菜单
Reference contract for creating, editing, or reviewing Kadmon Harness agent files (`.claude/agents/*.md`). Covers the mandatory/recommended/optional section taxonomy, the opus-vs-sonnet model decision tree, K-first naming convention, anti-patterns enforced by `/medik` Check
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Freeze the task list and project state to disk when a session ends, and thaw it back when the next one starts — the cryo-mirror of /kompact. /kompact compacts the context window *within* one session (you stay, its summary lives on in-thread); kontinuum crosses a hard session boundary — a /exit into a brand-new session in a fresh process, where the conversation thread breaks and an on-disk note is the only channel that survives. One trigger, two modes auto-detected by whether SESSION-HANDOFF.md exists at the repo root: FREEZE / SAVE (write the note on the way out, anchored to the exact commit sha) or THAW / RESTORE (git pull --ff-only FIRST, print a drift report of what landed while the session was frozen, reconcile WORK_COORDINATION.md claims when the repo has one, re-pin the OPEN tasks, print the summary, delete the note only once the restore succeeded). Use this skill whenever the user is about to /exit and start a new session, or says "kontinuum", "freeze the session", "congela la sesion antes de salir", "
Track lifecycle for multi-developer parallel work coordinated through WORK_COORDINATION.md. Two modes — /sprint claims a track + branches from main against the Next planned table; /sprint close runs the verification gate, opens a PR with a structured body, auto-merges when eligible, then post-merge syncs WORK_COORDINATION and flips the plan-NNN frontmatter status. Use this skill whenever someone says start plan, claim track, open track, ship plan, close plan, ready for PR, post-merge sync, flip track done, sprint close, or partial ship — even when they do not name the skill. First use in a project: instantiate WORK_COORDINATION.template.md (same directory) at the repo root.
Rewrite Hebrew marketing, landing, and UI copy so it reads as native Israeli writing instead of machine-translated Hebrew. Fixes the two things that instantly mark Hebrew copy as translated: gendered second-person address (the default-masculine אתה trap that excludes half your customers) and literal English calques. Produces a native-quality, gender-safe, professional-warm draft. Use this whenever writing, reviewing, localizing, or polishing ANY customer-facing Hebrew — a landing page, hero copy, a marketing site, UI microcopy, a bot or voice persona, ad or WhatsApp copy — and especially when the Hebrew was translated from English/Spanish, when someone says the Hebrew "sounds translated / stiff / generic / off", when it mixes masculine and plural address inconsistently, or when you need Hebrew that works for both male and female readers. Trigger even if the user only says "fix the Hebrew" or "make the Hebrew sound native". It raises the floor to ~native; it does not replace a final native-speaker glance on sh
De-slop marketing, UX, and landing copy: strip the "tells" that make prose read as AI-generated (em-dash overuse, "not just X but Y", rule-of-three triads, slop vocabulary like seamlessly / robust / elevate / unlock / delve, hedging, and generic openers) and tighten it to concrete, human, on-brand writing. Use this whenever polishing or reviewing ANY customer-facing prose — landing pages, hero copy, marketing sites, product emails, UX microcopy, app-store descriptions, ad copy, social posts — and especially when the user says "make this sound human", "less AI", "punchier", "tighter", "this feels generic", or "clean up the copy", even if they never say the word "de-slop". Works per-language (English, Spanish, and others). It is a quality-and-voice tool, NOT an AI-detector-evasion tool.
Offer the user an explicit choice of response depth (25% / 50% / 75% / 100%) before answering, with token estimates based on prompt complexity. Use this skill whenever the user says "token budget", "token count", "response length", "answer depth", "short version", "tldr", "brief answer", "detailed answer", "exhaustive answer", "cuántos tokens", "ahorrar tokens", "respuesta corta", "dame la versión corta", "responde al 50%", or clearly asks to control how much context/response to consume. Do NOT trigger when "token" refers to auth/JWT/payment tokens, when the user has already set a depth this session (maintain it silently), or when the answer is trivially one line.
Scan all `.claude/skills/` to extract cross-cutting principles that appear in 2+ skills, and distill them into rules — append to existing rule sections, revise outdated content, or create new rule files. Use this skill whenever the user says "extract rules from skills", "distill patterns", "update rules", "pull common principles", "what's repeated across skills", or periodically (monthly or after adding 5+ new skills) to keep `rules/` in sync with what the skills actually teach. Applies the "deterministic collection + LLM judgment" pattern — scripts collect facts exhaustively, then judgment produces verdicts.
| name | agent-authoring |
| description | Reference contract for creating, editing, or reviewing Kadmon Harness agent files (`.claude/agents/*.md`). Covers the mandatory/recommended/optional section taxonomy, the opus-vs-sonnet model decision tree, K-first naming convention, anti-patterns enforced by `/medik` Check |
| requires_tools | [] |
Authoritative reference for the agent contract used in the Kadmon Harness. The orchestrator and kody (during /chekpoint) consult this skill when creating, editing, or reviewing agent files. Content here was previously inlined in .claude/rules/common/agents.md and moved to on-demand loading 2026-04-26 to reduce auto-load footprint (see plan at ~/.claude/plans/se-puede-algo-entre-sunny-crayon.md).
/chekpoint verify → [ts-reviewer, py-reviewer, spektr, orakle] → kody (consolidate) → gate → commit
/skanner [arkonte (perf), kartograf (e2e)] → report
/abra-kdabra arkitect (if arch) → konstruct → feniks (if tdd)
/medik direct (16 checks) → [mekanik, kurator] (parallel analysis) → gate → repair → verify
/doks doks
/evolve alchemik
/nexus script execution
/kompact context compaction
/almanak almanak (single lookup)
/skavenger skavenger (single researcher, loads deep-research skill)
/forge unified instinct pipeline (preview gate)
/release release-hygiene pipeline (ADR-037); composes /doks for Layer-1 count sync
Not every skill is owned by an agent. Some skills are loaded directly by commands because the work is deterministic and indirection through an agent adds no value. These are not routing bugs — doks and audit tools should treat them as intentional.
| Skill | Loaded by | Why no agent |
|---|---|---|
verification-loop | /chekpoint (Phase 1) | Build → typecheck → lint → test is a deterministic sequence. A reviewer agent is already invoked in Phase 2; verification is the command's job. |
strategic-compact | /kompact | The compaction-decision matrix is deterministic enough that routing through an agent adds no value. The skill is loaded directly when the user (or another skill) needs the decision guide. |
skill-creator:skill-creator (plugin) | /evolve step 6 Generate (PROMOTE proposals only) | Alchemik proposes SkillSpec objects in a JSON fence but NEVER invokes the mutator or plugin itself (ADR-008 Q2). The /evolve command parses the fence, renders the approval gate, then invokes the skill-creator plugin for PROMOTE proposals or calls applyEvolveGenerate for non-skill types. This orchestration lives at command level to keep alchemik pure-analysis and to centralize collision handling. |
council | /abra-kdabra Step 1.5 (ambiguity detected) + main orchestrator ad-hoc | The skill's core mechanism is spawning 3 fresh sibling sub-agents via Task for anti-anchoring. Anthropic's observable pattern is orchestrator-driven spawning; nesting Task inside a sub-agent is neither documented nor endorsed. Keeping council at the command/orchestrator level matches that pattern and avoids granting Task to planner agents (konstruct, arkitect) that don't otherwise need it. Moved here 2026-04-23 after empirical test proved konstruct's declared ownership was unexecutable (lacked Task). |
fable-prompt | Operator, directly via the Skill tool — no command or agent owner | Not a command-level skill in the strict sense (there is no command to attach it to) but the same "intentional orphan" logic applies. Fable 5 prompt-authoring is a manual, ad-hoc step the operator takes before kicking off an autonomous run — it doesn't fit any deterministic pipeline. Added f912181 2026-07-06; documented here so future orphan audits recognize it as by-design rather than accidental. |
When adding new command-level skills, document the rationale here so future audits don't flag them as orphaned.
requires_tools: frontmatter (plan-029)Skills that invoke sub-agents, WebFetch, or other tools outside their owner agent's default grant SHOULD declare a requires_tools: field in frontmatter as a YAML flow ([Task]) or block list:
---
name: council
description: ...
requires_tools: [Task]
---
Purpose: /medik Check #14 (capability-alignment) compares each skill's requires_tools: against its owner agent's tools: field. Mismatches emit severity FAIL — a skill whose owner lacks a required tool is silently unexecutable (the council pre-2026-04-23 pattern). Seed adopters: council ([Task]), deep-research ([Task, WebFetch, WebSearch]).
Opt-in: when requires_tools: is missing, the check falls back to a heuristic body-scan (regex on Task(, WebFetch, Bash() and emits severity WARN with a suggestion to formalize. Explicit declaration always beats heuristic.
Every new agent file MUST derive from the canonical skeleton at .claude/agents/_TEMPLATE.md.example. The template uses the .md.example extension so Claude Code's plugin loader and the frontmatter linter (both .md-only scanners) skip it — this replaces the original underscore-prefix-only convention (ADR-017, 2026-04-19), which ADR-019 dogfood 2026-04-20 proved the Claude Code plugin loader does NOT respect. The linter at scripts/lib/lint-agent-frontmatter.ts still filters _-prefix as an additional safety net.
| # | Section | Reference agent |
|---|---|---|
| 1 | Frontmatter block (YAML with name, description, model, tools, memory, skills) | all 16 |
| 2 | Opening identity paragraph (first-person, after frontmatter) | arkitect, skavenger |
| 3 | ## Output Format (synonyms allowed: ## Plan Format, ## Review Output Format) | konstruct (Plan Format), kody (Review Output Format) |
| 4 | ## Memory block with canonical Read-before/Append-after wording | all 16 |
| # | Section | Reference agent |
|---|---|---|
| 5 | ## Expertise (bullet list of domain competences) | arkitect, konstruct |
| 6 | ## Workflow (synonyms: ## Review Workflow, ## Review Process, ## Planning Process, ## TDD Workflow, ## Analysis) | konstruct (Planning Process), feniks (TDD Workflow) |
| 7 | ## no_context Rule (short paragraph — how the agent refuses to invent evidence) | arkitect, konstruct, skavenger |
| Section | Include when | Reference agent |
|---|---|---|
## Security | Agent fetches/executes external content OR processes untrusted input | skavenger, spektr, almanak |
## Pipeline Contract (/command) | Agent is part of a /command chain with file-based hand-off | konstruct, feniks, kody |
## Examples | Workflow branches non-obvious (Routes, multiple modes) | skavenger |
## Red Flags | There is a taxonomy of BAD outputs to reject | arkitect, konstruct |
Artifact template (e.g. ## ADR Template) | Agent produces a specific artifact consumed downstream | arkitect (ADRs), konstruct (plans) |
## Execution Caps / ## Depth Modes / ## Self-Evaluation | Agent has multiple invocation modes or bounded resource budgets | skavenger |
typescript-reviewer, python-reviewer (language-targeted reviewers), doks (role spells itself out)./chekpoint on agent edits.Q1: Does the agent produce architectural decisions, critique code quality at scale,
or analyze security across systems?
YES -> opus. (arkitect, konstruct, spektr, alchemik, doks)
NO -> go to Q2.
Q2: Does the agent execute a well-defined workflow (review, debug, test, refactor,
research, fetch docs)?
YES -> sonnet.
NO -> revisit agent scope; it likely needs decomposition.
NEVER -> haiku (explicitly forbidden; rule enforced at review, not mechanically today).
Mechanically rejected today (lint-agent-frontmatter.ts, Check #8 of /medik):
skills: a, b, c — comma-separated scalar (ADR-012 requires YAML block list)..claude/skills/<name>.md instead of .claude/skills/<name>/SKILL.md (ADR-013).Flagged by kody in /chekpoint review (deferred to linter until Option B review 2026-10-19):
model: haiku or any value other than opus/sonnet.description: with an unquoted embedded colon.## Memory block.Skill listed in the agent's tools: frontmatter (plugins are command-level — see "Command-Level Skills" table above).scripts/lib/lint-agent-frontmatter.ts (Check #8 of /medik). The _-prefix filter skips template/example files; all other .md files in .claude/agents/ are checked./chekpoint (full tier when multiple agents change, lite tier for single-file edits — per the /chekpoint Tiers table in development-workflow.md).Source rule file: .claude/rules/common/agents.md — sections moved here 2026-04-26 to reduce auto-load footprint. See plan at ~/.claude/plans/se-puede-algo-entre-sunny-crayon.md for context.