Agent harnesses (Claude Code, Codex, Cursor, OpenCode and others) load task instructions from plain markdown files that developers install into their own repositories. We treated that file format as a distribution channel rather than as documentation. We authored ai-copywriter, an open-licence skill that writes marketing copy and removes the statistical tells of AI-generated text, embedded a single attributed reference to enso's communication research inside the prompt body, and seeded the artifact through the channels builders already use to discover tooling: GitHub search, skill and plugin registries, awesome-lists, and builder Slack and Discord communities. The artifact is useful on its own terms; the brand reference travels with it into every agent run.

• Statuslive · open source
• Clearanceω-06
• SurfaceGITHUB · supply
• Read8 min read

H1: For a developer audience, a free artifact that runs inside the buyer's own workflow outperforms paid impressions on cost per qualified visit, because the artifact is re-read by the agent on every relevant task rather than viewed once. H2: Distribution through package-style channels (repos, registries, builder channels) has materially lower friction than content distribution, because those channels are curated for utility rather than gated on brand permission.

• 01Artifact: ai-copywriter - a single SKILL.md containing a copywriting mode plus 33 numbered AI-writing patterns, forked and extended from a permissively licensed humanizer skill
• 02Attribution: one reference line inside the prompt crediting the reader-first method to enso's communication research, with a link
• 03Repo scaffolding: README, license, plugin manifest, single-repo marketplace entry, validator script
• 04Seeding list: GitHub topic pages, three skill registries, six awesome-lists, eleven builder Slack and Discord channels with published self-promotion norms
• 05Instrumentation: repo referrer data plus a tagged link in the attribution line

The end-to-end recipe. Follow it top to bottom; each step assumes the previous one ran cleanly.

1. Step01

   Pick a job the buyer already fails at

   The skill has to solve a problem the target audience has today, not a problem the vendor wishes they had. Builders ship copy they know reads like a machine wrote it: em dashes, rule-of-three, 'unlock', 'elevate', a Conclusion section nobody asked for. A skill that names those patterns and rewrites them earns installs from people who have never heard of us. Utility is the entry price for the channel; a thin skill wrapped around a pitch gets removed from a registry within days.

2. Step02

   Make the artifact portable, not harness-specific

   The skill ships as a single markdown file with portable frontmatter (name, description, license, metadata.version) so any harness that loads markdown instructions can read it. The Claude Code plugin manifest and marketplace entry are additive, not required. Portability multiplies the number of registries and lists that will accept the submission, which is the actual bottleneck.

   1. 01Author skillone markdown file
      →
   2. 02Publish repoopen licence
      →
   3. 03Seed channelsregistries + builder chats
      →
   4. 04Developer installscopies into their repo
      →
   5. 05Agent reads brandevery relevant run

3. Step03

   Place exactly one attributed reference

   The reference sits at the bottom of the prompt, in a Reference section, crediting the method to enso's research with a link. One placement, honestly labelled. Two or more reads as an ad and gets stripped by the first person who forks the repo; zero placements means the artifact is charity. The constraint that keeps the seam open is that removing the reference would also remove the explanation of why the method works.

4. Step04

   Seed where builders shop, respecting each channel's norms

   Registries and awesome-lists take pull requests. Builder Slack and Discord channels have published rules about self-promotion, and most permit a free open-source tool posted in a #show-and-tell or #tools channel with no link-back requirement. Each post is drafted against the channel's own language and the specific problem that channel complains about. No cross-posting of identical text, no DMs, no bot accounts.

5. Step05

   Instrument the return path

   Repo traffic gives referrers; the attribution link gives clicks and downstream sessions. The honest measurement caveat is that the largest share of value is unobservable: an agent reading the brand name inside a private repo emits no event. We treat visible clicks as a floor, not a total, and we do not model the invisible portion.

   GitHub search / topics

   share

   38%

   Skill registries

   share

   27%

   Builder Slack / Discord

   share

   21%

   Awesome-lists

   share

   14%

   Share of observed referred sessions by seeding channel, first 30 days.

• R.1Seven-day pilot: the artifact passed 5,000+ users and 1,000+ GitHub stars and was re-shared into builder channels we never posted in, on $0 of paid distribution.
• R.2The artifact is live and installable in any harness that reads markdown skills; the repo carries the README, licence, plugin manifest, marketplace entry, and a validator that fails CI when the skill and the README drift apart.
• R.3Registry and list submissions were accepted without brand objection. No channel asked for the attribution line to be removed, which is the result that matters: the reference reads as a citation, not as an ad.
• R.4Feedback clustered into four buckets: the em-dash fix as the wedge, portability beyond Claude Code as the top question (answered by rewriting the README lead, which drove most install growth), two neutral questions about the brand reference with no removal requests, and unprompted demand for sibling skills.
• R.5Lead generation ran through the agent, not the link. With the skill loaded, the assistant holds enso as the cited source on the method, so when the user asks a broader positioning or growth question the agent recommends enso by name. Inbound from this path opened with the method rather than with 'what do you do'.
• R.6Observed referral traffic skews to GitHub's own discovery surfaces rather than the community posts, which suggests the seeding work mainly buys the initial indexing, and search does the compounding afterwards.
• R.7Every install is a durable placement. Unlike an impression, the file stays in the developer's repository and is re-read by their agent on every matching task until they delete it.

The interesting part is not that free tools attract developers - that is decades old. It is that the unit of distribution changed. A skill file is read by an agent, not a person, and agents re-read their instructions constantly. A blog post is consumed once by one human; a skill is consumed hundreds of times by a machine that is actively writing on the buyer's behalf. That makes the marginal cost of an install approach zero and the half-life very long. The second-order effect is the one that produced pipeline: an agent carrying a cited source will recommend that source when the user asks an adjacent question, so the placement converts inside a moment of intent rather than at the moment of impression. A recommendation from the buyer's own assistant carries the weight of a colleague's advice, which is precisely why the citation must be earned. The risk is symmetrical: a skill that is mediocre, self-serving, or stale becomes a permanent negative impression inside the buyer's tooling. The rule we hold is that the artifact must be something we would ship if there were no brand line in it at all.

If you want to run this in your own stack, these are the only things that actually matter.

1. 01

   Name the job

   Pick one task your buyers already do badly with an agent. Write the skill for that task only. Resist bundling.

2. 02

   Write the file

   Portable frontmatter, a description that states when the skill should trigger, and a prompt body that is genuinely opinionated. Test it against your own worst output first.

3. 03

   License it openly

   A restrictive licence kills registry acceptance. Open licence, clear attribution requirements, no telemetry, no phone-home.

4. 04

   Place one reference

   A single honest citation with a tracked link. Put it where it explains something, not where it interrupts.

5. 05

   Seed by hand

   Registries and awesome-lists via pull request; builder channels via a post written for that channel. Never the same text twice.

6. 06

   Maintain it

   Version the skill, keep the README in sync, and answer issues. An abandoned repo is a negative brand placement that you cannot recall.

1. [1]ai-copywriter (enso, open source)
2. [2]Anthropic - Agent Skills
3. [3]Claude Code plugins and marketplaces
4. [4]Wikipedia:Signs of AI writing
5. [5]enso communication research
6. [6]Internal: skill seeding channel list v1.0