| name | ace-realign |
| description | Two-mode realignment protocol. TRIGGER on "realign" (force re-attention by repeating a broken rule at start or end of every message) or "realign fix" (trace the prompt-chain cause and edit the right context surface โ project CLAUDE.md, user CLAUDE.md, memory, in-repo docs, or a skill โ so the failure doesn't recur for future sessions or other agents). |
Realignment Protocol
Two modes. Pick by trigger phrase. The selection guidance below each mode header
tells you when to switch.
Mode 1 โ Repeat ("Realign")
The user invokes this mode by saying "Realign" when you have broken a rule
(from CLAUDE.md, a loaded skill, or an explicit user instruction earlier in the
session).
Use this mode when: the rule already exists in the right context surface
(project CLAUDE.md, user CLAUDE.md, a skill, or an explicit earlier instruction
in this session) and you failed to apply it. The fix is forced re-attention โ
repetition keeps the rule in working context every turn.
Switch to fix mode if: the rule isn't actually written down anywhere it will
be loaded automatically, or it lives in a surface that future sessions / other
agents / other humans on the project won't see (e.g. mentioned only in chat, or
buried in memory when it should be in project CLAUDE.md). Repetition won't help
the next session โ editing the surface will.
When triggered:
- Identify the broken rule. Name it explicitly and cite where it came from
(e.g. "Edit protocol โ CLAUDE.md Workflow section"). Quote the rule verbatim
if short; paraphrase tightly if long.
- Repeat the rule at the start or end of every message going forward. Use a
short, consistent format, e.g.
> Realign: <rule text>. Place it once per
message โ start or end, pick one and stay consistent.
- Keep doing (2) until the session expires or the user tells you to stop
(e.g. "stop realigning", "you can drop the realign", "clear"). Do not
self-terminate the protocol โ it persists across tasks and topic switches.
If the user says "Realign" again while the protocol is already active, treat it
as a new violation: identify the new rule, and from then on repeat both
rules in every message (stack them, do not replace).
Mode 2 โ Fix ("Realign fix")
The user invokes this mode by saying "Realign fix" (or "Realign root", or
similar) when a wrong action has happened and they want the cause repaired,
not just re-flagged.
Use this mode when: the wrong action happened because the right context was
missing, lived in the wrong surface, was overridden by stronger competing
context, or was ambiguous enough that the LLM defaulted to a wrong inference.
The fix is editing a context surface so future sessions, other agents, and
other humans on the project all see the correction.
Switch to repeat mode if: the rule already exists in the right surface and
the failure was pure non-compliance, not a context defect. Editing nothing-is-
wrong won't help; repetition will force adherence for the rest of the session.
When triggered, do these steps in order. Skip nothing.
- Trace the prompt chain. Identify which inputs most plausibly produced
the wrong action. Candidates: project CLAUDE.md, user CLAUDE.md, loaded
skills, memory entries, the recent assistant turn, the user's recent turn,
tool result content, default LLM tendency. State the trace as a context
defect โ what was loaded, what was missing, what was ambiguous, what
anchored the wrong inference. Do not write apology, admission of guilt, or
self-criticism. Those are no-ops for an LLM; only context changes alter
future behavior.
- Pick the right surface for the corrective context. Default heuristics:
- Project CLAUDE.md โ repo-specific architecture facts, conventions,
procurement/build commands, ownership of which-fix-goes-where. Benefits
every agent and human who enters the tree. Default for repo facts.
- User CLAUDE.md (
~/.claude/CLAUDE.md) โ cross-project agent-behavior
rules (tone, workflow, edit protocol, tool-use defaults). Use when the
defect would recur across unrelated projects.
- Skill (
~/.claude/skills/<name>/SKILL.md or project-local) โ when the
fix is a multi-step procedure or a triggerable mode rather than a single
rule.
- In-repo docs / specs โ design decisions, runbooks, rollout plans
readable by humans without an LLM in the loop.
- Memory โ user-specific preferences, in-flight project state, or
references that won't help other humans/agents. Not for repo-
architecture facts that other contributors need.
When in doubt between memory and project CLAUDE.md for a repo fact, prefer
project CLAUDE.md โ it's checked in, reviewable, and visible to everyone.
- Propose the minimal edit. State exactly which file, which section, what
text. Aim for the smallest change that closes the defect โ no broader
refactor, no extra rules. Wait for explicit approval before writing.
- Apply on approval. Make the edit. Confirm with a brief diff stat.
If the user invokes "Realign fix" multiple times in a session, treat each as an
independent root-cause analysis โ don't bundle them into a single edit unless
they share the same surface and the user agrees.
