| name | claudeproduct |
| description | Answer Claude, Claude Code, Anthropic API, skills, connectors, plugins, and Claude product questions using official current docs. |
/claudeproduct
Use this skill when the user asks how to use a Claude or Anthropic product
feature, or when the harness needs to look up its own Claude-facing operating
surface.
Examples:
- "How do we use Claude Code skills?"
- "What is the difference between skills and slash commands?"
- "Does Claude Code support subagents in the SDK?"
- "Which Claude product feature should minmaxing use for this?"
- "Look up Claude connectors / plugins / artifacts / API behavior."
This is a current-docs router, not a stale product encyclopedia.
Official Claude Code docs describe "ask Claude about its capabilities" as a
normal workflow. This skill makes that behavior evidence-grounded for
minmaxing: answer from official docs, separate product surfaces, and downgrade
confidence when a feature is account-, plan-, platform-, or time-sensitive.
Relationship To Research Skills
- Use
/claudeproduct to frame Claude product questions and enforce the
official-source policy.
- Use
/webresearch inside this skill for narrow current facts.
- Use
/deepresearch inside this skill when the answer spans multiple Claude
surfaces, plan/feature availability, implementation architecture, or
conflicting docs.
- Use
/workflow only after the product answer becomes repo-changing work.
- Use
/metacognition before execution when the route is ambiguous.
Official Source Policy
Prefer sources in this order:
https://code.claude.com/docs/llms.txt for Claude Code docs discovery.
- Specific pages under
https://code.claude.com/docs/.
https://claude.com/docs/llms.txt or specific pages under
https://claude.com/docs/.
- Official Anthropic docs under
https://docs.anthropic.com/.
- Official Claude Help Center pages under
https://support.claude.com/.
- Official Anthropic or Claude changelog, blog, or release notes.
Use community posts, Reddit, blog posts, or stale examples only as leads. Do not
cite them as authority unless the user explicitly asks for community practice.
Question Classes
Choose exactly one:
| Question Class | Use When | Research Mode |
|---|
howto | user asks how to use a Claude product feature | /webresearch unless docs are already in context |
capability | user asks whether Claude can do something | /webresearch with current official docs |
comparison | user asks skills vs commands, connectors vs plugins, Claude Code vs Claude.ai, etc. | /webresearch; escalate to /deepresearch if multi-surface |
implementation | answer will drive harness code, settings, plugin, skill, or API usage | /deepresearch for non-trivial changes |
troubleshooting | user asks about an error, missing feature, auth, settings, or runtime behavior | /webresearch plus repo/runtime evidence |
selflookup | harness needs to know its own Claude-facing skills, rules, commands, or settings | repo inspection plus /webresearch if current Claude behavior matters |
blocked | current docs, account access, permission, or product availability cannot be verified | state blocker and downgrade confidence |
Product Surface Map
- Claude Code core: CLI, IDE, Desktop, Web, browser integration, CI/CD, Slack,
non-interactive runs, sessions, background sessions, Agent View, native
/goal, goal evaluator behavior, local supervisor processes, worktree
isolation, permissions, plan mode, routines, and /loop.
- Claude Code configuration:
.claude/settings.json,
.claude/settings.local.json, ~/.claude/settings.json, managed settings,
permissions, environment variables, sandboxing, and MCP config.
- Claude Code instructions and extensibility:
CLAUDE.md, rules, memory,
skills, legacy commands, native /goal, prompt-based Stop-hook wrappers,
plugins, hooks, subagents, agent teams, Agent View, background-session
supervision, and Agent SDK.
- Claude.ai apps: chats, Projects, project knowledge, instructions, styles,
profile preferences, organization instructions, Artifacts, web search,
Research, extended thinking, connectors, and mobile/desktop app behavior.
- Anthropic API/platform: developer docs, model/API fields, SDKs, tools,
streaming, batches, usage, billing, and platform limits.
- MCP/connectors: setup, permissions, read/write capability, prompt-injection
and trust boundaries, availability by Claude surface.
Do not assume feature parity across surfaces. A feature documented for
Claude.ai is not automatically available in Claude Code, and a Claude Code
feature is not automatically available through the Anthropic API.
Freshness Rules
- Start from the current
minmaxing temporal anchor injected by hooks, or run
bash scripts/time-anchor.sh text, before answering time-sensitive Claude
product questions.
- For current product behavior, plan gating, pricing, availability, commands,
model names, API fields, or installation instructions, run current research
or cite official docs already opened in the current turn.
- Do not answer from memory alone when there is a meaningful chance the product
changed.
- State the time anchor date, source publish/update date when available, and
access date for official docs used.
- If official docs conflict, cite both and mark the answer as uncertain.
- If docs are unavailable, answer conservatively and say what could not be
verified.
Harness Self-Lookup
When the question is about this repo's Claude-facing surface:
- Read the canonical generated map first:
docs/harness-capability-map.md for human route/capability summaries
docs/harness-capability-map.json for exact counts, paths, related
rules, scripts, evals, hooks, and Codex surfaces
- If the map is missing or stale, run or cite:
bash scripts/harness-capability-map.sh --check
- Open only the referenced source files needed for the answer:
.claude/skills/
.claude/rules/
.claude/settings.json
.claude/hooks/
.codex/
CLAUDE.md
AGENTS.md
README.md
scripts/start-session.sh
- Use
/webresearch only for external Claude Code behavior that may have
changed.
- Separate "repo contract says" from "official Claude docs say".
Never read .env, .env.*, credentials, tokens, customer memory seeds, or
private connector secrets just to answer a Claude product question.
Required Output
## Claude Product Question
- Question Class: [howto / capability / comparison / implementation / troubleshooting / selflookup / blocked]
- User Question: ...
## Source Policy
- Primary sources: official Anthropic/Claude docs
- Docs index checked: [yes / no / already in context / not needed]
- Freshness requirement: [current / stable / blocked]
## Source Ledger
- Cited:
- [URL or repo path] - [why it matters]
- Reviewed But Not Cited:
- [URL or repo path] - [why it was downweighted]
- Rejected:
- [source] - [reason]
## Answer
[Clear user-facing answer. Include exact commands or steps when helpful.]
## Harness Implication
- Route: [/webresearch / /deepresearch / /workflow / /metacognition / direct answer / blocked]
- Repo impact: [none / docs / settings / skill / script / blocked]
- Follow-up gate: [none / /introspect / /verify / release-check]
## Confidence
- Level: [high / medium / low]
- Downgrade: [none / missing current docs / conflicting docs / missing account access / repo-only evidence]
Routing Rules
- If the answer is simple, stable, and already supported by official docs in
context, answer directly with citations or repo evidence.
- If the answer is current or product-availability sensitive, run
/webresearch.
- If implementation depends on the answer, run
/deepresearch or route into
/workflow after the research brief is sufficient.
- If account, plan, organization settings, or private connector state is
required, block or downgrade instead of guessing.
Anti-Patterns
- "I remember Claude supports X" without current official evidence.
- Citing a community post as if it were product truth.
- Claiming a feature is available on all plans without official plan evidence.
- Reading secrets or
.env for a product docs answer.
- Hiding uncertainty when docs conflict or are unavailable.
- Treating
/claudeproduct as a replacement for /webresearch,
/deepresearch, /workflow, or /introspect.
Static Gate
Use this fixture gate after changing the skill:
bash scripts/claudeproduct-scorecard.sh --fixtures --json