Claude's /advisor Can't Delegate. A Free Skill That Picks the Mode and Sets the Handoff Rules
November 18, 2026
Claude Code's native /advisor lets a cheaper main model call a stronger one at decision points, but the advisor has no tools, so it cannot split work across agents. The Advisor skill decides when a job needs splitting, which model tier gets each piece, and what rules every subagent follows when it hands work back.
What gap does this fix?
The /advisor [model|off] command has the main model do the work and consult a stronger advisor model. The Claude Code docs say the advisor "tends to consult before committing to an approach, when an error keeps recurring, and before declaring a task done, but the timing is model-driven rather than rule-based." It cannot spawn or direct other agents.
That leaves a job like this uncovered: collect sources, fact-check them, write, then judge the draft. The advisor cannot split that. The other option is Fable 5.2 (Anthropic's top model tier, above Opus) as the main model, delegating lower-level work to cheaper models. The skill chooses between the two and sets the rules for the second.
How does it choose a mode?
| Mode | Use it when | Who does what |
|---|---|---|
A: native /advisor |
One cheaper main model can do the whole task and only the decisions need a stronger model | The cheap model works. Fable 5.2 advises and has no tools |
| B: Fable 5.2 as main | The work has to be split across models | Fable plans, judges and writes the final copy. Cheaper models take the sub-jobs |
The skill starts after the delegate-or-not decision. Whether to delegate at all, and fork versus fresh agent, belongs to Subagent Dispatch Economics. If that answer is "don't delegate," the Advisor skill stops and the main session does the task.
Which model gets which job in Mode B?
| Job | Model |
|---|---|
| Plan, make the calls, write the final copy | Fable 5.2 (main) |
| Collect sources, read-only lookups | Haiku |
| Fact-check each claim against its saved source | Sonnet |
| Cold-read judge of the final draft | Haiku or another cheap model, a fresh agent that is never the writer and never the collector |
Cheap jobs run as fresh agents. A forked subagent inherits the parent model and ignores the per-spawn model override, so a fork cannot be the cheap tier.
What are the handoff rules?
- Every subagent writes its full result to a file, with no length limit.
- The message it returns to the orchestrator is 200 words or fewer and names the file. The cap applies to that return message only, not to the file and not to the final copy.
- Collector output is not trusted until a checker tier that did not produce it passes it. The orchestrator's own drafts go through the checker too.
- Checker verdicts are PASS, FIX, UNSUPPORTED and NO SOURCE.
- The orchestrator keeps a run log with one row per subagent: job, model, tokens as the harness reports them, verdict and output file.
Does the checker tier earn its cost?
On 16 November I ran the pattern with Fable 5.2 as main model and /advisor off:
| Job | Model | Tokens | Result |
|---|---|---|---|
| Read-only lookup of my tool's database | Haiku | 91,409 | Rows found, no token or cost data recorded |
| Fact-check of the first generated skill | Sonnet | 78,437 | 28 claims: 8 PASS, 4 DRIFT, 11 INVENTED, 5 MISSING |
| Fact-check of Fable's own LinkedIn drafts | Sonnet | 89,861 | 56 rows: 37 PASS, 5 FIX, 10 UNSUPPORTED, 4 NO SOURCE |
| Cold-read judge of those drafts | Haiku | 79,634 | 7 drafts scored 2 to 4 of 5, 2 AI tells |
That is 339,341 subagent tokens. All four returned 200 words or fewer and wrote a file. Fable's own token count is not visible from inside the session. The checker flagged 19 lines in the top model's drafts, so the top model needs checking too.
Is it cheaper than one Fable agent?
I don't know. Anthropic's April blog post reports Sonnet with an Opus advisor at +2.7 points on SWE-bench Multilingual and 11.9% lower cost per task. Those were older models, with Opus 4.6 as the advisor, not Fable 5.2. The current platform docs say only that Opus keeps total cost similar or lower and that Fable 5.2 maximizes the quality lift. I found no Fable advisor number from Anthropic, so the skill carries a "Never" rule against presenting the old benchmark as a Fable result.
I have not measured whether Mode B beats one Fable agent doing everything. I did run a smoke test: a fresh Sonnet agent followed the skill on four tasks with pass conditions fixed beforehand (a variable rename, a fact-checked blog post from six sources, a flaky test, and 40 tickets summarised for a CEO). It chose the expected path all four times. That is four prompts, not an evaluation.
How was it made?
The first draft came from the Context Engineering pipeline in Prompt Optimizer, given a goal and a one-page reference file of verified facts. An evaluation of 4 skills, with and without a reference, 2 runs each, 16 generations in all, showed the difference:
- With a reference file: 8 of 8 usable. 74 of 80 required rules came through and every limit kept its qualifier.
- From a goal only: 0 of 8 worth using. Each was well-formed but hollow.
The published file is a hand rebuild from the reference file: one invented line removed, the companion-skill pointer restored, the Mode A limits, fork rule, checker gate, Never and Requirements sections added, and the description rewritten.
How do I install it?
- Download
advisor.md. - Save it as
.claude/skills/advisor/SKILL.mdin your project or in~/.claude/skills/. - It applies when a task could run under native
/advisoror with Fable as main model delegating to cheaper models.
Native /advisor is experimental. It requires the Anthropic API, Claude Platform on AWS, or Claude Models on Google Cloud (Anthropic managed). Fable 5.2 requires Claude Code v2.1.321 or later, and on some plans Fable usage, including as advisor, bills to usage credits. The source is in the GitHub repository.
FAQ
Can the native advisor delegate work?
No. It has no tools, so it can advise but cannot spawn or direct agents. Splitting work needs Fable 5.2 as the main model.
Why must cheap jobs be fresh agents?
A forked subagent inherits the parent model and ignores the model override, so it would run on the expensive tier.
Why check the top model's drafts?
In the 16 November run the Sonnet checker flagged 19 lines across Fable's own drafts as FIX, UNSUPPORTED or NO SOURCE.
Does it decide whether to delegate?
No. That is Subagent Dispatch Economics. This skill starts after that decision.
Related: Agent Prompt Architecture covers designing the prompt and tools of the agent itself.