| name | writing-marketing-readme |
| description | Use when writing or revising the README, landing text, release announcements, or any outward-facing text meant to attract new users. Evidence-backed rules for converting a technical hobbyist audience. |
Writing marketing text for a technical hobbyist audience
These rules come from a research pass (2026-07). Sourcing, stated honestly: the correlational README/popularity study arXiv 2206.10772 backs the structural basics (a one-line purpose, usage and install sections, images, lists, license, and contribution/reference sections all correlate with repo popularity; correlation, not causation). PostHog's own developer-marketing writing backs the anti-hype and specificity rules. A third-party analysis of Tailscale's Hacker News launches (markepear.dev) backs the peer-voice tone claims. The READMEs of ripgrep and fzf are the concrete pattern examples (fzf for hero plus early screenshot, ripgrep for honest limitations and data-driven comparisons); OrcaSlicer is an example only for community-terminology and feature-list conventions. Rules not traceable to those sources are marked as heuristics. These rules apply to the README and any outward-facing text; in-app text is governed by CLAUDE.md rule 7 instead.
The register: a knowledgeable peer, not a manual and not an ad
The voice that converts technical hobbyists is a third thing between neutral manual prose and marketing copy: a practitioner talking directly to another practitioner. Here is the problem, here is what this does about it.
- First person is an asset when it carries a real technical story (a real maker who hit a real problem). It becomes a liability the moment it turns into persuasion or self-congratulation.
- Zero hype: no superlatives, no "revolutionary", "effortless", "blazing", "magic". Developers detect marketing spin instantly and find it patronizing. Assume the reader is smart.
- Specifics carry the weight that adjectives cannot: exact numbers ("differs by 0.05% in X"), exact commands, exact method names ("gradient centroid sub-pixel edge estimator", "Taubin circle fit"). Naming the established method is credibility signaling.
- Section split: the hero line and the "why this exists" section may use peer voice and first person. Everything below (how-to steps, how it works, requirements, limitations) is plain technical prose in the same register as the in-app text.
Structure (in order)
- Hero: one factual spec-like sentence stating function, audience, and payoff mechanism. No slogan. The visitor decides in seconds whether to keep reading.
- Badges: build status, license, version. Near-universal in the example READMEs (ripgrep, fzf) and a cheap credibility signal.
- Proof above the fold: a screenshot or GIF of the tool actually working, plus the concrete output artifact (a code block with a real pasteable result, e.g.
SET_SKEW or M221 S...). Pattern example: fzf. The "visitors bounce without early proof" framing is practitioner heuristic, not study-backed, but images correlating with popularity is study-backed.
- CTA with friction-killing microcopy: the live link plus the genuine trust signals ("free, open source, runs entirely in your browser, nothing uploaded"). Stating local processing as differentiation is a heuristic (it is true and relevant for a scanning tool), not a case-study finding.
- Why / how it works: the measurement principle is itself the selling point when the mechanism is not self-evident. Explaining the method credibly is what separates the tool from eyeballed alternatives and earns the skeptical crowd's trust.
- What you get: name the exact outputs using community terminology (Klipper
SET_SKEW, Marlin M900 K, RRF M572, flow %). Specificity doubles as searchable keywords for the exact phrases hobbyists Google.
- Honest limitations: a plain section on requirements and what the tool cannot do (ripgrep's "Why shouldn't I use ripgrep?" pattern). Owning limits up front reads as expertise, not weakness, and pre-empts the skeptical forum reply.
- Quickstart / build / license, plus a contributing or reference section: plain technical prose. Contribution guidelines and reference/documentation sections correlate with higher popularity in the arXiv study; include at least a short pointer.
Proof beats claims
- Balanced factual comparisons against the status quo (calipers, eyeballed test prints) are strong trust devices; FUD destroys trust (PostHog).
- Heuristics (general practice, not sourced studies): real user quotes beat zero testimonials, but never fabricate and add social proof only as it accrues; do not build the pitch on vanity metrics such as star counts.
Scannability
Nobody reads top to bottom. Short sections with headers, short paragraphs, bolded key terms, one visible code or output block early. No prose walls.
Discipline carried over from the project rules
- Terminology: CLAUDE.md rule 7's terminology clause binds all outward-facing text with no exceptions (slicer and firmware setting names verbatim, one term per concept, no invented synonyms).
- Honesty: no setup-specific claim stated as a general truth, no capability overclaim. Every number quoted must be a real measured one.
- No em-dash (rule 6) and no AI attribution (rule 4) anywhere, including the README.