catalog present?
βœ… adversarial verification (majority-confirm), judge panels, loop-until-dry with seen-set, multi-criteria sweeps, completeness critic
$ three execution modes with selection criteria?
βœ… mode table keyed to control-flow determinism, plus a decision tree in execution-modes.md
$ v2 runtime premise vs the installed Claude Code 2.1.152 binary (string-table grep)
βœ… SendMessage 57, TaskCreate 32, TaskUpdate 27, resumeFromRunId 18, agentType 439 occurrences; TeamCreate strings persist at 9 and cannot be classified statically
$ external surfaces live?
βœ… revfactory.github.io/harness 200, animation site 200, v1 repo claude-code-harness 200
$ README EN/KO parity?
βœ… 125 lines each, same section count

Functional pass rate: 6 of 6. One method correction during the run: the first premise check grepped a directory for *.js files, but Claude Code 2.1.152 is a single 204MB Mach-O binary, so the corrected run greps the binary directly and the published log is the corrected one.

Full functional log β†’

What the runs tell you

The pack installs exactly as documented and every static contract holds, with one cosmetic validation gap and no release engineering. What no static run can confirm is the factory's output quality: generating an agent team from a sentence happens inside a live Claude Code session, and the only published efficacy numbers are the author's own v1 measurements. Treat the machinery as verified and the outcome claims as promising but unproven.

Setup Walkthrough

  1. In Claude Code, run /plugin marketplace add revfactory/harness.
  2. Run /plugin install harness@harness-marketplace. Verify with /plugin list (harness v2.1.0, enabled).
  3. Trigger it with a domain sentence: "build a harness for this project" or the Korean trigger phrases listed in the SKILL.md description. Expect a Phase 0 audit report, a mode and architecture proposal, then agent and skill files under .claude/.
  4. After real usage, "evolve the harness with this feedback" routes to the evolve skill, which generalizes feedback and appends to the CLAUDE.md change log.
  5. Watch cost: the quickstart warns that multi-agent runs can consume tens to hundreds of thousands of tokens per task, and generated orchestrators keep default scale small unless you ask for a thorough sweep or pass a budget like "+500k".

Post-install notes: the strict-validation warning is cosmetic and does not block the verified install flow. The quickstart asks for "latest Claude Code" without stating a version floor; the runtime primitives the contract references are all present in the 2.1.152 binary we checked.

Alternatives

  1. obra/superpowers - a general workflow-skill framework (reviewed 4/5) with breadth across brainstorming and planning; it does not generate domain-specific agent teams.
  2. yusufkaraaslan/Skill_Seekers - converts existing source documents into skills (reviewed 4/5); use it when the knowledge already exists and you want extraction rather than team design.
  3. disler/super-simple-software-factory - a deterministic SDLC factory with typed envelopes and SQLite event streams (reviewed 4.5/5); single-process discipline rather than per-domain team architecture.
  4. levi-qiao/octopus-skill - long-horizon orchestration as a markdown library (reviewed 4.5/5); it structures one durable task graph where harness designs the team that would run such work.
// review provenance
reviewed by
GearScope
tested
2026-09-26 · macOS (Apple Silicon)
last verified
2026-09-26
depth
HANDS-ON
sponsorship
none, ever
// share this review
// feedback
was this review helpful?

Don't install your next skill blind.

Every week: the shortlist of skills worth installing β€” and the ones to skip β€” from 100+ hands-on tests. No spam, no affiliate links.