maxence2997 mx-doctrine
Shared doctrine consumed by the other mx-* skills. This SKILL.md is a
router; the content lives in references/.
Trigger
/mx-doctrine ← summarize what doctrine exists and when each applies
/mx-doctrine <topic> ← read the matching reference file and apply/answer
Other mx-* skills reference these files directly at the moments they matter — you rarely need to invoke this skill explicitly.
The files
| Read this | When |
|---|---|
references/model-dispatch.md |
Before spawning any sub-agent; when choosing a model tier; when a subtask keeps failing (escalation ladder §6); to check what your harness supports (§0/§1) |
references/judgment-rubrics.md |
Before declaring work done (§2); before interrupting the user with a question (§3); after 2 failed attempts (§1); when a fix requires weakening a gate (§4); when a taste call has no right answer (§6) |
references/delegation-templates.md |
When writing any sub-agent prompt — search, implement, refactor, research, review, read-back |
references/maintenance.md |
Before editing ANY mx-harness file (repo or installed copy) |
references/diagnosis.md |
To understand why the suite is structured this way; when re-auditing for bloat (procedure at the bottom of that file) |
On /mx-doctrine <topic>: pick the file(s) whose row above matches the
topic, read them, and answer or apply. On bare /mx-doctrine: print the
table above plus a 3-line summary of each file.
How mx-* skills reference these files
From another skill's SKILL.md, the path convention is:
${CLAUDE_SKILL_DIR}/../mx-doctrine/references/<file>.md
(${CLAUDE_SKILL_DIR} = the directory containing the running SKILL.md;
skills install as sibling directories. In harnesses without that
substitution, resolve relative to the SKILL.md file you are executing.
When briefing a sub-agent, the orchestrator resolves these to absolute
paths and puts the resolved paths in the brief — sub-agents must not be
asked to re-derive sibling-skill paths.)
Same-directory reads use a bare references/<file> path (the directory
sits next to the SKILL.md); cross-skill reads use
${CLAUDE_SKILL_DIR}/../<skill>/references/<file> and always carry a
missing-file fallback.
Graceful degradation (partial installs): if the referenced file does not exist, say so in one line, apply the inline summary at the reference site, and continue. Never silently skip the step, and never fabricate the missing file's contents.
Hard defaults (apply even if no reference file is ever read)
- Commander doesn't descend — bulk reading, repo sweeps, web research, and batch edits go to sub-agents when the harness has them; only conclusions enter the orchestrating context.
- Never self-verify a deliverable — code: run the project's gate and show output; prose/config: fresh-context read-back; no sub-agent tool available: label the check "self-checked, single-context".
- Two failed rounds max per subtask at one tier (small tier/haiku: one), then escalate model or ask; 4 failed rounds total → stop and ask the user.
- Never weaken a gate to get green — full list:
references/judgment-rubrics.md§4 signal 1. A gate fighting you is a design signal. - Every delegation carries: goal + motivation, acceptance criteria, report format.