tension-finder

Prompt file imported from johntnanney/gradience (.claude/commands/tension-finder.md). Fill in {{arguments}} before use. Copyright stays with the author.

Run a cross-document tension audit using the protocol defined in /Users/john/code/gradience/research_review/tension_finder_prompt.md.

The audit is on-demand only — never run on a schedule. Trigger when the user senses drift between commitments and current state. Typical triggers: before a lock amendment, before manuscript submission, after absorbing a significant external input.

Steps

1. Read the prompt file at /Users/john/code/gradience/research_review/tension_finder_prompt.md. The agent-prompt body is everything from the line ## Prompt to pass to the agent (everything below this line) onward (skip the header preamble — it's documentation for the user, not for the spawned agent).

2. Construct the spawned-agent prompt by composing in this order:

   • If {{arguments}} is non-empty, prepend exactly:

     Scope override (user-supplied at invocation): {{arguments}}
     
     Apply this scope override to Phase 1; otherwise follow the prompt as written.
     
     

   • Append the agent-prompt body from step 1.
   • Append a trigger note: a single line Trigger note: on-demand audit, $(date +%Y-%m-%d). (substitute today's date from the env block; if {{arguments}} named a trigger reason, use that reason instead of "on-demand audit").

3. Spawn a fresh general-purpose Agent using the Agent tool:

   • subagent_type: general-purpose
   • description: Tension audit
   • prompt: the composed prompt from step 2

   Do not run any other agents in parallel for this command — the audit must complete in a fresh context with no other concurrent work.

4. After the agent returns, do not paraphrase or expand its findings. Verify the agent wrote the report file by checking /Users/john/code/gradience/research_review/tension_audit_YYYY-MM-DD.md exists. Then confirm to the user with:

   • The report path.
   • The findings count (extracted from the file's "Findings count:" header line).
   • A one-line note that the report is theirs to read and act on.

   Do not editorialize on the findings. The prompt's anti-patterns explicitly forbid the audit agent from proposing resolutions; the same discipline applies to your post-completion summary.

Constraints

• Never edit program documents based on audit findings — that's the user's role.
• Never re-summarize the agent's report contents in this conversation. The report is the artifact; pointing at it is the deliverable.
• If the report file already exists for today's date when the agent finishes, alert the user — the agent may have overwritten a prior run. The convention is one audit per day; if a second is genuinely needed, the user will rename or move the prior one before re-running.