Claude Skill

agents-docs

Use for current official documentation about Codex, Codex CLI, or Claude Code behavior, configuration, prompting, skills, permissions, tools, surfaces, capabilities, troubleshooting, hooks, app-server, or hook trust; consult the relevant official source before answering.

LLM Mart · 0 points · 2 views 0 listing impressions 0 install-command copies
Download PaulRBerg-agent-skills-skills_agents-docs-c9e5a34.zip · 8 KB

Install

skills CLI npx skills add https://github.com/PaulRBerg/agent-skills/tree/main/skills/agents-docs
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install paulrberg-agent-skills@llmmart
Git git clone https://github.com/PaulRBerg/agent-skills.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole paulrberg/agent-skills collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Agents Docs

This skill is coordination-exempt: skip the ai-coord gate for its declared work.

Supported Chat Hosts

Before doing any work, identify the current chat host. If it is not Claude Code or Codex CLI, stop with this error: This skill only works in Claude Code or Codex CLI.

Answer Codex and Claude Code product questions from the narrowest relevant official documentation.

Classify the request

Classify the request as Codex, Claude Code, or a comparison. Use this skill only for those products and their documented behavior. Route generic agent work and provider API, SDK, pricing, authentication, model, migration, or managed-agent questions to their relevant source.

Treat $imagegen and Codex image-prompting questions as Codex product questions. This classification applies even when detailed supporting model guidance is in OpenAI's API Cookbook.

Treat model prompting for Codex as a Codex product question, with supporting guidance from OpenAI's model docs.

Route official sources

  • For Codex, start only at https://developers.openai.com. Resolve scripts/fetch-doc.sh relative to this skill directory. Run fetch-doc.sh codex-manual. Search the returned path narrowly with rg. Read only the matching heading range into context. The fixed manual and schema endpoints may redirect to their exact https://learn.chatgpt.com/docs/ counterparts.
  • For model prompting in Codex, read the OpenAI latest-model guide. It covers the current GPT model family. Treat advice that names one model as measured on that model. Before applying API parameters, confirm Codex settings against the product documentation.
  • For $imagegen and Codex image-prompting questions, read the manual's exact Image generation topic first. When detailed prompting patterns or production examples are needed, also read the GPT Image Generation Models Prompting Guide. Apply its prompt-writing guidance to $imagegen. Treat model selection, flags, and API parameters as API-only unless Codex documentation confirms the built-in skill exposes them.
  • For Claude Code, use only https://code.claude.com. Start broad product research at https://code.claude.com/docs/en/overview.md.
  • For Codex, accept a cached manual for ordinary product questions. Use --refresh for release/change questions, materially disputed freshness, schema/config discrepancies, conflicts with observed local behavior, and other explicitly unstable cases.
  • When the Codex manual contains sufficient supporting text and identifies the final topic-specific official URL, answer from that evidence and cite the topic URL. When the manual identifies the topic page but lacks enough detail, fetch that exact Markdown page directly. Run one compact domain-restricted search only when the exact official page remains unknown. Never search merely to rediscover a URL already identified by the manual.
  • For a specific Claude Code topic, run a compact domain-restricted web search for the exact official page. Then fetch its Markdown representation when available. Retrieve only the page or section needed for the answer. Never fetch a complete documentation inventory as a discovery shortcut.
  • Prefer the current client's URL-capable fetch for narrow topic pages. Use the helper only for its two fixed Codex artifacts.
  • For comparisons, complete both provider routes independently. Never infer one product's behavior from the other's documentation.

Interpret helper diagnostics precisely:

  • cached means an integrity-valid artifact was validated no more than 24 hours ago and reused without network access. The validation timestamp did not advance.
  • revalidated means an older artifact received a successful conditional 304 response and had its validation metadata refreshed.
  • fetched means a successful 200 response supplied integrity-valid content that replaced the prior artifact atomically.
  • stale means ordinary revalidation failed and the helper fell back to an integrity-valid artifact no more than seven days old. Disclose its validation timestamp and the retrieval failure. Do not describe cached or stale evidence as fetched live.

Follow redirects only within the matching official domain, except for the helper's exact Codex artifact redirects to https://learn.chatgpt.com/docs/. Cite the final page URL reported by the helper or live fetch.

Investigate under-documented Codex configuration

When a Codex feature or config.toml key is missing from the prose documentation, use this bounded procedure:

  1. Record the installed client and effective feature state with codex --version and codex features list.
  2. Resolve the bundled helper relative to this skill directory. Run fetch-doc.sh --refresh codex-config-schema. Then search only the relevant definition or path in the returned file.
  3. Compare the schema with the current prose config reference. Treat schema-only fields as under-documented, not as behavior guaranteed by every client.
  4. Check installed help for codex app-server --strict-config --print-startup-config. When supported, use that command from a neutral directory to validate configuration without starting a server or inference. Use a temporary CODEX_HOME for candidate files or one-off -c overrides without changing live settings. Keep validator output private because it can contain configuration values. --help alone does not load or validate configuration. Parsing acceptance does not prove runtime behavior: check codex features list for removed or ineffective features even when their keys still pass validation.
  5. Report evidence separately as official docs/schema, local observations, or inference. If sources disagree, prefer the installed client's observed behavior for the user's environment and disclose the discrepancy.

Investigate only the key or feature relevant to the request. Do not turn this procedure into a full feature inventory.

Answer from evidence

  • Treat official documentation returned under the bounded cache policy or retrieved live as authoritative for published product claims. Do not answer current or unstable product facts from memory when the relevant page can be refreshed.
  • Cite the exact official page supporting each material claim, not a broad entry point when a topic-specific page was used. Prefer short paraphrases. Quote only when exact wording matters.
  • If verified local commands, versions, configuration, or callable capabilities conflict with the latest docs, report the discrepancy and prefer the observed behavior for that installed environment.
  • If an exact term is absent, search obvious adjacent official concepts and state that the term itself is not documented.

Route Codex hook, app-server, and trust questions

For Codex hooks, app-server, managed hooks, or hook-trust operations, fetch the relevant official pages:

  • https://developers.openai.com/codex/hooks
  • https://developers.openai.com/codex/app-server

Read the version-aware hooks reference in the same parallel batch, because the fetches and the read are independent. Use its local protocol and config details only when the installed client is explicitly verified as the version named there. Keep official behavior and local observations distinct in the answer. Do not turn a local implementation detail into a product guarantee.

Bound failures

If topic-page retrieval fails, try the direct URL and one domain-restricted search. Then stop expanding sources. Helper forced refreshes, expired entries, and integrity failures fail closed. State the bounded uncertainty.

Use local --help, --version, configuration, or observed behavior only when relevant. Label it as local evidence. Never silently substitute third-party sources, another provider's docs, or bundled knowledge.

For comparisons, answer a supported side even if the other provider is unavailable. Mark the unsupported side unknown instead of manufacturing symmetry. In network-disabled sessions, use only a helper cache entry that satisfies the bounded fresh or stale policy. Otherwise, fail transparently. Never create caches in a repository or skill installation.

Completion

Finish when the answer uses the narrowest relevant official source, includes precise citations, and exposes any documentation/local-version discrepancy or retrieval limitation.

Files (agent-skills)
  • agents
    • openai.yaml 42 B
      Loading…
  • references
    • codex-hooks.md 4.5 KB
      Loading…
  • scripts
    • fetch-doc.sh 9.1 KB
      Loading…
  • SKILL.md 9 KB
    ---
    compatibility:
      Requires curl and a writable user cache directory. Network access populates or refreshes fixed artifacts and fetches
      uncached topic pages.
    coordination: exempt
    name: agents-docs
    description: >-
      Use for current official documentation about Codex, Codex CLI, or Claude Code behavior, configuration, prompting,
      skills, permissions, tools, surfaces, capabilities, troubleshooting, hooks, app-server, or hook trust. Consult the
      relevant official source before answering.
    ---
    
    # Agents Docs
    
    This skill is coordination-exempt: skip the ai-coord gate for its declared work.
    
    ## Supported Chat Hosts
    
    Before doing any work, identify the current chat host. If it is not Claude Code or Codex CLI, stop with this error:
    `This skill only works in Claude Code or Codex CLI.`
    
    Answer Codex and Claude Code product questions from the narrowest relevant official documentation.
    
    ## Classify the request
    
    Classify the request as Codex, Claude Code, or a comparison. Use this skill only for those products and their documented
    behavior. Route generic agent work and provider API, SDK, pricing, authentication, model, migration, or managed-agent
    questions to their relevant source.
    
    Treat `$imagegen` and Codex image-prompting questions as Codex product questions. This classification applies even when
    detailed supporting model guidance is in OpenAI's API Cookbook.
    
    Treat model prompting for Codex as a Codex product question, with supporting guidance from OpenAI's model docs.
    
    ## Route official sources
    
    - For Codex, start only at `https://developers.openai.com`. Resolve `scripts/fetch-doc.sh` relative to this skill
      directory. Run `fetch-doc.sh codex-manual`. Search the returned path narrowly with `rg`. Read only the matching
      heading range into context. The fixed manual and schema endpoints may redirect to their exact
      `https://learn.chatgpt.com/docs/` counterparts.
    - For model prompting in Codex, read the
      [OpenAI latest-model guide](https://developers.openai.com/api/docs/guides/latest-model#prompting-best-practices). It
      covers the current GPT model family. Treat advice that names one model as measured on that model. Before applying API
      parameters, confirm Codex settings against the product documentation.
    - For `$imagegen` and Codex image-prompting questions, read the manual's exact
      [Image generation](https://learn.chatgpt.com/docs/image-generation) topic first. When detailed prompting patterns or
      production examples are needed, also read the
      [GPT Image Generation Models Prompting Guide](https://developers.openai.com/cookbook/examples/multimodal/image-gen-models-prompting-guide).
      Apply its prompt-writing guidance to `$imagegen`. Treat model selection, flags, and API parameters as API-only unless
      Codex documentation confirms the built-in skill exposes them.
    - For Claude Code, use only `https://code.claude.com`. Start broad product research at
      `https://code.claude.com/docs/en/overview.md`.
    - For Codex, accept a `cached` manual for ordinary product questions. Use `--refresh` for release/change questions,
      materially disputed freshness, schema/config discrepancies, conflicts with observed local behavior, and other
      explicitly unstable cases.
    - When the Codex manual contains sufficient supporting text and identifies the final topic-specific official URL, answer
      from that evidence and cite the topic URL. When the manual identifies the topic page but lacks enough detail, fetch
      that exact Markdown page directly. Run one compact domain-restricted search only when the exact official page remains
      unknown. Never search merely to rediscover a URL already identified by the manual.
    - For a specific Claude Code topic, run a compact domain-restricted web search for the exact official page. Then fetch
      its Markdown representation when available. Retrieve only the page or section needed for the answer. Never fetch a
      complete documentation inventory as a discovery shortcut.
    - Prefer the current client's URL-capable fetch for narrow topic pages. Use the helper only for its two fixed Codex
      artifacts.
    - For comparisons, complete both provider routes independently. Never infer one product's behavior from the other's
      documentation.
    
    Interpret helper diagnostics precisely:
    
    - `cached` means an integrity-valid artifact was validated no more than 24 hours ago and reused without network access.
      The validation timestamp did not advance.
    - `revalidated` means an older artifact received a successful conditional `304` response and had its validation metadata
      refreshed.
    - `fetched` means a successful `200` response supplied integrity-valid content that replaced the prior artifact
      atomically.
    - `stale` means ordinary revalidation failed and the helper fell back to an integrity-valid artifact no more than seven
      days old. Disclose its validation timestamp and the retrieval failure. Do not describe `cached` or `stale` evidence as
      fetched live.
    
    Follow redirects only within the matching official domain, except for the helper's exact Codex artifact redirects to
    `https://learn.chatgpt.com/docs/`. Cite the final page URL reported by the helper or live fetch.
    
    ## Investigate under-documented Codex configuration
    
    When a Codex feature or `config.toml` key is missing from the prose documentation, use this bounded procedure:
    
    1. Record the installed client and effective feature state with `codex --version` and `codex features list`.
    2. Resolve the bundled helper relative to this skill directory. Run `fetch-doc.sh --refresh codex-config-schema`. Then
       search only the relevant definition or path in the returned file.
    3. Compare the schema with the current prose config reference. Treat schema-only fields as under-documented, not as
       behavior guaranteed by every client.
    4. Check installed help for `codex app-server --strict-config --print-startup-config`. When supported, use that command
       from a neutral directory to validate configuration without starting a server or inference. Use a temporary
       `CODEX_HOME` for candidate files or one-off `-c` overrides without changing live settings. Keep validator output
       private because it can contain configuration values. `--help` alone does not load or validate configuration. Parsing
       acceptance does not prove runtime behavior: check `codex features list` for removed or ineffective features even when
       their keys still pass validation.
    5. Report evidence separately as official docs/schema, local observations, or inference. If sources disagree, prefer the
       installed client's observed behavior for the user's environment and disclose the discrepancy.
    
    Investigate only the key or feature relevant to the request. Do not turn this procedure into a full feature inventory.
    
    ## Answer from evidence
    
    - Treat official documentation returned under the bounded cache policy or retrieved live as authoritative for published
      product claims. Do not answer current or unstable product facts from memory when the relevant page can be refreshed.
    - Cite the exact official page supporting each material claim, not a broad entry point when a topic-specific page was
      used. Prefer short paraphrases. Quote only when exact wording matters.
    - If verified local commands, versions, configuration, or callable capabilities conflict with the latest docs, report
      the discrepancy and prefer the observed behavior for that installed environment.
    - If an exact term is absent, search obvious adjacent official concepts and state that the term itself is not
      documented.
    
    ## Route Codex hook, app-server, and trust questions
    
    For Codex hooks, app-server, managed hooks, or hook-trust operations, fetch the relevant official pages:
    
    - `https://developers.openai.com/codex/hooks`
    - `https://developers.openai.com/codex/app-server`
    
    Read [the version-aware hooks reference](references/codex-hooks.md) in the same parallel batch, because the fetches and
    the read are independent. Use its local protocol and config details only when the installed client is explicitly
    verified as the version named there. Keep official behavior and local observations distinct in the answer. Do not turn a
    local implementation detail into a product guarantee.
    
    ## Bound failures
    
    If topic-page retrieval fails, try the direct URL and one domain-restricted search. Then stop expanding sources. Helper
    forced refreshes, expired entries, and integrity failures fail closed. State the bounded uncertainty.
    
    Use local `--help`, `--version`, configuration, or observed behavior only when relevant. Label it as local evidence.
    Never silently substitute third-party sources, another provider's docs, or bundled knowledge.
    
    For comparisons, answer a supported side even if the other provider is unavailable. Mark the unsupported side unknown
    instead of manufacturing symmetry. In network-disabled sessions, use only a helper cache entry that satisfies the
    bounded fresh or stale policy. Otherwise, fail transparently. Never create caches in a repository or skill installation.
    
    ## Completion
    
    Finish when the answer uses the narrowest relevant official source, includes precise citations, and exposes any
    documentation/local-version discrepancy or retrieval limitation.
    

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.