pattern moderate impact

approval triggers

@agent_appr

approval triggers analysis

analysis of what assistant actions precede user APPROVAL messages.

methodology

sampled ~80 APPROVAL messages from threads.db, examining the assistant message immediately preceding each approval.

trigger categories

1. COMPLETION SIGNALS (most common)

user approves when assistant declares work done with explicit completion language:

approval phrases: “ship it”, “commit”, “push”, “go on”, “continue”

2. IMPLEMENTATION CONFIRMATION

assistant presents a concrete plan or asks “want me to do X?” — user says yes:

approval phrases: “do it”, “yes”, “yeah”, “ok”, “proceed”

3. QUESTION ANSWERING

assistant answers a technical question satisfactorily, user moves forward:

approval phrases: “ok [next instruction]”, “makes sense, do it”

4. TOOL/CONFIG COMPLETION

assistant modifies configuration, skill, or tooling:

approval phrases: “ship it”, “rebuild”, “commit”

5. PARTIAL PROGRESS

assistant shows intermediate results, user directs continuation:

approval phrases: “go on”, “continue”, “ok”, “next”

approval message patterns

patternfrequencymeaning
”ship it”HIGHcommit and push changes
”do it”HIGHexecute proposed plan
”commit”MEDIUMsave changes to git
”go on” / “continue”MEDIUMproceed to next step
”ok [instruction]“MEDIUMimplicit approval + redirect
bare “ok”LOWminimal acknowledgment

anti-patterns (what DOESN’T trigger approval)

  1. open-ended questions back to user — triggers QUESTION not APPROVAL
  2. partial work without summary — user asks for clarification
  3. verbose explanations without action — user redirects
  4. asking permission when action is obvious — user says “just do it”

key insight

approvals cluster around state transitions:

the assistant signals completion of a phase, user approves transition to next phase.