Skip to content

Latest commit

 

History

History
71 lines (49 loc) · 6.39 KB

File metadata and controls

71 lines (49 loc) · 6.39 KB

Landscape research

Why Kitbash exists, derived from what already exists and where it stops.

The field today

Viral prompt packs

  • Ponytail (~75k stars in weeks, MIT) — injects a "lazy senior dev" ruleset: a ladder of questions the agent must climb before writing code (does this need to exist? does it already exist? …). Measured ~54% less code, ~20% cheaper on real sessions. Ships adapters for a dozen assistants by hand-maintaining a copy per format.
  • Caveman — compresses agent output style to cut token spend; hook-based activation, intensity levels, session stats.

What they prove: single-behavior skills go viral; developers will install agent modifiers in huge numbers; measured claims (Ponytail's %) build trust. Where they stop: each is one hardcoded philosophy. Distribution is copy-paste or per-assistant plugin. No shared format, no tests, no versioned trust. Every one of them re-solves the "N assistants, N formats" problem by hand — that is the tell that infrastructure is missing.

Distribution directories

  • skills.sh (Vercel) — npx skills add owner/repo copies SKILL.md folders into 70+ agents; open directory with an install-count leaderboard. The strongest prior art on distribution and proof of demand at scale. Where it stops: it distributes the SKILL.md convention as-is. No manifest — no budgets, permissions, artifacts, dependencies, or gate semantics. Installs are unpinned copies: no lockfile, no diff-on-update, no integrity checking, no tests. Ranking is install counts, i.e. popularity. Distribution without engineering. Kitbash's answer is interop, not rivalry: a bare SKILL.md folder is KSF-minus-manifest, so kitbash install accepts any skills.sh skill directly — flagged as unmanifested, defaults applied. Their catalog is our funnel; our manifest, lock, and evals are the upgrade path.

Platform-native extension systems

  • Claude Skills / plugins — richest model: progressive disclosure via frontmatter, scripts, hooks, subagents, marketplaces. Claude-only.
  • Cursor Rules (.mdc, glob-scoped, auto/manual attach) — good scoping model. Cursor-only, no scripts, no tests.
  • Copilot instructions (repo-wide + path-scoped *.instructions.md) — GitHub-native. Static text only.
  • Windsurf workflows / Cline rules / Roo modes / Continue blocks / OpenHands microagents — each a partial re-invention: triggers, scoping, sometimes tool constraints, always locked in.
  • Codex CLI + AGENTS.mdAGENTS.md is the closest thing to a standard: plain markdown read by many agents. Lowest common denominator: no triggers, no scoping, no permissions, no budget, no composition. It's a README for robots, and that's all it can ever be.

What they prove: the capability set skills need is known — lazy loading, scoped triggers, script execution, tool permissions. Where they stop: every capability dies at the platform boundary. A team using three assistants maintains three divergent rule sets that drift apart.

Config conventions

  • Aider conventions files, dotfile rule stacks, CLAUDE.md/AGENTS.md hierarchies — accreting text files with no ownership, no measurement, and silent context cost on every request.

The gap map

Capability Prompt packs Platform systems skills.sh Kitbash
Cross-assistant distribution manual copies none yes — same file copied compiler + adapters (native idioms, visible degradation)
Versioned, pinned installs git clone at best marketplace-ish (one platform) unpinned copies lockfile + content hashes
Behavior testing none none none eval tiers
Trust / injection review none partial review none install-time diff, permissions manifest, signing
Context cost accounting none none none declared budgets, lint-enforced
Skill composition none ad-hoc chaining none typed artifacts + pipelines
Ranking stars stars install counts measurement badges
Durable repo knowledge none per-platform memory none lore layer

skills.sh solved distribution. Everything below the first row is still empty — that's the product: skills.sh distributes skills; Kitbash makes them engineering.

First principles: what developers actually want

Developers don't want "AI skills." They want: fewer review cycles, faster onboarding, fewer bugs, less repetitive work, better architecture decisions, less context switching. A skill is only worth installing if it moves one of those. Two consequences:

  1. Grounding beats prompting. A generic "write tests" skill is a commodity — the model already does that. Value comes from project-specific grounding (this team's review standards, this repo's invariants) and enforcement (gates that block, evals that measure). Hence lore and gates are core primitives, not add-ons.
  2. Trust is the adoption ceiling. Teams won't standardize on unreviewable prompt injections into their codebases. Hence lockfiles, diff-on-update, and permission manifests are v0 features, not v2.

Ideas considered and rejected

  • Personalities as a product — cosmetic; output style doesn't change outcomes. Allowed as config flavor, never a pillar.
  • Generic single-task skills (test generator, doc writer, commit writer) as standalone offerings — commodity; every frontier model does this unprompted. Only shipped when grounded in lore or enforced as gates.
  • A new agent/framework — competing with Claude Code/Cursor is a losing war and betrays the point: portability of the developer's investment.
  • "Adaptive prompts" / auto-optimizing magic — unfalsifiable marketing. Replaced by the measurable version: lore-driven context injection with evals.
  • Marketplace-first strategy — a store of untested prompts is a race to the bottom and a prompt-injection watering hole. Registry ships after lint + evals + hashing exist.

Sources