ai-native-toolkit

An AI contributor is always the new hire. Make the codebase supply the tenure.

A Claude Code plugin, and standalone skills for any AI assistant, built so a fresh agent works like an engineer who has been in the org eighteen months: a navigable map, load-bearing contracts made explicit, and guardrails that make the wrong change hard.

Install in Claude Code

Run these inside a Claude Code session, not a shell. /plugin is a Claude Code command.

Claude Code prompt
/plugin marketplace add https://github.com/bjcoombs/ai-native-toolkit
/plugin install ai-native-toolkit@ai-native-toolkit

Skills appear namespaced, for example /ai-native-toolkit:assess. Update with /plugin update ai-native-toolkit.

No Claude Code? Download the standalone skill ZIPs from the latest release and upload them to claude.ai, Claude Desktop, Cowork, or any assistant that reads the Agent Skills format.

The difference is context, not capability

Every agent session starts with an empty head and one narrow context window. The question is how much of a tenured engineer's implicit map the codebase has made explicit.

The map

Where things live, which contracts are load-bearing, where the minefields are. The aim is a codebase legible enough that the relevant slice fits one context window, with the agent's answers anchored to code a person can check in ten seconds.

The guardrails

When a contributor makes a mistake, ask what made it possible and what would make it impossible next time. Linters, architecture tests, CI gates, coverage and review bots protect the contributor by design, the way staged environments protect a new engineer.

Legibility you can trust, not omniscience you can't.

Skills

Invoke each one in Claude Code as /ai-native-toolkit:<name>. The standalone ZIPs fire on the same slash name or on a natural-language request.

Portable, also as standalone ZIPs

/assess
Scores a codebase's readiness for AI contributors against a 0-8 layered contract model, with a complexity hotspot treemap and a doc-navigability graph. Writes a report with the Top 3 Actions, naming files.
/huddle
Structured multi-perspective deliberation using Six Thinking Hats, sized by Fibonacci from solo to board.
/deslop
Detects and removes the telltale signs of AI writing, derived from Wikipedia's "Signs of AI writing". Runs as a silent gate or an explicit edit pass.
/skill-forge
Hardens a skill or agent instruction file through judge-panel refinement rounds until it clears a 3-tier promotion gate.
/semantic-compress
Shrinks an LLM-directed document while preserving what it does, gating whole-document distillation on an A/B behavioural-equivalence check.

Claude Code only

/ghsync
Bulk-clones and fast-forward syncs every GitHub repo you can access across an org or personal account.
/ghreport
Read-only org repo state report: open PRs, default-branch CI, security alerts and branch protection per repo.

Portable commands

/6hats
Solo Six Thinking Hats analysis: /huddle at team size 1.
/understand
Deep understanding mode (nemawashi): exhaustive context-gathering before action.

Workflow commands

/tm
Task Master orchestration: starts, reviews, or cleans up tasks based on current state.
/issues
Triages open GitHub issues, then runs the agent-ready ones to merge with Agent Teams.
/fix-pr
Iterates on a PR's CI failures and review comments until it is green.
/fix-develop
Fix loop for failing CI on the repo's default branch.

The workflow commands bake in one author's daily setup: a <repo>-main/ plus worktree/ layout, GitHub and the gh CLI, Task Master, review bots, and the Agent Teams flag. Read the adapting notes before relying on them elsewhere.

What /assess draws

Two views from one run on a real monorepo of about 650k lines, file paths sanitized: before, and after its maintainers worked through the report's Top 3 Actions. The analysis is conventional static analysis and git history, not the model, so the findings reproduce run to run. Open any image for the full interactive SVG.

Doc-navigability graph

Can an agent traverse the docs to the right place, and is what it finds still true?

Doc-navigability graph before the action sweep: 254 docs, 30% reachable from the entry point, 43 islands, 15 broken links.
Before. 30% reachable, 24% orphaned, 43 islands, 15 broken links.
Doc-navigability graph after the action sweep: 247 docs, 89% reachable from the entry point, 20 islands, 0 broken links.
After the Top 3 Actions. 89% reachable, 11% orphaned, 20 islands, no broken links.
Reachable from entry
30% 89%
Orphaned docs
24% 11%
Disconnected islands
43 20
Broken links
15 0

Complexity hotspot heatmap

Which files are riskiest to change next week? The heatmap stays the same repo's complexity profile throughout: it shows which territory is dangerous, while the graph shows whether the map is honest.

Complexity hotspot heatmap before the action sweep.
Before. Size is estimated tokens, hue is complexity, saturation is churn.
Complexity hotspot heatmap for the same repo after the action sweep.
After. The same repo, same encoding.

Reading the doc graph

  • Position is reachability. The centre is the entry doc, rings are link distance from it, the rim is unreachable, and a dashed ring marks an orphan.
  • Colour is staleness. Vivid red is a frozen doc beside churning code: a lying map.
  • Size is file length. Solid edges are links, dotted edges are references.

Reading the heatmap

  • Size is estimated tokens (about chars/4, the context-window unit). Lines of code are in the tooltip.
  • Hue is cyclomatic complexity, red for high. Saturation is recent git churn, vivid for active.
  • The vivid-red blocks are what an agent or a person is most likely to break next week. Both views use the colour-blind-safe OrRd ramp, with no red-green.

Elsewhere

  • Huddle explainerA narrated visual walkthrough of how /huddle sizes a team and seats its hats.
  • AI-Readiness Assess GateThe /assess deterministic core as a GitHub Action. Runs on every PR with no model and no API keys; warn-only unless .assess/config.toml opts in to failing.
  • Source on GitHubThe README, the Map of Content, and every skill's SKILL.md.
  • Latest releaseRelease notes and the link to that version's standalone skill ZIPs.