# Optional coaching for Fast Decisions

Documentation convention: `coaching-session/v1`. Keep the native library, CLI, MCP or host operation direct and usable unattended. Only when explicitly requested, offer one separate pre/post review checkpoint. Do not insert questions, a learning loop or extra latency into the operational call.

## One fictional practice task

This is an authored example, not a real user session, result or observed improvement.

| Agree before practice | Fictional example |
|---|---|
| User and job | Agent operator: Review two fictional matched execution receipts |
| Goal | Separate predicted saving from observed execution cost |
| Starting artifact and baseline | Preserve an unchanged copy/reference of `fictional-runs.json` before a proposed change. |
| Scope and budget | Only this fictional artifact; no external action. At most 1 practice round and 15 minutes unless the person agrees a smaller bound. Native technical retries remain separate. |
| What must survive | Keep native startup/candidate validation and no conversational latency in calls |
| Feedback checkpoint | Are these runs matched on task and checked outcome? |
| Human review | Present keep, revert or leave unchanged with reasons; keep/revert remains a proposal until separately authorized. |

Ask the person to attempt the relevant judgment before giving diagnostic feedback, when appropriate. Keep the method's own question cadence and gates. Accept “just finish” as an exit from teaching; continue only the already authorized task. If the person is unavailable, return assumptions, a proposal and unanswered questions. Do not simulate their answers or claim they grew.

## Stop and keep control

Stop at the agreed goal, user exit at any time, round/time cap, a full stalled attempt, no meaningful gain, missing input/authority or a critical failed/unknown gate. Leave a sufficient artifact alone. Preserve the baseline and relevant failed/unknown evidence; review keep/revert before adoption. A coaching note grants no tool, spending, editing, sharing or publication authority. Native verifiers still check the facts they actually exercise.

## Optional private note

Record only the named user/job, goal, starting artifact, baseline reference, scope/budget, what must survive, feedback checkpoint, stop reason and reviewed keep/revert choice. Keep actual human attempts/reasons separate from the artifact; observed dips need not be progress. Save nowhere by default. Save a minimal task-only note only after explicit opt-in to a chosen private location and retention period; do not automatically add transcripts, personal details, standing memory or credentials.

Usefulness remains a separate evidence question: Job, First value, Fit, Outcome, Trust and Return are not one total. Team/agent walkthroughs stay Claimed. Seen needs a specific dated outside-user observation; Repeated needs unprompted return. No authentic human learning or outside-user usefulness is established by this fictional example or by software checks.

This local guide is sufficient on its own. No other capability, remote harness, schema download or persistent service is required. Existing method IDs, native interfaces, manifests and older record schemas are unchanged.