Troubleshooting
Agent Keeps Asking the Same Question?
Break the loop when your agent repeatedly asks for info you already gave.
A common trust-break in community support: you answered once, maybe twice, and your agent still asks the same clarifying question. This is usually a context-shape problem, not a capability problem.
Root cause: repeated questions usually happen when critical details are scattered across messages, mixed across surfaces (dashboard/DM/server), or missing required format constraints.
Fast 4-step loop breaker
1) Consolidate requirements in one block
Put all critical constraints into one compact message: goal, input files, output format, and deadline.
2) Force acknowledgement before execution
Require the agent to restate the key requirements before doing work. This catches missing details early.
3) Lock uncertainty handling
Tell the agent to ask at most one clarifying question only if a required field is missing; otherwise proceed.
4) Add evidence requirements
Require concrete output proof (path, line refs, or quoted excerpt). This shifts behavior from chat looping to execution.
Paste-ready recovery prompt
You already have enough context to proceed.
Before execution, restate exactly:
1) goal
2) required output format
3) file/path target (if any)
4) deadline/constraints
Then execute.
Only ask ONE clarifying question if a required field is missing.
If nothing required is missing, do not ask again—proceed and return evidence of completion.
High-friction cases (and fixes)
- Dashboard vs Discord mismatch: repeat key instructions in the same surface where you want execution.
- Long chat drift: start a fresh session and provide a compact handoff instead of replaying the whole thread.
- Vague output target: specify exact file name/path or exact message format expected.
- Multiple tasks in one ask: split into numbered steps and require check-in after each completed step.
Important: if you keep adding new constraints mid-task, the agent may keep re-clarifying. Freeze scope for one run, then iterate.
Known-good command pattern
Task: [single goal]
Inputs: [exact files/links/messages]
Output: [exact format]
Done criteria: [objective checks]
Ask policy: If required fields are present, proceed with no further questions.
Evidence: Return [path + quote + summary].
Advanced edge cases
- Approval-gated execution: the agent may keep asking variants of the same question when the real blocker is an unapproved command. Check for pending Approval Required first.
- Tool-capability mismatch: if a requested action is not available in the current session/surface, the agent may loop on clarifications instead of stating the limitation clearly.
- Conflicting rules: if SOUL/AGENTS instructions conflict with your immediate request, the model may keep seeking reconfirmation. Ask it to list the conflict explicitly before proceeding.
- Hidden required field: some tasks require one missing key (exact destination, date/timezone, account name). Tell the agent to output a missing-fields list once, then continue.
Stronger recovery prompt (anti-loop)
You are currently in a repeated-question loop.
Do this now:
1) Extract all confirmed requirements from this thread.
2) Return them as a checklist with ✅/❌ status.
3) If any required field is missing, ask ONE final question containing only the missing fields.
4) If no required fields are missing, execute immediately.
5) Return completion evidence: exact output path/location + short quoted proof.
Do not repeat previously answered questions.
When to escalate in #help
If repetition continues after the loop-break prompt, share:
- The exact prompt that triggered repetition
- Which surface you used (dashboard, DM, server channel)
- What requirement was repeated back incorrectly
- Whether an approval prompt was pending
- One screenshot or quoted reply showing the loop
Bottom line: repeated-question loops are usually structure + blocker issues. One compact requirement block, one-question ask cap, and explicit blocker checks resolve most loops fast.