Claude Advise

Claude Advise

Use Claude as an external text consultant. Keep Codex responsible for gathering context, judging the response, and reporting what is genuinely useful.

Workflow

1. Build a compact text brief.

   • State the audience, objective, current text or concept, constraints, and exact questions.
   • Distinguish established conclusions from open questions so Claude cannot earn credit by repeating the brief.
   • Include only the minimum factual context required.
   • Never send repository contents, source code, credentials, private raw correspondence, or unrelated files.

2. Announce the paid external call and its controls.

   • Default to one direct-API call with claude-fable-5, low effort, and a 4,096-token output ceiling for routine writing or advice.
   • Use medium or high effort only when the task needs deeper synthesis and the added latency and spend are justified.
   • Use Haiku only to verify a previously untested installation or credential path. Do not rerun the connection check every session once the setup is known to work.
   • Use the CLI path when the user specifically asks for Claude Code. The wrapper replaces its coding prompt and removes every built-in tool.

3. Run the bundled wrapper from the skill directory.

   scripts/claude-advise.sh --api --model claude-fable-5 --effort low --max-tokens 4096 brief.txt
   

   For the answer-only Claude Code path:

   scripts/claude-advise.sh --cli --model claude-fable-5 --effort low --max-tokens 4096 --max-budget-usd 0.50 brief.txt
   

   For a one-time cheap connectivity check:

   scripts/claude-advise.sh --check
   

4. Inspect the receipt before using the advice.

   • Confirm the returned model or CLI modelUsage names the requested model.
   • Require stop_reason: end_turn for a complete API answer.
   • Treat max_tokens as a truncated result. Narrow the follow-up or deliberately raise the ceiling; do not silently present it as complete.
   • Treat refusal as a real result. Fable can classify benign life-science wording as bio. Remove subject-matter detail that is irrelevant to an editorial task and retry at most once. Never add an automatic fallback model unless the user authorizes it.
   • Report CLI total_cost_usd when present. For direct API calls, report input, output, and thinking-token usage and calculate cost only from current official Anthropic pricing.

5. Audit the answer rather than forwarding it.

   • Separate genuinely new findings from ideas already supplied in the brief.
   • Identify advice that is actionable, advice that needs evidence, and advice that conflicts with product facts or user direction.
   • Never turn invented examples, unsupported claims, statistics, testimonials, or provenance into product copy.
   • Lead the user with the new conclusions, then summarize the model, usage, spend, refusal, or truncation details.

6. Delete temporary briefs and responses unless the user asks to retain them. Never commit site-specific prompts or model output to a skill repository.

Security and cost rules

• Read ANTHROPIC_API_KEY from the environment or ~/.env; never print, copy, or commit it.
• Treat every prompt as data sent to an external provider. Remove secrets, personal data, and sensitive raw material.
• Keep one model turn, no session persistence, no tools, and a deliberate output limit.
• Do not use Claude to inspect or modify code under this skill. Codex may summarize relevant behavior into prose first.
• Stop after one failed substantive call and diagnose the failure before spending again.

Wrapper behavior

scripts/claude-advise.sh defaults to the direct Messages API because it exposes the exact model, stop reason, refusal category, and token usage. Its CLI mode uses --bare, a replacement system prompt, --tools "", --disallowedTools "*", one turn, a dollar ceiling, and no saved session.