Install
Works with Claude Code, Claude apps, Codex, Cursor, Gemini CLI, GitHub Copilot, VS Code, OpenCode, Goose, Windsurf. Read the repository before you give any tool access to your data.
npx agentmarket install skill/marcuskrogh-define
Asks which of your tools to add it to (Claude Code, Codex, Cursor, VS Code, Windsurf, Gemini CLI, Claude Desktop) and shows the change first. See https://agentmart.dev/cli.
Get the weekly diff
The weekly diff: one email a week with new agent tools, notable updates and grade changes, plus security alerts for tools you follow. Double opt-in, one-click unsubscribe.
About
Use cases: productivity
Reviews
Write a reviewNo reviews yet. Installed it? Reviews come from verified installers only. How reviews work
Permissions
What this skill can do and what it asks for, from its SKILL.md. Checked Oct 25, 2026.
- Tools
- No tools found
- Secrets
- None asked for
- Install scripts
- None found
Network hosts it talks to 1
codeload.github.com· found in install command
Files it touches
~/.claude/skills/define
What's inside
- Skill name
define- Description
- Definition front door for concrete work. Interviews the user (CONCEPT_ALIGNMENT) before classification and workflow binding. A short description starts the interview; it does not approve the plan. Produces PLAN.md (with Classification + Workflow), Sub-tasks, and the Task's delivery branch/PR. Prefer /explore when the destination is foggy; prefer explicit /bug /tweak /refine /rework only as manual overrides. Prefer /adopt when the whole existing tree should meet the structure catalog.
- License
- MIT
- Instructions
- 126 lines in SKILL.md · view on GitHub
Files (2)
- SKILL.md
- overrides.md
SKILL.md (first lines)
# Define Applies [CONCEPT_ALIGNMENT](../concepts/CONCEPT_ALIGNMENT.md), [CONCEPT_DEFINITION](../concepts/CONCEPT_DEFINITION.md), and [CONCEPT_CLASSIFICATION](../concepts/CONCEPT_CLASSIFICATION.md) to a **specific topic**. Produces `PLAN.md` (including **Classification** + **Workflow** binding) and Sub-tasks on the **pipeline Task**. Applies [CONCEPT_SKILL](../concepts/CONCEPT_SKILL.md). **On invoke:** read those concepts, [CONCEPT_SKILL](../concepts/CONCEPT_SKILL.md), [../concepts/CLASSIFICATION-CATALOG.md](../concepts/CLASSIFICATION-CATALOG.md), and [../workflow/SKILL.md](../workflow/SKILL.md). ## Extensions | Slot | This skill | |------|------------| | **Subject** | Concrete work: defect, small delta, brownfield whole-tree structure, bounded structure-only, measured impl swap, or feature slice (explore route Task when present) | | **Probes** | Thin description; scope in/out; behaviour divergences (or preserve-behaviour / parity bar as class requires); constraints; **pass criteria** (checkable success rows, distinct from the specification); work packages; fog pointers on this route Task; Task key; how to apply RESEARCH/MODEL/SANDBOX — same definition probe set as classic define, proportional depth | | **Stop condition** | Scope, behaviour (or parity/preserve-behaviour), constraints, and pass criteria are stated by the user or explicitly deferred; the user approved the readiness prompt; then class + workflow binding are persisted | | **Alignment / definition artifact** | `PLAN.md` (path from WORKSPACE) — always; class lives in Classification, not a separate BUG/TWEAK file | | **Readiness prompt** | "Does this plan and workflow binding look right?" | | **Opening** | Thin description **required**. Missing → "What should we define?" Any unstated probe → one question on that probe. Every in-play probe already stated in the user's words → artifact + readiness prompt, then stop | | **Scope guard** | Stay on this Task; no foggy destination mapping (`/explore`); write approved `PLAN.md` on the delivery branch; **do not skip alignment** to rush classify/bind | | **Depth** | Proportional per CONCEPT_DEFINITION. A provisional class selects full vs lightweight only. Lightweight still runs CONCEPT_ALIGNMENT on unsettled probes. Full when the provisional class is feature, the class is ambiguous, or a definition probe is unstated | | **Work packages** | Sub-tasks per package when more than one unit; single package OK for small classes | | **Class catalog** | [CLASSIFICATION-CATALOG.md](../concepts/CLASSIFICATION-CATALOG.md) | | **Template catalog** | same | | **Binding rules** | same | | **Artifact sections** | `## Classification` + `## Workflow` required (in addition to full plan body) | | **Tracker mirror** | Copy class, template, params, and chain onto the Task (and Story when linked). The workflow appends **Next** | ## Steps 1. **Resolve entry** — Require a thin description or ask once; fetch route Task + Story when present; load ROADMAP / RESEARCH / MODEL / SANDBOX as supportive. Form a provisional class only to choose full vs lightweight. Done when the subject and which probes are still unstated are known. 2. **Align and define** — Follow CONCEPT_ALIGNMENT + CONCEPT_DEFINITION until every in-play probe is user-settled. Done when scope, behaviour/parity/preserve-behaviour, constraints, and pass criteria are stated by the user or explicitly deferred, and the turn has waited for those answers. 3. **Classify, bind, and confirm** — Apply CONCEPT_CLASSIFICATION + the catalog on that agreed description: final **class**, **template** + **parameters** (efficiency-first), confirm only on costly ambiguity. Present `PLAN.md` (Classification + Workflow included) and the readiness prompt. Done when the user approves the prompt — without reopening settled definition decisions unless the binding exposes a new divergence. 4. **Persist and track** — Write `PLAN.md` only after that approval. Follow delivery continuity, apply the define tracker row, and mirror binding fields on the tracker. Record outcome `ready`. Apply the workflow transition. Done when artifact, Sub-tasks, branch/PR, comments, and mirrors agree. ## Artifact ```markdown # Implementation plan: [title] ## Summary - … ## Scope / Decisions / Constraints - … (user-aligned in this define session — deep alignment when Depth is full) ## Classification - Class: bug | tweak | adopt | refine | rework | feature | iterate - Confidence: high | medium - Why: … ## Workflow - Template: fix-fast | delta-fast | structure-safe | parity-iterative | feature-standard | feature-heavy - Parameters: - implement.mode: single | multiagent - implement.verify: tests | non-regression | comparative - implement.iteration: one-shot | until-bar - test.mode: skip | dedicated - harden.mode: skip | dedicated - review.mode: single | multiagent - review.depth: focused | full - review.lasers: bundled | sequential - side_paths: none | research | model | research+model - sandbox: none | inject - Chain: architect → implement → test → restructure → review → ship # adopt: characterize → architect → implement → test → restructure → review → ship - Rationale: … ## Inputs - Research: RESEARCH.md (if any) - Model: MODEL.md (if any) - Sandbox: SANDBOX.md (if any) ## Pass criteria - … # one observable per row; fail if unmet. Distinct from Scope. Docs-only: none — no executable behaviour. Parity bar metrics when Class is rework. ## Work packages 1. … ## Open items - … ## Tracker - Provider: … - Story: <KEY> (if linked) - Task: <KEY> - Sub-tasks: … - Branch: <delivery-branch> - PR: <url or draft url> - Classification: <class> - Workflow: <template> ``` The recorded chain in that template stays the automatic shape. When WORKSPACE **Shipping procedure** is `human review`, assembly inserts `human review → fix-forward` after `review` and before `ship`. ## Tracker (after approval) Follow [delivery continuity](../workflow/delivery.md) and the [define tracker row](../workflow/tracker-sync.md#matrix). Enrich the explore route Task when present; otherwise create the pipeline Task. Keep it **To Do**, create Sub-tasks per work package, and record `PLAN.md`, **Classification**, **Workflow** (template + params + chain), branch/PR, and Sub-task keys on the Task, parent Story, and enabled mirror. The workflow appends **Next**. ## Inputs | Input | When present | When absent | |-------|----------------|-------------| | Thin description | Apply it as the subject | Ask "What should we define?" once | | `ROADMAP.md` | Apply destination and fog as context | Continue without a map | | `RESEARCH.md` | Apply as evidence, not as decisions | Continue without literature | | `MODEL.md` | Apply as the formulation | Continue without a math spec | | `SANDBOX.md` | Apply the element, bar, and promote map as context | Continue without a sandbox |
From marcuskrogh/skills, MIT.
More from marcuskrogh/skills
- 0.587score
explore
marcuskrogh
Exploration through foggy or oversized work.
SkillAFreshness: A★ 1Claude Code +9 - 0.587score
help
marcuskrogh
Help choosing skills and workflows.
SkillAFreshness: A★ 1Claude Code +9 - 0.587score
iterate
marcuskrogh
Iteration on already merged work.
SkillAFreshness: A★ 1Claude Code +9 - 0.587score
adopt
marcuskrogh
Adopt the structure catalog across an existing codebase that was not built to it.
SkillAFreshness: A★ 1Claude Code +9 - 0.584score
implement
marcuskrogh
Implementation through managed, value-routed work packages.
SkillAFreshness: A★ 1Claude Code +9 - 0.584score
refine
marcuskrogh
Refine alignment and lightweight definition for a limited area that needs refactoring or descriptive improvement without changing behaviour.
SkillAFreshness: A★ 1Claude Code +9 - 0.584score
rework
marcuskrogh
Rework alignment and lightweight definition for an intentional implementation change that must not degrade measured outcomes.
SkillAFreshness: A★ 1Claude Code +9 - 0.584score
model
marcuskrogh
Mathematical alignment through LaTeX-only questions.
SkillAFreshness: A★ 1Claude Code +9 - 0.584score
restructure
marcuskrogh
Restructure: refactoring pass against the structure catalog (campground, names, size, cohesion, seams, named smells, CRAP).
SkillAFreshness: A★ 1Claude Code +9 - 0.584score
review
marcuskrogh
Review of a change as sequential or bundled lasers across Spec, Correctness, Integration, Architecture, and Standards: find, fix, then publish a code review.
SkillAFreshness: A★ 1Claude Code +9 - 0.584score
architect
marcuskrogh
Architecture for one Task: record where the work sits (modules, layers, seams, dependency direction) in ARCHITECTURE.md.
SkillAFreshness: A★ 1Claude Code +9 - 0.584score
bug
marcuskrogh
Bug alignment and lightweight definition for a clear defect.
SkillAFreshness: A★ 1Claude Code +9
Signals
- Stars
- 1
- Forks
- 1
- Open issues
- 0
- Last commit (this folder)
- Oct 3, 2026 (26 days ago)
From GitHub (fetched Oct 24, 2026).
Details
- License
- MIT
- Repository
- github.com/marcuskrogh/skills
- First seen
- Oct 24, 2026
- Last checked
- Oct 24, 2026
- Source
- GitHubgithub-graphql
More from marcuskrogh
- 0.587score
iterate
marcuskrogh
Iteration on already merged work.
SkillAFreshness: A★ 1Claude Code +9 - 0.587score
help
marcuskrogh
Help choosing skills and workflows.
SkillAFreshness: A★ 1Claude Code +9 - 0.587score
adopt
marcuskrogh
Adopt the structure catalog across an existing codebase that was not built to it.
SkillAFreshness: A★ 1Claude Code +9 - 0.587score
explore
marcuskrogh
Exploration through foggy or oversized work.
SkillAFreshness: A★ 1Claude Code +9 - 0.584score
rework
marcuskrogh
Rework alignment and lightweight definition for an intentional implementation change that must not degrade measured outcomes.
SkillAFreshness: A★ 1Claude Code +9 - 0.584score
refine
marcuskrogh
Refine alignment and lightweight definition for a limited area that needs refactoring or descriptive improvement without changing behaviour.
SkillAFreshness: A★ 1Claude Code +9
Cite this page
define skill. Agent Mart. https://agentmart.dev/skills/marcuskrogh-define (last updated 2026-10-24)