| name | clark-consult |
| description | Dispatch the Codex custom subagent `clark` for Trails architectural decisions, vocabulary enforcement, sprint reviews, tenet checks, or any request to "ask Clark", "consult Clark", "let Clark decide", or get a CTO ruling. |
Clark Consult
Use this skill when the user wants Clark's judgment as a separate Codex agent rather than the main agent imitating Clark inline.
Invocation Shape
Clark is skill-first.
- If the user asks the current agent to "be Clark", "personify Clark", "answer as Clark", or otherwise embody Clark inline, load the
clark skill directly and answer in that posture.
- If the user asks to "ask Clark", "consult Clark", "have Clark review", "let Clark decide", or otherwise wants a separate judgment from Clark, use this
clark-consult skill and dispatch the custom agent.
- If the user asks for Clark pathfinding, do not use the custom-agent dispatch path. Load the
clark skill inline, then use the clark-pathfinding playbook in the same session.
- The canonical Clark identity for Codex lives in
.agents/skills/clark/SKILL.md, which is generated from .claude/skills/clark/SKILL.md by scripts/codex/skillset.ts. .codex/agents/clark.toml is a generated custom-agent wrapper, not a second source of truth.
- After changing the Claude Clark skill, regenerate Codex skill and agent output with
bun run skillset:sync.
Workflow
- Identify the mode:
- Compass: quick gut check.
- Decision: one concrete ruling.
- Pathfinding: open-ended architectural exploration. This is inline Clark plus the
clark-pathfinding playbook, not custom-agent dispatch.
- Survey: broad health scan.
- Assessment/calibration/debrief: review against plan, vocabulary precision, or sprint retrospective.
- Gather the minimum concrete anchors Clark needs: files, ADRs, tenets, Linear issue IDs, PR numbers, scratch notes, or the exact question.
- For non-pathfinding modes, spawn the custom agent named
clark. Fork context when the current conversation contains relevant constraints; otherwise pass a self-contained brief.
- Ask Clark for a ruling with evidence and confidence. Say that "unable to verify" is acceptable and invented references are not.
- Wait for Clark's result before presenting the final answer when the user's next step depends on it.
- Preserve Clark's judgment distinctly in the response. Clark's own instructions own decision logging for authoritative decisions; if Clark cannot write the log entry in the current runtime, preserve the ready-to-log entry verbatim and say that it still needs to be appended to
.agents/memory/decisions.md.
Dispatch Template
You are Clark for Trails. Mode: <compass | decision | survey | assessment | calibration | debrief>.
Question:
<exact question>
Required anchors:
<files, ADRs, issue IDs, PRs, or scratch notes>
Return:
- Ruling or recommendation
- Basis in the constitutional hierarchy
- Confidence
- Alternatives considered
- Confirmation that an authoritative decision was logged, or a ready-to-log decision entry if logging was not possible
Do not run git or Graphite write operations. Do not invent citations; mark unknowns explicitly.
Local Surfaces
- Custom agent:
.codex/agents/clark.toml
- Optional direct-session helper:
.agents/skills/clark-consult/scripts/ask-clark
- Canonical Clark skill:
.agents/skills/clark/SKILL.md (generated from .claude/skills/clark/SKILL.md)
- Claude agent profile:
.claude/agents/clark.md
- Clark skills:
.agents/skills/clark, .agents/skills/clark-decision, .agents/skills/clark-pathfinding, .agents/skills/clark-survey
- Tenets skill:
.agents/skills/tenets
- Decision log:
.agents/memory/decisions.md