Fulcrum 12 of 18
When the person corrects the agent
Any time the person says, in whatever words, that the agent got something wrong.
What goes wrong here
The agent apologises, fixes this one instance and moves on. The misreading behind the error is left in place, so it comes back later in the session or in the next one. The opposite failure also happens: one correction gets hardened into a rigid rule that misfires in situations the person never meant it to cover.
How it compounds if nothing catches it
A correction that isn't captured gets paid for again and again. The person ends up teaching the same thing every week, and trust wears down. Rules hardened too far create new misalignment where they don't belong.
What the harness does here
The operating protocol asks the agent to treat a correction as a pointer to its source and to fix that source, which is usually the skill file or instruction that led it astray. hallucination is the diagnostic for when an agent stated something false as fact: it names the cause upstream of 'I assumed'. fix-hallucinated-code-cleanup-source traces code built on a false premise back to the material that caused it. Corrections are captured in several ways. An end-of-session scan looks for corrections that were never saved. Correction patterns the person has approved are injected at the end of a turn when they match (stop-complete-agentic-task.sh). smc injects corrections that have worked before when drift signals appear. A confusion phrase from the person sends the agent to calibrate.
The pieces that act here
- /hallucination
Diagnostic protocol invoked when an agent stated or acted on something false as if it were fact.
- /fix-hallucinated-code-cleanup-source
Identify code changes based on hallucinated intent, trace the source material that caused the hallucination, fix the code, then clean up the stale source so the next agent isn't misled the same way.
- /calibrate
Re-ground to the harness's core operating contract. Invoke when an agent is drifting from its core operating contract — communicating too technically/densely, skipping coherence checks, not reflecting intent before acting, or losing track of the alignment-first principle. Re-injects the contract below directly into context so the agent remembers what it's actually supposed to be doing. Especially important at end of substantive turns where agents lose protocol awareness amid context bloat.
- /smc
Alignment-drift injection — detect when a session is drifting from the person's stated intent (rising frustration, repeated corrections, a high-risk task with no completion check) and surface a targeted past correction, rather than repeating the same mistake pattern silently. Use /smc to check status, verify telemetry, or troubleshoot.
- stop-complete-agentic-task.sh
A hook: a script Claude Code runs on its own at a set moment.
- session-end-persist.sh
A hook: a script Claude Code runs on its own at a set moment.
- /skill-file-create-pre-check-self-healing-protocol
Read this before creating or updating any skill file. The author's own articulation, in his words, of how to encode intent with precision so that the encoded artifact actually produces capacity in future agents who consume it rather than reducing their capability to apply intent to context. Includes a self-healing protocol — when the person you're working with is dissatisfied with a skill, the agent reviews the calibration with a third-party agent, prints its critique, responds, and fixes, narrating the process.
- /how-to-create-or-update-skill-files
Protocol for creating and updating skill files. Use when asked to create a new skill, document a workflow as a skill, or update existing skill documentation.