DeepWorkPlan — Author

Author and maintain the current repository's agent kit: reusable skills, agents, and
commands. Reason about the repo — never copy a generic kit. This sub-skill is also the executor of
the skills reconciliation the Final Review performs (DWP spec 2.3.0 §6.1) — and, in a plan created
under an earlier lifecycle, of that plan's separate "Skills & Agents Discovery" task.

---

Read scope

• Guide (essential — read for this flow): `../guide/skills-integration.md` §11 — how skills and agents are referenced from plans and tasks.
• Guide (conditional — read only when the trigger fires): none. `../guide/GUIDE.md` is the routing index.

Concepts

• Skill — a reusable, parameterized procedure invoked in-session (e.g. /fix-lint). Encodes
  "how to do X". Lives in <skills-dir>/<name>/SKILL.md.
• Agent — a specialized worker definition (role, model tier, tools, system prompt) dispatched to
  handle a class of tasks (e.g. reviewer, executor). Encodes "who does X". Lives in
  <agents-dir>/<name>.md.
• Command — a thin slash-command entry point that routes to a skill or agent. Logic stays in the
  skill/agent; the command is a delegator.

Need	Create
A repeatable procedure run in-session	Skill
A persistent role with its own model tier / tools	Agent
A shortcut to invoke a skill or agent	Command

Frontmatter fields (cross-agent)

The tool-agnostic core every harness reads: name (kebab-case, English, unique) and
description (one line, starts with a verb — harnesses use it for relevance scoring).
Everything else is a per-harness opt-in: keep only the keys the host repo's neighboring
files already use, and never require a key the repo does not use.

Field	Applies to	Meaning
name	skill / agent	kebab-case identifier; the registry/harness address
description	skill / agent	one line, starts with a verb; used for relevance scoring
version	skill	quoted SemVer ("1.0.0") — quote it so YAML keeps it a string
documentation_url	skill	canonical docs URL; replaces the legacy homepage (some harnesses treat homepage as a re-fetch source — never use it)
user-invocable	skill	true makes /<name> a slash command where the harness supports it
allowed-tools	skill	tool allowlist (e.g. Bash, Read, Grep, Glob, Edit, Write) where the harness enforces one
model	agent	abstract tier (light / standard / heavy) — never a vendor model ID (see "Model tiers")
tools	agent	the agent's tool list; keep it as narrow as the role allows

Do not invent harness features: if a field is not already used in this repo (or documented by the
target harness), leave it out rather than guessing a key name.

---

Step 0 — Detect the repo layout (do NOT assume)

Before authoring anything, discover where this repo keeps its kit. Do not hardcode any single repo's
conventions.

1. Find the agent root: look for .agents/, then .claude/ or .cursor/ (often symlinks to .agents/), then any
   AGENTS.md / CLAUDE.md at the repo root for documented paths.
2. Within it, locate skills/, agents/, commands/, and a catalog under docs/.
3. Inspect 1-2 existing skills/agents to learn the repo's local conventions (frontmatter keys it
   uses, naming, body structure, whether commands are thin delegators). Match them.

ls -d .agents .claude 2>/dev/null
ls .agents/skills .agents/agents .agents/commands .agents/docs 2>/dev/null
ls AGENTS.md CLAUDE.md 2>/dev/null

If the repo has no .agents/ layout yet, route the developer to the onboard sub-skill first
(onboard/SKILL.md) — onboarding scaffolds the directories this sub-skill writes into.

Trust boundary (write scope)

allowed-tools includes write-capable Edit, Write, and Bash.

Writes: new or updated files only under the repo's .agents/ kit —
skills/*/SKILL.md, agents/*.md, commands/*.md, docs/ catalogs — plus the
catalog index entries that keep them discoverable. Broad additions (a new skill
family, restructuring the kit) are proposed to the developer before creation.

It MUST NOT: edit files outside .agents/ (a skill's content may
document anything; this sub-skill writes only kit files), weaken the frontmatter
conventions (name, quoted version:, documentation_url, user-invocable,
allowed-tools, kebab-case), delete an existing skill/agent/command without
explicit approval, or commit/push.

---

Flows

Pick the flow that matches the developer's intent.

A. Create a skill

1. Audit fit — confirm a real, repeatable workflow exists (see "Repo-fit rubric" below). Skip
   generic skills that do not match an actual workflow.
2. Name it — kebab-case, English, unique. Check for collisions in the skills dir.
3. Scaffold — copy templates/SKILL_TEMPLATE.md into <skills-dir>/<name>/SKILL.md. Adapt the
   frontmatter to the repo's local convention (keys it already uses).
4. Fill — one-procedure focus; clear Goal, When-to-use, Steps, Validation.
5. Keep commands thin — if it needs a slash command, add a delegator (see Flow C).
6. Catalog — update the repo's catalog under the docs dir (see "Keep the catalog in sync").
7. Validate — naming, structure, frontmatter, and that any command is thin.

B. Create an agent

1. Confirm a recurring role with distinct model/tools needs (not just a one-off procedure → that
   is a skill).
2. Name it — kebab-case, English, unique.
3. Scaffold — copy templates/AGENT_TEMPLATE.md into <agents-dir>/<name>.md.
4. Choose a model tier — reason about it (see "Model tiers" below). Do NOT hardcode vendor model
   IDs inside the agent body; keep the abstract tier and map it in repo config.
5. Fill — Role, Inputs, Process, Output, escalation rules.
6. Catalog — update the catalog under the docs dir.
7. Validate — naming, structure, tier chosen with justification.

C. Create a command (thin delegator)

1. Confirm the target skill or agent exists.
2. Scaffold from templates/COMMAND_TEMPLATE.md into <commands-dir>/<cmd>.md — a ~20-line
   delegator: read the target skill/agent fresh and follow it, passing along args. Do NOT embed
   logic — logic lives in the skill/agent so updates propagate.
3. Reference the new command in the catalog / commands reference.

D. Update an existing skill / agent / command

1. Read the existing file and the repo's conventions first.
2. Make the smallest change that satisfies the request; preserve frontmatter the repo relies on.
3. Keep delegators thin; keep skills single-procedure.
4. Update the catalog if name, description, or surface changed.

E. Evaluate the catalog (skills reconciliation)

This is the flow the Final Review invokes for its skills-reconciliation pass. A plan
created under the pre-2.3.0 lifecycle invokes the same flow from its separate
"Skills & Agents Discovery" task; the procedure is identical either way.

1. Enumerate every skill (<skills-dir>/*/SKILL.md) and agent (<agents-dir>/*.md).
2. For each, capture name, one-line description, and model tier (if any).
3. Cross-check against the repo's catalog under the docs dir: flag missing entries, stale
   descriptions, orphaned commands, and duplicates.
4. Reconcile: update the catalog so it matches reality. Report drift found and fixed.

---

Repo-fit rubric (before creating)

Gather: stack & tooling (languages, frameworks, package manager, test runner, linter); workflows (how
code is built, tested, reviewed, released); pain points (repetitive manual steps); the existing kit
(avoid duplicates). Then:

• Create a skill for a repeatable procedure people do by hand.
• Create an agent for a recurring role with distinct model/tools needs.
• Create a command only as a thin entry point.
• Skip anything generic that does not match a real workflow.

Anti-patterns to avoid: generic kits that do not match the repo; fat commands with embedded logic;
duplicates of existing skills/agents; agents pinned to vendor model IDs in their bodies.

---

Model tiers (for agents)

Skills inherit the session model; agents may pin an abstract tier.

Tier	Use for	Examples
light	Mechanical, well-specified, low-judgment tasks	formatting, lint fixes, renames, list/status commands
standard	Most engineering work; moderate reasoning	feature work, refactors, test writing, reviews
heavy	High-judgment, architectural, ambiguous tasks	architecture, security audits, complex planning

How to choose: default to standard; drop to light only if mechanical and well-specified; raise
to heavy only for deep reasoning or high blast radius. Keep tiers abstract — map them to concrete
model IDs in the repo's runtime/config, in one place, never inside skill bodies.

---

Keep the catalog in sync

After any create/update, reconcile the repo's catalog (commonly under the docs dir, e.g. a
skills/agents catalog and a commands reference). Add or update the row for the affected skill/agent/
command: name, one-line description, tier (agents), and the command that invokes it. The catalog must
always match what is on disk.

---

Templates

• templates/SKILL_TEMPLATE.md — skill scaffold.
• templates/AGENT_TEMPLATE.md — agent scaffold.
• templates/COMMAND_TEMPLATE.md — thin-delegator command scaffold.

Reference them by these relative paths. Adapt the frontmatter to the host repo's local convention
(keep only the per-harness opt-in keys neighboring files already use — version,
documentation_url, user-invocable, allowed-tools; see "Frontmatter fields").

---

Validation

• Names are kebab-case, English, unique.
• Frontmatter matches the repo's convention; descriptions are a single line starting with a verb.
• Skills are single-procedure; agents declare a justified tier (no hardcoded vendor model IDs).
• Commands are thin delegators with no embedded logic.
• The catalog under the docs dir reflects every change.

---

Notes

• Reason about the repo; never copy a generic kit.
• Propose before creating when scope is broad — confirm the short list with the developer first.
• All references inside the skill are relative; nothing outside the skill root is required at runtime.