Mindful conduction
Skill finnfotschki/mindful-agent-collaboration/skills/mindful-conduction
Installable skills, gates, hooks and templates for deliberate human-agent work — drops into your existing Claude Code or Codex setup.
npx -y skills add finnfotschki/mindful-agent-collaboration --skill mindful-conductionAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Structure recurring human-agent work to preserve meaningful human task exposure through deliberate execution and monitoring, active domain-knowledge maintenance, and critical output checking. Use when a human is drifting into confirm-only review, when the same domain judgment or verification task has been delegated repeatedly, at consequential assistive or human-final judgment nodes, or when a reviewer must remain capable of independently assessing agent output. This is an unvalidated design routine, not a guarantee of competence preservation.
SKILL.md
4.4 KB, 754 tokens by cl100k_base, as published. Nobody here has run it
Mindful Conduction
Overview
Plan recurring opportunities for the human to perform, monitor, and critique domain work instead of relying on passive approval. Treat the routine as a design hypothesis whose appropriate cadence and effect must be evaluated in the actual context.
When to use
- When repeated delegation has reduced the human role to accepting conclusions.
- At consequential
assistiveorhuman-finaljudgment nodes. - When independent verification skill is important to the control design.
When NOT to use: every routine autonomous step; trivial reversible choices; or as evidence that competence has actually been maintained. Use scheduled samples rather than adding friction everywhere.
Process
- Name the capability at stake. Specify the domain judgment, execution, or verification skill the human needs to exercise and why it matters.
- Set a human-owned cadence. Schedule an output-independent sample or rotation point. Record the date, owner, activity, and next due date; do not claim a universal frequency.
- Rotate deliberate execution. On selected instances, let the human perform the relevant task or critical substep without an agent answer in view. The agent may review afterward.
- Maintain domain criteria. Ask the human to explain or update the standards used to distinguish acceptable from unacceptable output. Propose persisted checklist or rule changes for human approval; never write them silently.
- Require critical output checking. For verifiable work, have the human apply the criterion to a sample and explain any discrepancy. At a consequential judgment node, record the human's hypothesis or weighting before showing an agent interpretation.
- Respect delegation modes. In
assistive, present neutral evidence and symmetric options first; offer a non-binding assessment only after the human has formed a view. Inhuman-final, limit the agent to evidence gathering and consequence simulation while the human frames, weights, decides, and approves. - Record the decision in
assets/decision-record.mdwhen the exercise produces a consequential choice. Capture rationale and a revisit condition. - Evaluate, do not assert. Track whether the routine was completed and what corrections or criteria emerged. Describe any claimed effect on competence, control quality, or calibration as an unvalidated design hypothesis unless a suitable evaluation supports it.
Rationalizations
| Excuse | Rebuttal |
|---|---|
| "The human can override my recommendation." | Seeing a recommendation first can anchor the judgment; capture the human view first when consequence warrants it. |
| "Every step needs practice." | Forced participation in exact routine work can consume the delegation benefit; sample deliberately. |
| "Completing the routine proves competence was preserved." | Completion is process evidence, not an effectiveness result. |
| "The agent can update the checklist from corrections." | Persisted criteria steer future work and require human approval. |
Red flags
- The human only approves agent-selected options.
- No independent task sample or criterion explanation occurs over the agreed cadence.
- A
human-finalrecord contains an agent-made decision. - Routine completion is reported as proof of competence preservation.
- Human corrections disappear without a governed proposal to update criteria.
Verification
The routine is complete when the capability and cadence are recorded; the human performed the scheduled independent activity; criteria and corrections were captured; consequential judgments preserved the proper mode boundary; persisted changes received human approval; and claims about effects are separated from evidence about whether the routine occurred.
What ships with it: 3 files
3.3 KB alongside SKILL.md
agents/
- openai.yaml229 B
assets/
- decision-record.md768 B
evals/
- evals.json2.3 KB