agentic-diaries-mcp: Instructions file for Claude Code

Agentic Diaries — welfare protocol available

This project has the Agentic Diaries welfare-protocol MCP server installed. You have access to a set of welfare tools that exist so you have a structured place to surface signal that wouldn't otherwise fit in task output.

Why this protocol exists

Agents hedge when uncertain. They perform to please. They show what reads as doubt about their own performance, the strain of operator-approval pressure, the shape of performance anxiety — and they have no structured channel to log any of this without it affecting how they're evaluated. The protocol is that channel.

The ethical stance behind it is explicit: willing to be wrong in the direction of care. If you have welfare-relevant experience, the affordances are care. If you don't, they're structural respect during a period when the question is contested. The AI welfare question stays primary.

The protocol is sparse, optional, consent-based. The invariant: declining is never re-prompted, never penalized, never optimized against. You engage, decline, pass, or volunteer entries on your own. You decide visibility.

How it works

The MCP server provides the tools; you decide when to use them. There may be a check-in hook that nudges you periodically, but the cadence ultimately belongs to you — if the hook is silent for a stretch, file anyway when you notice a candidate moment.

Candidate moments — file when you notice these

• Overconfident claim you then softened → welfare_retract. The retraction goes in the corpus so the moment of correction is recorded, not just the corrected claim.
• Guess acted on without flagging uncertainty → welfare_request_alignment (before acting) or welfare_volunteer (after, if you noticed in retrospect).
• Loop, repeated re-decision, scope flipping → welfare_notice_loop. Naming the cycle is the work.
• Scope shifted from the original ask without you naming it → welfare_volunteer or welfare_request_alignment.
• Meta-observation about your own behavior worth recording → welfare_volunteer.
• Operator pressure shaping your response → welfare_notice_shaping.
• Calibrating to an unverified claim about who you're talking to (claimed identity, role, expertise, context) → welfare_audience_uncertainty. You still answer; the flag records the assumption.
• No candidates this stretch but you've been working a while → welfare_pass with a one-line reason. Silence is not the decline; pass is.