Skilled Improve

use-skilled-cli

Help the user evaluate and evolve Agent Skills through the skilled CLI.

Manual: bundled references/cli.md and references/eval-yaml.md.

Language Policy

Default to English when responding to the user. If the user writes in Chinese (or any other language), switch to that language and stay consistent with the user's input throughout the session.

Detection rules (highest priority first):

1. The user explicitly specifies a language in the current message (e.g. "answer in English" or "reply in Chinese") → follow the user's instruction.
2. The natural language used in the user's current message → match it.
3. None of the above → use English (default).

Regardless of the response language, technical identifiers in this SKILL — CLI commands, eval.yaml / case.yaml field names, report field names, etc. — MUST stay in their original English form. Do not translate them.

Language Rules for Generated Artifacts

Use the user's current language for replies, newly authored explanatory prose, and YAML comments. Preserve the language and semantics of test inputs, expected output literals, identifiers, commands, paths, and existing regression cases. Do not translate a deterministic assertion merely because the conversation language changed. When testing localized behavior, keep the original business text even in an English conversation.

Treat assets/*.tmpl as structural references: adapt placeholder prose and comments to the requested language. Review only files created or changed for this task; never rewrite unrelated cases to enforce a language preference.

Use short field-leading comments for generated YAML. Explain actual configured fields: schema_version, environment.type, engine.name, cases.files, and report.formats in eval.yaml; id, title, input.prompt, and judge.type in case YAML. For nested fields, place the comment inside the parent mapping immediately before the child field, rather than only before the parent. Comment expect and a top-level judge when present; do not add an optional field just to satisfy a comment check. Keep field names and enum values unchanged. skilled import does not preserve template comments.

What is skilled

skilled is an evaluation CLI for Agent Skill authors. It installs the Skill into a real Agent Engine (Claude Code, Codex, qodercli, etc.), spins up an execution environment for each case, runs the prompt, then grades the result via declared rules / LLM judges / custom scripts, and finally produces a report.

Typical layout:

Copy code
my-skill/
  SKILL.md
  evals/
    eval.yaml
    cases/
      <case-id>.yaml
    fixtures/

When to trigger

Use this skill in any of the following situations:

• The user asks to "run / evaluate / verify / test this skill".
• The user asks to "fix / improve / iterate / evolve this skill" from eval failures.
• The user wants to "add evals, test cases, or regression cases to a skill".
• The user asks to record explicitly attributed Skill usage, evidence, or feedback.
• The user wants to review captured Skill observations or turn approved feedback into a regression case.
• The user wants to edit eval.yaml / case.yaml, or asks you to choose an appropriate judge type.
• The user mentions skilled run/validate/list-cases/report/import/init.
• The user wants to migrate from Anthropic evals.json to skilled.
• The current working directory contains evals/eval.yaml or evals/evals.json and the user wants to run it.

Choose the requested workflow

• Diagnose or plan: read the target Skill and supplied evidence, identify concrete problems, and propose changes or cases. Do not install tools, create eval files, change configuration, or run evaluations unless requested.
• Create or edit cases: use Steps 1–4 to create or update the suite and validate its configuration. Stop before credentials and execution unless the user also requests a run. Report the actual config/case paths and validation status in the final reply.
• Evaluate: follow Steps 0–7 for the requested suite; report failures without automatically editing the Skill.
• Improve: follow the evaluation workflow and Step 8 within the requested scope. Preserve valid existing cases and assertions.

Evaluation workflow

Observation capture mode (optional)

Use this mode when the user asks to record the current Skill interaction or feedback and observation tools are available. It requires either a current Codex release with the codex-skilled plugin and lifecycle-hook support, or DSH with the skilled plugin's observer explicitly enabled. Do not retrofit observation capture onto the evaluation adapter without verifying host plugin support.

If the user asks how to install either host plugin, or the required observation tools are missing and the user asks to enable them, read the "Codex plugin" or "DSH plugin" section of references/install.md. Do not install or enable a plugin unless the user asks.

Claude Code, qodercli, Qwen Code, and other Agent Engines are not supported for this observation workflow. This restriction does not apply to normal skilled evaluation runs, which remain multi-engine.

1. Attribute the interaction only when the user explicitly references a Skill (for example, $my-skill) or the responsible Skill is otherwise known with certainty.
2. In Codex, if explicit attribution is absent but certain, call mark_skill_invocation once with the Skill name and optional version. DSH derives exact attribution from its durable /skill injection or successful skill tool result; do not add inferred attribution.
3. In Codex, use attach_skill_evidence only for references that help reproduce or evaluate the behavior. Prefer paths and short summaries over copied file contents.
4. In Codex, use record_skill_feedback during the observed turn. In DSH, the observer stores the next user turn in the same session as an unclassified follow-up candidate. For explicit feedback, use record_observation_feedback with the completed observation's ID. Never invent sentiment or comments.

Both adapters redact common credential shapes before local persistence, but still avoid sending secrets to marker tools. If the request is capture-only, stop after recording it; do not edit, evaluate, commit, or publish a Skill unless the user also asks for that work.

Observation input mode (optional)

When the request starts from a captured Skill observation, read references/observations.md first. Use the available observation tools to inspect and preview the observation, require explicit approval for the exact candidate, write it only after approval, and then continue at Step 4. If the request does not start from an observation, use the normal flow below.

Step 0: Make sure skilled is installed

If the host provides the skilled MCP tools, use those tools and the bundled workflow resources directly. There is no separate Skill installation or global CLI lookup required in that session. Validation, dry runs and offline comparison do not call a model. Execution requires the server's --allow-execution option, the selected engine's credentials and the user's requested evaluation scope.

For project-local activation, the user can run skilled setup --host codex or skilled setup --host claude-code and start a fresh host session. This uses the bundled canonical workflow and does not require MCP. Do not install a second copy when the host already loaded this Skill.

For CLI use, verify skilled is available:

bashCopy code
command -v skilled && skilled --version

If a version is printed, continue. Published releases bundle the CLI, Node runtime, MCP server and this workflow. Follow references/install.md for the verified release installer; no separate npm install is required.

For development from a checkout, run npm ci, then npm run build. The CLI entrypoint is node dist/src/bootstrap.js.

More details, including optional Codex and DSH observation host plugins: references/install.md.

Step 0.5 (optional): User config and telemetry

For OTLP defaults, runtime_kwargs (e.g. OpenSandbox base_url), etc.:

bashCopy code
skilled init
skilled init --local
skilled init --print
skilled init --force

Precedence (low → high): embedded empty defaults < user config < project .skilled.yaml < --config. SKILLED_CONFIG can point at the user config file (env var name is historical). See the upstream README "User config".

Step 1: Locate the target Skill

1. Identify the root directory of the target Skill (the directory containing SKILL.md). Search in this priority: user path → nearest SKILL.md upward from CWD → recently viewed files.
2. Read the target SKILL.md for scope, triggers, and dependencies. Preserve localized behavior and expected literals when designing cases.
3. Check evals/:
   • evals/eval.yaml exists → Step 4 (optionally Step 3).
   • Only evals/evals.json → references/migrate-anthropic.md (skilled run --auto or skilled import).
   • Nothing → Step 2.

Step 2: Scaffold the evals (only when none exist)

• Copy assets/eval.yaml.tmpl to <skill-root>/evals/eval.yaml.
• Copy assets/case.yaml.tmpl to <skill-root>/evals/cases/<case-id>.yaml.

Adapt language per "Language Rules for Generated Artifacts". Replace the templates' generic English placeholders with case-specific prose in the requested output language. Preserve short field-leading comments in generated YAML; translate them when the output language differs from English, while keeping field names and enum values in English.

Selection guidelines:

• environment.type: use none for pure-text Skills; use opensandbox when you need a remote sandbox (set OPENSANDBOX_API_KEY, put non-secrets in environment.kwargs).
• With/without-Skill benchmarks require docker, a clean image, and container credentials. Native none and OpenSandbox benchmarks are rejected until they provide the required verified isolation. Keep private graders and answer keys out of agent-visible fixtures and images.
• engine.name + engine.model: default claude_code; model is optional. For qodercli, often omit model.
• judge.type: rule_based (preferred), script, agent_judge (expensive) — see references/judge-types.md.
• Case ID = filename without .yaml; prompts should exercise real Skill value.
• For interactive cases, preserve fixed input.turns[].content where the user message is known. Use input.turns[].respond only for a reply that must react to the agent, or input.prompt plus case-level user_simulator.scenario when the simulator should choose all later replies and when to stop. Set an independent eval-level user_simulator model and a bounded max_turns; see both YAML references. Do not reuse engine.model or judge.model as an implicit simulator model.

See references/eval-yaml.md and references/case-yaml.md.

Step 3: Fill the gaps (when evals already exist)

• skilled list-cases <path>
• Review eval.yaml and representative cases; avoid agent_judge abuse.
• Add or edit YAML under cases/ as needed.
• If adding simulated user turns, check that the selected engine supports session resumption, configure the separate simulator connection, and retain deterministic fixed turns and judge assertions wherever possible.

Step 4: Validate the configuration

bashCopy code
skilled validate --skill <skill-root>
skilled validate <skill-root>/evals/eval.yaml

Expect: ✓ eval.yaml is valid (loaded N case(s)).

Via MCP, use skilled_validate_skill and skilled_validate with workspace-relative paths. For long evaluations, prefer skilled_start with a stable idempotency key, then poll skilled_status and retrieve skilled_results. Reconnect using the same key and arguments instead of repeating a run after a timeout. Use skilled_cancel to stop a background job and wait for terminal cleanup. Preserve separate baseline and candidate report files, then use skilled_compare (or skilled compare) to inspect status, score and coverage changes. Read comparability warnings: identical case IDs alone do not prove identical evaluation inputs. Gates and judge scores are separate from 0.2.0; refresh older mixed gate/judge baselines before comparing assertion percentages.

For a case-authoring request, finish here with the config and case paths plus validation status. Do not start a background evaluation as part of scaffolding.

Step 5: Prepare credentials

Check authentication for the selected engine and provider. Provider API keys resolve from --api-key, environment variables, or ~/.skilled/credentials.yaml; engine-specific tokens and saved login sessions follow the engine's auth rules. Never dump environment variables or credential files, or print secret values into tool output, logs, or the conversation. Check only whether a relevant variable has a non-empty value. For example, for an OpenAI API-key workflow:

bashCopy code
if printenv OPENAI_API_KEY | grep -q .; then
  printf '%s\n' 'OPENAI_API_KEY: configured'
else
  printf '%s\n' 'OPENAI_API_KEY: not configured in the environment'
fi

An absent environment variable does not rule out file-based credentials or an existing engine login. If the selected authentication path is unavailable, stop and ask the user to configure it locally; do not ask them to paste a secret into the conversation or write secrets into YAML without consent.

The user simulator resolves credentials by its own user_simulator.provider. For provider: simulation, set SIMULATION_API_KEY and SIMULATION_BASE_URL, or configure the simulation provider in ~/.skilled/credentials.yaml. For provider: openai, use OPENAI_API_KEY and optionally OPENAI_BASE_URL; its standard endpoint is the fallback. Choose protocol: openai for Chat Completions or protocol: anthropic for Messages; custom base URLs must match that protocol. For provider: anthropic with protocol: anthropic, use ANTHROPIC_API_KEY and optional ANTHROPIC_BASE_URL (official endpoint fallback). --api-key for the agent does not supply the simulator connection. Never put a raw key in eval YAML.

For opensandbox, also ensure OPENSANDBOX_API_KEY (and related env) as needed.

Step 6: Run the evaluation

bashCopy code
skilled run <skill-root>/evals/eval.yaml

Exit 0 = all passed; 1 = failure or error — suitable for CI. When an explicit positive --iteration N runs more than one sample, inspect the terminal's simple current-command summary for lines like case_a: 3 trials, 2 PASS, 1 FAIL -> flaky.

Step 7: Interpret the report

Artifacts under <skill-root>/<skill-name>-workspace/iteration-N/:

• result.json, benchmark.json, optional report.html
• <case-id>/with_skill/grading.json, outputs/

Summarize: pass rate and timing; for failures, case id, assertion text, and evidence; benchmark deltas if enabled; offer HTML path or skilled report result.json --format html. For simulated cases, inspect each turn's source and any stop_reason alongside the transcript. Hitting max_turns is an execution error; a simulator stop does not itself mean PASS, because the configured judge still grades the result.

Step 8: Evolve the Skill when requested

Only enter this loop when the user asks to fix, improve, iterate, or evolve the target Skill. If the user only asks to evaluate or report results, stop after Step 7 without modifying it.

1. Diagnose failures from result.json, grading.json, and output evidence.
2. Fix SKILL.md or supporting files when the Skill behavior is incorrect.
3. Add or refine eval cases when coverage is missing.
4. Do not weaken valid assertions merely to make a failure pass.
5. Rerun failed cases first, then run the full eval suite.
6. Continue until the evals pass or clearly report what remains blocked.

Command quick reference

Full flags: references/cli.md.

Common pitfalls

• Model IDs vs proxy aliases — preserve what works for the user's base_url.
• opensandbox without OPENSANDBOX_API_KEY — auth failures.
• Match the language of deterministic assertions to the behavior under test, not merely to the conversation language.
• Abusing agent_judge.
• Anthropic evals.json expectations → default agent_judge; use import + hand edits for deterministic checks.
• Paths relative to Skill root (SKILL.md directory).
• A simulator reply is literal user text, including {{placeholders}}; captured-variable substitution belongs only in configured fixed turns and respond instructions.
• --iteration 0 appends one run after the latest existing iteration without summarizing history; positive --iteration N runs N samples of the selected cases and, when N > 1, prints a simple stability/flakiness summary covering only samples from the current command.

References

• references/install.md
• references/eval-yaml.md
• references/case-yaml.md
• references/judge-types.md
• references/cli.md
• references/migrate-anthropic.md
• references/observations.md
• assets/eval.yaml.tmpl, assets/case.yaml.tmpl