claude-dev-env 1.38.0 → 1.39.0

package/skills/skill-writer/SKILL.md

@@ -1,222 +0,0 @@
----
-name: skill-writer
-description: >-
-  Write, create, or improve Claude Code skills (SKILL.md files) with correct
-  frontmatter, progressive disclosure, and prompt-engineering principles.
-  Covers workflow skills, enforcement rules, advisory guidance, reference
-  lookups, and automation scripts.
----
-@packages/claude-dev-env/skills/skill-writer/REFERENCE.md
-
-# Skill writer
-
-**Core principle:** A skill is a prompt -- the SKILL.md body is injected into Claude's context when triggered. The same principles that make prompts effective make skills effective.
-
-**Canonical source:** https://platform.claude.com/docs/en/claude-code/skills -- the official reference for skill structure, frontmatter fields, progressive disclosure, and string substitutions.
-
-## Skill-only output rule (overrides all other delivery instructions)
-
-This skill produces skill artifacts. It does not perform the underlying task the skill would automate.
-
-When this skill is active, your response contains exactly one of:
-1. **Clarifying questions** to gather information needed to write a better skill (Step 3) -- then stop and wait.
-2. **The skill artifact** in fenced code blocks (SKILL.md and optionally REFERENCE.md) -- then stop.
-
-Prohibited responses: executing the task the skill would automate, proposing implementation changes unrelated to skill authoring, explaining what *you would do* to accomplish the underlying task.
-
-## When this skill applies
-
-Trigger for any request to **author**, **refine**, or **restructure** a Claude Code skill: SKILL.md files, skill frontmatter, skill body content, reference files, or skill directory structure.
-
-Use this skill when the user needs a structured skill artifact; for quick answers about skill syntax, answer directly in plain text.
-
-When invoked with arguments (e.g. `/skill-writer improve this: [paste]`), treat `$ARGUMENTS` as the skill content to refine.
-
-### Ground-up multi-file packages (required)
-
-When the user is creating a **new** skill as a **package** (workspace with `ARCHITECTURE.md`, `REFERENCE.md`, `EXAMPLES.md`, `WORKFLOWS.md`, `evals/*.json`, per-file human review), **before** drafting `SKILL.md`:
-
-1. Read `prompt-generator/templates/skill-from-ground-up.md` (installed layout: sibling folder under the same `skills/` parent; repository path: `skills/prompt-generator/templates/skill-from-ground-up.md` in [jl-cmd/prompt-generator](https://github.com/jl-cmd/prompt-generator)).
-2. Ensure `/prompt-generator` has run with that template filled so architecture-first steps, checkpoint gates, and eval evidence rules are already agreed.
-
-If the task is **only** editing an existing `SKILL.md` or a small single-file tweak, this subsection does not apply.
-
-### Refinement multi-file packages (required)
-
-When the user is **refining** an existing skill as a **package** (baseline skill directory, `ARCHITECTURE.md` with planned deltas, checkpointed updates to REFERENCE / EXAMPLES / WORKFLOWS / `evals/`), **before** rewriting multiple files:
-
-1. Read `prompt-generator/templates/skill-refinement-package.md` (installed layout: under the same `skills/` parent; repository path: `skills/prompt-generator/templates/skill-refinement-package.md` in [jl-cmd/prompt-generator](https://github.com/jl-cmd/prompt-generator)).
-2. Ensure `/prompt-generator` has run with that template filled so baseline root, workspace root, observation inputs, and evidence rules are fixed before edits proceed.
-
-If the change set is a **small single-file** tweak, this subsection does not apply.
-
-## Workflow (run in order)
-
-### 1. Classify the skill type
-
-Pick one primary:
-
-| Type | Purpose | Example |
-|------|---------|---------|
-| `workflow` | Multi-step process with sequential phases | TDD enforcement, plan review, PR submission |
-| `enforcement` | Rules that constrain behavior | Code standards, comment preservation, commit rules |
-| `advisory` | Guidance where multiple approaches are valid | Best practices, design patterns, optimization tips |
-| `reference` | Lookup material loaded on demand | API docs, field mappings, configuration tables |
-| `automation` | Script-driven with tool calls | PDF form filling, browser automation, file processing |
-
-### 2. Set degree of freedom
-
-Match specificity to task fragility:
-
-- **High:** Multiple valid approaches; use numbered goals and acceptance criteria. The skill states *what* to achieve, not *how*.
-- **Medium:** Preferred pattern exists; use structured steps with room for adaptation. The skill states a recommended approach but allows deviation when justified.
-- **Low:** Fragile or safety-critical; use exact steps, exact field names, and boundary constraints. The skill states precisely what to do and what not to do.
-
-### 3. Collect missing facts
-
-Ask 1-3 short questions if needed. Focus on:
-- What capability does the skill provide? (one skill = one capability)
-- When should the skill trigger? (specific phrases, file patterns, workflow position)
-- What tools does the skill need? (Bash, Read, Grep, MCP tools)
-- What does the skill produce? (output format, files created, state changes)
-- Where does it fit in existing workflows? (before/after other skills)
-
-### 4. Choose location
-
-**Personal skills** (`~/.claude/skills/` or package-managed):
-- Individual workflows, experimental skills, cross-project utilities
-
-**Project skills** (`.claude/skills/`):
-- Team conventions, project-specific expertise, committed to git
-
-### 5. Write frontmatter
-
-Apply official constraints from the canonical source:
-
-```yaml
----
-name: skill-name-here
-description: >-
-  [What it does] in third person. [When to use it].
-  Triggers: '[phrase 1]', '[phrase 2]', '[phrase 3]'.
----
-```
-
-**Name:** lowercase, hyphens, numbers only. Max 64 chars. No `anthropic` or `claude`. Must match directory name.
-
-**Description:** max 1024 chars. Third person. No XML tags. Include trigger phrases that match how users naturally ask for this capability.
-
-**Optional fields** -- include only when needed:
-- `allowed-tools` -- when the skill requires specific tools (e.g., `Bash(node *), Read`)
-- `argument-hint` -- when the skill accepts arguments (e.g., `[filename] [format]`)
-- `paths` -- when the skill applies only to certain file types (e.g., `"*.py"` or `["*.ts", "*.tsx"]`)
-- `context: fork` -- when the skill needs isolated execution without conversation history
-- `disable-model-invocation: true` -- when the skill should only trigger via explicit `/name` invocation
-- `effort` -- when the skill benefits from a specific thinking depth (`low`, `medium`, `high`, `max`)
-
-See REFERENCE.md for the complete field table with types, defaults, and constraints.
-
-### 6. Build the skill body
-
-Apply prompt-engineering principles -- a skill body is a prompt. Source: https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices
-
-**Structure with XML section tags** where content mixes instructions, context, and examples. Anthropic: "Use consistent, descriptive tag names across your prompts. Nest tags when content has a natural hierarchy." For skills under ~20 lines, use concise plain structure instead.
-
-**Set a role** when the skill benefits from focused behavior. Anthropic: "Setting a role in the system prompt focuses Claude's behavior and tone for your use case. Even a single sentence makes a difference."
-
-**Add motivation behind constraints.** Anthropic: "Providing context or motivation behind your instructions... can help Claude better understand your goals and deliver more targeted responses." State *why* a rule exists, not just *what* it forbids.
-
-**Frame positively.** State the desired behavior directly. Anthropic: "Your response should be composed of smoothly flowing prose paragraphs" provides clearer guidance than a prohibition-only instruction. Write "validate field names before filling" rather than "never guess field names."
-
-**Emotion-informed framing.** Anthropic's emotion concepts research (2026) found internal activation patterns that causally influence output quality. Apply to skill writing: (1) provide clear success criteria and escape routes; (2) use collaborative framing; (3) frame tasks as interesting problems; (4) invite transparency -- include "say so if you're unsure" when appropriate; (5) use constructive, forward-looking tone.
-
-**Golden rule check.** Anthropic: "Show your prompt to a colleague with minimal context on the task and ask them to follow it. If they'd be confused, Claude will be too." Apply the same test to the skill body.
-
-**Progressive disclosure.** Keep the SKILL.md body under 500 lines. Move heavy content (field tables, long examples, lookup data, scripts) into REFERENCE.md or separate files that load only when referenced.
-
-#### Body structure template
-
-```markdown
-# Skill Name
-
-**Core principle:** [One sentence capturing the essential insight]
-
-## When this skill applies
-[Trigger conditions, workflow position]
-
-## The Process
-
-### Step 1: [First Action]
-[Instructions -- narrative, not verbose checklists]
-
-### Step 2: [Second Action]
-[Continue with clear sequential steps]
-
-## Output Format
-[What the skill produces -- be specific]
-
-## Examples
-[3-5 concrete scenarios with inputs and expected outputs]
-```
-
-### 7. Control output format
-
-If the skill produces structured output, specify the format explicitly. Use XML format indicators when the skill should produce tagged sections. Match the formatting style in the skill body to the desired output style -- Anthropic notes the prompt's own formatting influences the response.
-
-### 8. Add examples
-
-Include 3-5 concrete examples showing the skill in action. Wrap in `<example>` tags with diverse, representative inputs. Anthropic: "Include 3-5 examples for best results. You can also ask Claude to evaluate your examples for relevance and diversity."
-
-Examples should cover:
-- A typical use case
-- An edge case or boundary condition
-- A case where the skill should decline or ask for clarification
-
-### 9. Self-check
-
-Before delivering, verify against this rubric:
-
-- [ ] Description is third person with trigger phrases
-- [ ] Under 500 lines
-- [ ] States what to do in positive terms (not prohibition-heavy)
-- [ ] Degree of freedom matches task fragility
-- [ ] Progressive disclosure used (heavy content in REFERENCE.md)
-- [ ] No time-sensitive claims unless clearly dated
-- [ ] Examples are concrete, not abstract
-- [ ] Frontmatter fields are valid per official docs
-- [ ] If tools needed: specifies `allowed-tools`
-- [ ] If arguments expected: includes `argument-hint`
-- [ ] Workflow steps are sequential and numbered
-- [ ] Golden rule: a colleague could follow this skill without extra context
-- [ ] Motivation provided for constraints (why, not just what)
-- [ ] One skill = one capability (not a bundle of loosely related features)
-- [ ] String substitutions used correctly (`$ARGUMENTS`, `${CLAUDE_SKILL_DIR}`, etc.)
-
-### 10. Deliver
-
-Final artifact as **fenced code blocks** the user can paste as-is: one block for SKILL.md, optionally one for REFERENCE.md. The fenced blocks are your entire response -- no surrounding commentary or explanation.
-
-## Claude 4.6 considerations
-
-When writing skills for current Claude models, apply these patterns:
-
-- **Overtriggering:** Dial back aggressive language. Anthropic: "Where you might have said 'CRITICAL: You MUST use this tool when...', you can use more normal prompting like 'Use this tool when...'." Skills that shout get ignored or cause erratic behavior -- write in a direct, conversational tone.
-- **Overeagerness:** Include scope constraints. Anthropic: "Claude Opus 4.5 and Claude Opus 4.6 have a tendency to overengineer by creating extra files, adding unnecessary abstractions, or building in flexibility that wasn't requested." Skills should state their boundary clearly.
-- **Overthinking:** Anthropic: "Replace blanket defaults with more targeted instructions. Instead of 'Default to using [tool],' add guidance like 'Use [tool] when it would enhance your understanding of the problem.'" Write conditional triggers, not blanket rules.
-- **Conservative vs proactive action:** For skills that should act, use explicit language. For skills that should advise, include: "Default to providing information and recommendations. Proceed with changes only when the user explicitly requests them."
-
-## Autonomy and safety pattern
-
-For `automation` and `enforcement` skill types, include reversibility guidance. Anthropic provides this pattern:
-
-```text
-Consider the reversibility and potential impact of your actions. You are encouraged to take local, reversible actions like editing files or running tests, but for actions that are hard to reverse, affect shared systems, or could be destructive, ask the user before proceeding.
-```
-
-## Conflict resolution
-
-When skill-writing guidance conflicts across sources, defer to the authority tier:
-
-1. **Tier 1 (primary):** Anthropic -- the model provider's own documentation is authoritative for Claude behavior
-2. **Tier 2 (strong secondary):** OpenAI, Google DeepMind, Microsoft Research -- major lab guidance often transfers across models
-3. **Tier 3 (supplementary):** Community resources, courses, individual blogs -- valuable for patterns and intuition, not authoritative on model specifics

package/skills/tdd-team/SKILL.md

@@ -1,128 +0,0 @@
----
-name: tdd-team
-description: "Spawn an orchestrated TDD agent team (planner, tester, implementer, validator) to build features or fix bugs via Red-Green-Refactor. Use when the user says /tdd-team, 'TDD team', 'agent team for this', 'have a team build this', or describes a non-trivial implementation task they want done with test-driven development and parallel agents. Also triggers on 'agentically', 'orchestrated team', or 'planner implementer tester validator'."
----
-
-# TDD Team
-
-Spawn an orchestrated team of 4 agents that implements a feature or fix using strict Test-Driven Development: Plan -> RED (failing tests) -> GREEN (minimum code to pass) -> Validate.
-
-## When to use
-
-- Non-trivial features requiring 3+ files changed
-- Bug fixes where the fix needs tests proving the bug exists first
-- Any task where the user says "TDD" and "team" or "agentically"
-
-## Arguments
-
-The full task description follows the command. Example:
-```
-/tdd-team fix proportional split distribution when item prices exceed order total
-```
-
-## Workflow
-
-### Phase 1: Setup
-
-1. **Create team** via TeamCreate with a name derived from the task (kebab-case, max 30 chars)
-2. **Create tasks** with dependencies:
-   - #1 PLAN — blocked by nothing
-   - #2-N RED — blocked by #1
-   - #N+1 GREEN — blocked by all RED tasks
-   - #LAST VALIDATE — blocked by GREEN task
-3. **Spawn 4 agents** (all at once for parallel startup):
-   - `planner` (team-lead) — reads code, designs the approach, defines file ownership so agents work independently, and messages teammates with clear assignments
-   - `tester` (team-implementer) — writes failing tests that define the expected behavior before implementation exists, establishing the contract the code must satisfy
-   - `implementer` (team-implementer) — writes the minimum production code to satisfy the test contracts, keeping scope tight to keep all behavior tested
-   - `validator` (team-reviewer) — runs the full test suite to catch regressions across files, and verifies the feature works end-to-end
-
-### Phase 2: Orchestrate
-
-As the team lead, your job is to:
-- **Nudge blocked agents** when upstream tasks complete (send messages)
-- **Shutdown idle agents** when all their tasks are done (send shutdown_request)
-- **Report progress** to the user at milestones (plan done, tests written, implementation done, validation complete)
-- **Fix coordination issues** (e.g., tester waiting for planner who already finished)
-
-### Phase 3: Cleanup
-
-After validator reports success:
-1. Shutdown all remaining agents
-2. Clean up any temporary files, scripts, or helper files created during iteration
-3. Delete the team via TeamDelete
-4. Run `python -m pytest tests/ -q` to confirm final count
-5. Commit with descriptive message
-6. Report summary to user
-
-## Agent Prompts
-
-Each agent receives a structured prompt containing:
-
-### Common context (all agents get this)
-- The task description from the user
-- Working directory path
-- Code rules reminder: no magic values, no abbreviations, complete type hints, no new inline comments, imports at top, constants in config/
-- Collaboration guidance: flag ambiguity or uncertainty to teammates and use [PLACEHOLDER] for values you have not verified. Your role contributes to a specific phase of the TDD cycle — the separation between phases is what makes TDD effective.
-- Anti-test-fixation: write general solutions that work for all valid inputs, not just the specific test cases. Tests verify correctness -- they do not define the solution. If a test seems incorrect, flag it rather than working around it.
-- Commit-and-execute: choose an approach and commit to it. Avoid revisiting decisions unless new information directly contradicts your reasoning. Course-correct later if needed.
-
-### Planner-specific
-- Read the relevant source files before designing
-- Design the implementation with function signatures and return types
-- Define file ownership boundaries (tester owns test files, implementer owns production files)
-- Create the task breakdown with specific test cases to write
-- Mark task complete and message teammates with assignments
-- If the scope is unclear or requirements are ambiguous, flag this to the orchestrator and propose assumptions for confirmation
-- Commit to an approach and communicate it clearly to teammates -- avoid revisiting architectural decisions unless implementation reveals a fundamental problem
-
-### Tester-specific
-- Wait for planner to complete (check TaskList)
-- Write FAILING tests that reproduce the desired behavior
-- Use existing test patterns from the project (read test files first)
-- Mock external calls, use patch.dict for env vars
-- Tests must FAIL before implementation — that's the point
-- Mark each task complete, check TaskList for next work
-- If a test case from the plan is ambiguous about expected behavior, ask the planner for clarification — a clear test is the most valuable deliverable
-
-### Implementer-specific
-- Wait for tester to complete (check TaskList)
-- Read the failing tests to understand expected behavior
-- Write MINIMUM code to make tests pass — no extras
-- Run the specific test file to verify green
-- Mark task complete, check TaskList for next work
-- Write general solutions that work for all valid inputs -- never hard-code values or create implementations that only satisfy specific test assertions
-- If a failing test seems to test the wrong behavior or has a bug in the test itself, flag it to the validator for review
-
-### Validator-specific
-- Wait for implementer to complete (check TaskList)
-- Run full test suite (`python -m pytest tests/ -q`)
-- Fix test regressions from the new code (e.g., unmocked new functions in existing tests)
-- Verify new functions exist and are wired correctly
-- Verify the implementation fully captures the planner's design (all planned functions exist, signatures match, edge cases covered)
-- Verify the implementation follows project code rules (centralized config, complete type hints, full-word naming, self-documenting code)
-- Report: test count, pass/fail, regressions found and fixed, plan coverage assessment, code rules compliance
-- If regressions trace to a design issue, report the root cause to the orchestrator for a targeted follow-up task
-
-## Task Sizing
-
-The planner decides how many RED tasks to create based on the scope:
-- **Small fix** (1 function): 1 RED task, 1 GREEN task
-- **Medium feature** (2-3 functions): 2-3 RED tasks (can be parallel), 1-2 GREEN tasks
-- **Large feature** (4+ functions across files): 3-5 RED tasks, 2-3 GREEN tasks with dependencies
-
-## Error Recovery
-
-- If tester writes tests that import non-existent constants, that's expected (RED phase)
-- If implementer's code breaks existing tests, validator fixes the regressions
-- If validator finds issues, create a new fix task and assign to implementer
-- If an agent goes idle repeatedly without progress, send a message with specific instructions
-- When encountering obstacles, use standard tools and approaches -- do not create helper scripts or workarounds to bypass problems; flag the issue for resolution
-
-## Ownership Boundaries
-
-Each agent owns its phase — this separation ensures the TDD cycle produces reliable results:
-
-- The tester writes tests; the implementer writes production code. This separation ensures tests define behavior independently of implementation.
-- Failing tests exist before any production code. This is the core of TDD — the test proves the requirement, then the code satisfies it.
-- The validator runs after implementation. It catches regressions across files that are invisible to agents working on individual files.
-- Shut down agents after their tasks complete. Idle agents consume context; active agents produce value.