Tarqd Diagnostics Skill Diagnostics Lang Skill Creator

Diagnostics Language Skill Creator

This skill is a guided workflow. It produces a new, packaged language-specific diagnostics skill that extends diagnostics-base. It must not be run silently or in one shot — it has explicit user checkpoints at each phase.

Before starting: verify that diagnostics-base is available in the user's skill library. If it is not, stop and tell the user:

"This workflow requires the diagnostics-base skill to be installed first. Please install it and then re-run this workflow."

Phase Tracker

Display this tracker at the start of each phase, updating status as you go. Use ✅ for complete, ▶ for current, ○ for pending.

○ Phase 1 — Detect language & application context
○ Phase 2 — Scan codebase for existing libraries
○ Phase 3 — Research: idiomatic libraries & patterns
○ Phase 4 — User decisions: library stack & application profile
○ Phase 5 — Generate skill
○ Phase 6 — Test the skill via subagents
○ Phase 7 — Review & package

Phase 1 — Detect language & application context

Goal: Know the language, know whether there's an existing codebase, and understand the application type before doing anything else.

Start by checking if the conversation or any uploaded files already tell you the language. If the user said "create a Rust diagnostics skill", the language is Rust — do not ask again.

If the language is not clear, ask:

"Which programming language should this skill target?"

Then determine whether there's an existing project to scan:

• Check bash_tool for common project files: package.json, Cargo.toml, go.mod, pyproject.toml, requirements.txt, pom.xml, build.gradle, *.csproj.
• If found, note the project root and continue to Phase 2.
• If no project is detected, tell the user and continue — Phase 2 will be shorter (no scan), but the web search in Phase 3 is still required.

Also note the application type if detectable from file structure:

• CLI (main.rs, cmd/, __main__.py, bin/ entry points)
• HTTP service (Express/Axum/Gin/Flask/Spring controller patterns)
• Background worker / queue consumer
• Library / crate / package (no main entry point)
• Monorepo (multiple of the above)

Checkpoint: Print the phase tracker (Phase 1 ✅) and a one-paragraph summary of what you found before proceeding.

Phase 2 — Scan codebase for existing libraries

Goal: Find every logging, error handling, and diagnostics dependency already in use. Detected libraries are the highest-priority signal — the generated skill must be compatible with what's already there.

Run the appropriate scan based on language:

TypeScript / JavaScript (package.json)

# Extract dependencies + devDependencies keys
cat package.json | python3 -c "
import json,sys
d=json.load(sys.stdin)
libs=list({**d.get('dependencies',{}), **d.get('devDependencies',{})}.keys())
keywords=['pino','winston','bunyan','log4','consola','debug','morgan',
          'neverthrow','ts-results','effect','zod','yup','valibot','fp-ts']
print([l for l in libs if any(k in l for k in keywords)])
"

Rust (Cargo.toml)

# Find logging/error crates
grep -E '^\s*(thiserror|anyhow|eyre|miette|snafu|tracing|log|env_logger|
  tracing-subscriber|color-eyre|serde|tokio)' Cargo.toml Cargo.lock 2>/dev/null | head -30

Go (go.mod)

grep -E '(zap|zerolog|logrus|slog|sentry|pkg/errors|multierr|go-multierror)' go.mod

Python (pyproject.toml / requirements.txt)

grep -Ei '(structlog|loguru|logbook|rich|pydantic|marshmallow|cerberus|
  result|returns|stamina)' pyproject.toml requirements*.txt setup.cfg 2>/dev/null

If the project structure is non-standard or multi-language, adapt — the goal is to find what's already wired in, not to run a rigid script.

Also look for existing logging configuration or setup files:

• logger.ts, logging.py, log/config.go, src/telemetry.rs, etc.
• Note any existing log format (JSON vs pretty, any custom fields).

Checkpoint: Print the phase tracker (Phase 2 ✅) and list every detected library with a one-line description. If nothing was found, say so explicitly.

Phase 3 — Research: idiomatic libraries & patterns

Goal: Find the current best-practice library landscape for this language with real star counts and recency signals. Even if libraries were detected in Phase 2, run this search — the project may be using a suboptimal choice, and the user should know the alternatives.

Run two web searches in parallel:

Search 1: {language} logging library best 2025 structured JSON Search 2: {language} error handling library idiomatic {year} github stars

For languages where the stdlib has evolved (Go slog, Python structlog, JS native Error.cause), also search: {language} stdlib logging error handling 2025 best practices

From results, build a comparison table for each category:

Logging libraries:

Error / Result libraries:

Validation libraries (if applicable):

For each entry, note:

• Whether it's the ecosystem default / blessed choice
• Whether it integrates with OpenTelemetry
• Whether it's appropriate for libraries vs. applications (or both)
• Any known maintenance concerns (abandoned, security issues)

Also find one concrete idiomatic example per major library: a short snippet showing an error type definition and a log call. These become the seed examples in the generated skill.

Checkpoint: Print the phase tracker (Phase 3 ✅), the comparison tables, and the seed snippets. Tell the user you'll use these in Phase 4.

Phase 4 — User decisions: library stack & application profile

Goal: Nail down the exact choices that drive the generated skill. Ask all questions in one batch — do not drip-feed them one at a time.

Present a summary of what you know so far:

• Detected language
• Application type
• Libraries already in use (from Phase 2)
• Top recommended libraries (from Phase 3)

Then ask the following questions. Use ask_user_input_v0 for any that have clean option sets; use free-text prompts for open-ended ones.

Required questions (always ask):

1. Logging library — which one should the generated skill teach? Options: the detected library if any, top picks from Phase 3, or "none / stdlib only". If the project already uses one, default to that but confirm.

2. Error / Result library — typed errors (Result/Either pattern) or exception-based? Which library? Or stdlib only?

3. Validation library — is there a schema validation library in use? (e.g., zod, pydantic, serde) The skill should cover how to surface validation errors well.

4. Application type (if not already clear):

   • Library/crate/package (will be published; callers will match on errors)
   • Application/binary/service (owns the error boundary; can use opaque errors)
   • Both (monorepo with shared libs and services)

5. Log output format:

   • JSON in all environments
   • JSON in production, pretty in development (recommended)
   • Pretty only (dev tools / CLIs)

6. Error rendering target — who sees the errors?

   • End users (CLI, human-readable output required)
   • HTTP API consumers (JSON response body)
   • Internal / ops only (logs and dashboards)
   • Mixed (need all three views)

Optional questions (ask only if not determinable from context):

7. Async runtime (Rust: tokio/async-std; JS: native promises/async-await; Go: goroutines). Affects tracing/span integration.

8. Are there any existing internal error conventions (error code scheme, specific field names, a shared error base type) the skill should know about?

Checkpoint: Print the phase tracker (Phase 4 ✅) and a "Decision summary" table with every chosen library and setting before generating anything. Confirm with the user: "Does this look right? I'll generate the skill based on these choices." Wait for explicit confirmation before Phase 5.

Phase 5 — Generate the skill

Goal: Produce the language-specific SKILL.md by filling in the template from references/skill-template.md with all discovered and decided information.

Read references/skill-template.md now. Fill every section using:

• Phase 2 findings (existing libraries, log format, any current config)
• Phase 3 research (star counts, concrete code examples)
• Phase 4 decisions (chosen library stack, application type, render targets)

Non-negotiable requirements for the generated skill:

1. diagnostics-base must be a hard dependency. SKILL.md frontmatter only supports name and description (a compatibility: key is silently ignored by the loader). Encode the dependency two ways:

   • In the description, end with the sentence: "REQUIRES diagnostics-base to be installed — this skill does not repeat base rules, it extends them."
   • As the very first line of the body, an enforced banner:

     "REQUIRED: Read diagnostics-base fully before applying any pattern in this skill. This skill is an extension — it does not repeat the base rules. All base rules apply."

2. Every pattern must reference the base rule it extends. When showing how to write a log call, note which base rule (e.g., "structured fields — base §Structured logging") it implements.

3. Library decision tree must be present. Show when to use each chosen library in a decision table or flowchart — the most common confusion point for every language is "which of my libraries do I reach for here."

4. Code examples must be real, runnable snippets — not pseudocode. The base skill uses pseudocode; this skill uses actual syntax for the chosen language and libraries.

5. The checklist addendum must include language/library-specific items on top of the base checklist.

Name the output file: diagnostics-{language-lowercase}.skill (e.g., diagnostics-typescript.skill, diagnostics-rust.skill).

Save the SKILL.md to: /tmp/diagnostics-{lang}/SKILL.md

Checkpoint: Print the phase tracker (Phase 5 ✅) and show the generated SKILL.md inline. Tell the user: "Here's the generated skill. Review it and tell me anything to change before I package it."

Phase 6 — Test the skill via subagents

Goal: Verify the generated skill actually changes a fresh agent's output before packaging. A skill that reads well but doesn't land in practice is a waste of context window. Subagent tests catch this gap.

Spawn three subagents using the Agent tool (subagent_type: claude or general-purpose), in parallel. Each receives the path to the saved SKILL.md plus the path to diagnostics-base/SKILL.md, and a task it must complete by applying both skills.

Test 1 — Reviewer test (catches anti-patterns). Hand the subagent a short, deliberately bad snippet in {LANGUAGE} containing at least four base-skill violations: a swallowed error, log-and-rethrow, an interpolated log message, an "ERROR" level on a 4xx user error, no cause preservation, no error code. Prompt:

"Read both skills, then review the snippet. List every violation you find, citing the skill rule by section name. Under 250 words."

Score: count cited violations. A well-written skill produces >= 4 distinct citations with section names. If the subagent misses obvious violations, the skill's anti-pattern table is not specific enough — fix and re-test.

Test 2 — Author test (writes from scratch). Give the subagent a tight spec: "Write a function loadConfig(path) in {LANGUAGE} using only the libraries declared in the skill's Library Stack table. Define its error type and one log call. Apply both skills." No bad seed code — just the spec.

Score: read the output. Does it have a stable code field on the error type? Does it preserve cause when wrapping I/O errors? Does the log call use structured fields? Does it use OTel-style field names? If any of these miss, the skill's positive examples are under-specified — add more concrete code.

Test 3 — Log-level test (taxonomy). Give the subagent five scenarios and ask for the log level + one-line justification for each:

1. Database connection refused at startup (process cannot continue)
2. Card declined returning 402 to the client
3. Retry succeeded on attempt 3 of 5
4. Background job completed processing 12,000 items
5. Deprecated config field still in use

Expected answers (per base rules):

1. FATAL (then exit) — process cannot continue
2. INFO or DEBUG — normal traffic, not service breakage
3. WARN — leading indicator worth knowing about
4. INFO — notable state change worth keeping in production
5. WARN — actionable in aggregate

Score: 5/5 must match. If the subagent picks ERROR for #2, the skill's log level rules need stronger language about user errors vs. service errors.

Checkpoint: Print the phase tracker (Phase 6 ✅) and a results table:

If any test fails, iterate on the skill and re-run that specific test (no need to redo passing tests). Tell the user the results before Phase 7:

"Subagent tests: [X passed, Y failed]. [If failures] I've patched the following sections: [...]. Re-running failed tests."

Do not proceed to Phase 7 until all three tests pass.

Phase 7 — Review & package

Goal: Incorporate any user feedback and produce the final packaged .skill file.

Apply any changes the user requests. Then run a self-review against this checklist before packaging:

• diagnostics-base named in the description AND the opening body banner (do not add a compatibility: frontmatter key — it is not recognized)
• Opening REQUIRED: Read diagnostics-base first instruction
• Library decision tree present
• Every code example is real syntax (not pseudocode)
• Log level rules consistent with base skill
• Cause-chaining pattern shown for chosen libraries
• Structured logging example uses OTel field names
• Single-log-at-boundary rule shown in context of chosen framework
• Security / redaction pattern shown
• Checklist addendum present
• No base-skill rules contradicted or weakened
• All three Phase 6 subagent tests passed

Once all items pass, package using the skill-creator's package script:

# Copy scripts to a writable location if needed
cp -r /mnt/skills/examples/skill-creator/scripts /tmp/skill-creator-scripts
cd /tmp && python3 -m skill-creator-scripts.package_skill /tmp/diagnostics-{lang}
# or, if the above module path fails:
python3 /mnt/skills/examples/skill-creator/scripts/package_skill.py /tmp/diagnostics-{lang}

Present the .skill file to the user and confirm install instructions:

"Install this by dropping diagnostics-{lang}.skill into your Claude skills directory alongside diagnostics-base.skill. Both skills will be loaded together whenever you work on error handling or logging in {Language}."

Print the final phase tracker (all ✅) and close the workflow.

Error recovery

If the user bails out mid-workflow or says "skip" for a phase:

• Skip only non-critical phases (Phase 2 if no codebase exists is fine to skip)
• Phase 4 decisions are required — the skill cannot be generated without them
• If the user skips Phase 4, ask for the minimum: logging library + application type

If web search returns thin results (Phase 3):

• Fall back to training knowledge for the language
• Note clearly in the generated skill which library recommendations are from search (with date) vs. from training knowledge (may be stale)

If diagnostics-base is not installed:

• Stop immediately at the opening check
• Do not attempt to inline the base skill content
• The dependency is intentional — the skills are designed to be used together

Source: tarqd/diagnostics-skill — distributed by TomeVault.