← ki-agentic-harness

Skills and journeys

The skills are the bulk of the harness today. Most are governance skills — each holds a house standard and ships the universal EDUCATE / AUDIT / CONFORM / REFRESH modes (plus skill-specific ones), backed by a tracked references/sources.md. A smaller, growing set are process skills — lightweight skills that drive an action or lifecycle rather than holding a standard (ADR-KI-HARNESS-SKILLS-006); ki-recap, ki-next, ki-plan, ki-implement, ki-accept, ki-batch, and ki-delegate form the current repository-delivery set. This page is the map: what a skill is, how the set fits together, and the shape they share; the per-skill entries are in the catalogue.

If you know the outcome you want but not the skill name, start with choose a skill by outcome, which routes a plain-language goal to the smallest capability that serves it.

The repository delivery journey

The process skills coordinate around durable repository roadmap items. They do not form an inheritance hierarchy: each owns one part of the journey and passes a governed work item to the next.

The delivered single-item workflow is:

ki-recap
  → ki-next
    → ki-plan
      → Ready
        → ki-implement
          → In progress
            → Acceptance
              → ki-accept
                → Done
                  → repeat

• ki-recap closes a working session, identifies unfinished or newly transferred work, and offers grounded actions to ki-next.
• ki-next grounds the repository portfolio, processes inbound handoffs, and confirms the next item or small compatible set.
• ki-plan creates or enriches selected items through the explicit Ready gate, then stops.
• ki-implement takes one approved Ready item through In progress, bounded execution, verification, and an evidence-backed Acceptance packet, then stops.
• ki-accept reviews one item at Acceptance, requires human approval by default, moves approved work to Done, and prunes only exact confirmed retained records.
• ki-delegate supports planning or implementation when work benefits from bounded specialist agents. It classifies, assigns, sequences, and gates delegated units; it does not own roadmap state or confer authority.

ki-batch coordinates the same ownership boundaries across an explicit set:

Preparation
named candidates → ki-next + ki-plan → reviewed batch authorisation

Implementation
approved authorisation → repeated bounded ki-implement cycles → Acceptance

Batch preparation selects and shapes only the named candidates; it does not start delivery. Batch implementation then runs independent items in dependency order, parks ambiguity or out-of-authority decisions, and records a per-item ledger plus recap. Its normal endpoint is Acceptance. ki-accept remains the closure owner unless the batch authorisation expressly grants acceptance authority for named items; pruning always requires separate explicit confirmation.

The keystone skills sit underneath this journey. ki-bootstrap establishes the user environment, ki-repo owns the repository contract and declared governance coverage, and ki-skills governs the quality and structure of the capabilities used throughout.

What a skill is

A skill is a self-contained capability an agent can load on demand — a name and description that tell the agent when to reach for it, and a body of instructions for what to do once it does. Skills are governed by the ki-skills standard, which sets the authoring rules; as a consumer you don't need those rules, only that a skill's name is what you invoke it by (via trigger or /name).

A SKILL.md follows the open Agent Skills standard, so it is not Claude-Code-specific: a second runtime such as OpenAI Codex CLI discovers the same SKILL.md files from its own path (.agents/skills, vs Claude Code's .claude/skills), though it reads project instructions from AGENTS.md rather than CLAUDE.md.

Every skill here is a Knowledge Islands skill, shipped as part of this system, but the set has two kinds (ADR-KI-HARNESS-SKILLS-006). Most are governance skills — each holds a house standard and ships the universal EDUCATE / AUDIT / CONFORM / REFRESH modes plus a mechanical checker; what tells governance skills apart is not their kind but what each governs: a repository's structure, an operation, or the machine itself. A smaller set are process skills — they drive an action or lifecycle rather than holding a standard and are exempt from the governance shape and universal modes. The repository-delivery process set separates recap, selection, planning, implementation, acceptance, batching, and delegation into the boundaries above. Human-led repository review is the REVIEW mode of ki-repo, so it follows the declared repository governance capability rather than being installed as a separate process skill. ki-work selects the repository's forward-work adapter: Roadmap by default for a Project repository, or Streams for a Knowledge Base. Kind and physical domain are separate axes in the map below.

The Agent Skills standard is more general than this, though. A skill need not govern a standard at all — it could equally encode a standalone workflow (a review process, a release checklist, a research harness) or target one specific project or recurring task. The process kind is the first step into that territory, and the set is expected to grow further over time.

ki-self: local governance for local concerns

Every governed repository may carry a repo-local ki-self skill for concerns that are real in that repository but do not yet belong to a shared Knowledge Islands skill. It is a governance skill, not a process wrapper: it holds the repository's local standard for repeatable checks, semi-regular housekeeping, and repository-specific maintenance. Author it once at .agents/skills/ki-self/SKILL.md, commit it with the repository, and project .claude/skills/ki-self to that source when Claude Code is used. It is never installed into a user-wide skills directory and is deliberately absent from the harness's shared implication graph and catalogue.

Its universal modes apply at local scale:

• EDUCATE establishes the local-concerns contract when needed: the repo-local skill, a judgmental recurring-work ledger such as HOUSEKEEPING.md, and documented focused procedures for repeatable checks.
• AUDIT is the repository's local-housekeeping audit. With no narrower concern it reviews the whole local-concerns contract; audit <concern> is an additive focused diagnostic command.
• CONFORM runs AUDIT first, then proposes and—only after confirmation—applies the appropriate source-state remediation.
• REFRESH retires resolved concerns and promotes either a recurring cross-repository pattern into a named Knowledge Islands skill or multi-step change work into the repository roadmap.

ki-housekeeping-claude recognises this boundary: it governs accumulated Claude machine state, while ki-self governs the repository-local concerns that sit beside it. Skills by outcome routes the same territory by what you are trying to do rather than by family, and keeps ki-self distinct as repository-local governance that can promote a recurring pattern into a named shared skill.

The skill domains

The source tree groups capabilities into six semantic domains:

1. Agentic systems — ki-communication and ki-subagents: the shared capability types that equip an agent.
2. Change management — ki-work, its Roadmap, Streams, GitHub Issues, and Linear adapters, and the short lifecycle process skills.
3. Environment — portable ki-binding and ki-tokenomics; their -claude and -codex runtime adapters; and ki-housekeeping-claude.
4. Governance — ki-agora, ki-authoring, ki-checkpoint, ki-decision-records, ki-delegation, ki-engineering, ki-git, ki-guides, ki-specs, ki-trade, and ki-trades: reusable standards and instruments that cut across repository shapes.
5. Keystone — ki-bootstrap, ki-repo, and ki-skills: the installation, repository, and skill-quality contracts that hold the set together.
6. Repository structure — the explicit primary structures ki-repo-project and ki-repo-kb, with composable deltas for harnesses, MCPs, websites, specifications, tools, and Knowledge Base zones.

Interdependencies

The domains group the skills by concern. A second relationship runs across them: which governance capability each skill necessarily requires. Formal composition is declared in ki-depends-on: frontmatter; selecting a dependent runs its prerequisites first. A repository explicitly declares both the selected skill and every dependency. Textual list order has no meaning, and separately coverage-detected standards do not appear as dependency edges. The tree below mirrors the declarations rather than expanding coverage.

ki-bootstrap

ki-work
├─ ki-work-roadmap
├─ ki-work-github-issues
└─ ki-work-linear

ki-repo-kb
├─ ki-repo-kb-activities
├─ ki-repo-kb-live-artifacts
└─ ki-repo-kb-streams

ki-repo-website
├─ ki-repo-website-content
├─ ki-repo-website-app
└─ ki-repo-website-cloudflare

ki-repo-harness

ki-repo-mcp

ki-repo-plugins

ki-repo-specifications

ki-specs

ki-git

ki-housekeeping-claude

ki-tokenomics-claude
└─ ki-tokenomics

ki-tokenomics-codex
└─ ki-tokenomics

ki-tokenomics

ki-binding-claude
└─ ki-binding

ki-binding-codex
└─ ki-binding

ki-binding-chezmoi
├─ ki-binding
└─ ki-dotfiles-chezmoi
   └─ ki-authoring

ki-delegate

ki-accept

ki-batch

ki-engineering

ki-homebrew-tap

ki-implement

ki-next

ki-plan

ki-recap

ki-repo
├─ ki-authoring
└─ ki-git

ki-tools

The governance-skill shape

Governance skills share one layout, so a reader can move between them; the layout and modes are codified in the ki-skills rubric:

• references/standards-<topic>.md — the normative, quotable reference: what good looks like and why.
• references/rubric.md — the generated readable publication of the structured catalogue.
• references/sources.md — tracked provenance and refresh cadence where the standard depends on moving sources.
• scripts/rubric/items/index.ts — the canonical structured catalogue hosted by ki; its family modules carry deterministic checks and safe proposals.

The universal modes are AUDIT, CONFORM, EDUCATE, and REFRESH, with HELP as the safe explanation path and skill-specific modes where they fit. The verified ki host executes governance catalogues; skills do not carry compatibility runners.