Improve advisor tool steering via workflow gate triggers and system-prompt reinforcement.
Current behavior
The advisor tool description steers by task category (architecture, algorithms, concurrency, etc.) and two loose temporal signals ("before committing to a plan" and "before declaring work complete"). This underspecifies when in a turn the advisor should be reached for, and omits common high-value trigger moments like before making a code edit or before posting a PR.
Gap
Anthropic's "think" tool research found that moment-based workflow triggers outperform category lists, with an optimized system-prompt section producing a 54% relative improvement over the tool description alone. The current advisor description:
- Does not explicitly name "making a code change" as a trigger, despite being the most common hard action
- Uses vague temporal signals ("after initial orientation") instead of concrete workflow gates
- Does not reinforce triggers in the system prompt, where the model is more likely to act on them
Proposed changes
-
Revise tool description — restructure around three explicit workflow gates:
- After initial exploration, before committing to a plan or writing code
- Before making any non-trivial code change (especially architecture, concurrency, security, broad refactors)
- Before declaring complex work done, posting a PR, or reporting results
-
Add <advisor-policy> block to the system prompt — a short reinforcement section near <tool-policy> that names the three gates explicitly, so the model encounters the guidance during reasoning, not just at tool-selection time.
Prior art
Improve advisor tool steering via workflow gate triggers and system-prompt reinforcement.
Current behavior
The advisor tool description steers by task category (architecture, algorithms, concurrency, etc.) and two loose temporal signals ("before committing to a plan" and "before declaring work complete"). This underspecifies when in a turn the advisor should be reached for, and omits common high-value trigger moments like before making a code edit or before posting a PR.
Gap
Anthropic's "think" tool research found that moment-based workflow triggers outperform category lists, with an optimized system-prompt section producing a 54% relative improvement over the tool description alone. The current advisor description:
Proposed changes
Revise tool description — restructure around three explicit workflow gates:
Add
<advisor-policy>block to the system prompt — a short reinforcement section near<tool-policy>that names the three gates explicitly, so the model encounters the guidance during reasoning, not just at tool-selection time.Prior art