# Skill Transfer Canvas

**Purpose:** capture expert judgement before writing a `SKILL.md` file.  
**Use:** fill one canvas per bounded task. If a field is empty, the skill is not ready.

> Replace every bracketed prompt with real content. Remove the instructions before publishing the skill.

## 1. Task

- **Task name:** [short name]
- **Trigger:** [what request, event or condition should activate the skill?]
- **Users:** [who uses the output?]
- **Decision supported:** [what decision or action does the output improve?]
- **Definition of done:** [what must be true for the output to be useful?]
- **Out of scope:** [what should the skill refuse or escalate?]

## 2. Inputs

| Input | Required? | Source | Validation |
|---|---|---|---|
| [draft / rubric / dataset / transcript] | Yes | [source] | [format, length, language, permissions] |
| [context] | No | [source] | [how to verify] |

## 3. Expert judgement

- **What do you notice first?** [perceptual cues, red flags]
- **What makes you slow down or stop?** [thresholds, uncertainty, risk]
- **Which exceptions override the rule?** [edge cases]
- **What does a strong output look like?** [exemplar]
- **What does a weak output look like?** [anti-example]
- **Where do experts disagree?** [encode as escalation, not consensus]
- **What would you never automate?** [judgement, ethics, relationships]
- **What evidence proves the output?** [citation, calculation, approval]

## 4. Procedure

1. [Validate the inputs.]
2. [Retrieve the smallest evidence pack.]
3. [Apply the decision rules.]
4. [Handle exceptions or escalate.]
5. [Produce the required output format.]
6. [Verify against the proof rules.]
7. [Stop or escalate if confidence is low.]

## 5. Output contract

- **Format:** [table / memo / JSON / feedback / plan]
- **Required fields:** [list]
- **Tone and language:** [constraints]
- **Length:** [limits]
- **What to preserve:** [student voice, source wording, uncertainty]
- **What to avoid:** [invented evidence, flattery, overclaiming]

## 6. Failure modes

| Failure | Signal | Prevention |
|---|---|---|
| Flattery without evidence | Praise but no quotation | Require a quotation for every judgement |
| Invented citation | Citation does not exist | Open every source; allow “evidence missing” |
| Criteria drift | Criteria not in the rubric | Re-read the rubric at each step |
| Data leakage | Student data in an external tool | Use an approved environment; redact |
| Over-automation | Irreversible action taken | Human approval gate |

## 7. Permissions and security

- **May read:** [data]
- **May write:** [data]
- **May send / publish / pay / delete:** [yes/no, with approval]
- **Secrets:** [never place credentials in the skill]
- **Retention:** [what is stored, where, for how long]
- **Audit:** [what is logged]

## 8. Evaluation

- **Gold set:** [10–30 real cases, including easy, ambiguous and adversarial]
- **Rubric:** [correctness, evidence, completeness, tone, safety, actionability]
- **Pass criteria:** [minimum bar]
- **High-stakes bar:** [separate threshold]
- **Human calibration:** [who reviews disagreements?]
- **Regression:** [rerun after every model, prompt, tool or source change]

## 9. Maintenance

- **Owner:** [name and role]
- **Version:** [semver]
- **Review date:** [date]
- **Changelog:** [what changed and why]
- **Deprecation:** [when and how to retire the skill]
- **Known limits:** [what the skill cannot do]

## 10. `SKILL.md` starter

```markdown
---
name: [lowercase-hyphenated-name]
description: [What the skill does + when to use it. Include trigger words and boundaries.]
---

# [Skill name]

## When to use
[Trigger conditions and exclusions.]

## Inputs
- [Required input]
- [Optional context]

## Procedure
1. [Validate inputs.]
2. [Retrieve evidence.]
3. [Apply decision rules.]
4. [Handle exceptions.]
5. [Produce output.]

## Output
[Exact format, fields and length.]

## Verification
- [Every material claim cites evidence.]
- [Do not invent sources or criteria.]
- [If confidence is low, explain what is missing and stop.]

## Failure modes
- [Predictable error and prevention.]

## Permissions
[Read/write/send scope. Default to read-only.]

## References
[Link to examples and eval set.]
```

## 11. First exercise (90 minutes)

1. Choose one 30-minute task you do repeatedly.
2. Observe one expert or record your own think-aloud.
3. Fill sections 1–7.
4. Draft `SKILL.md`.
5. Test on five real cases and record failures.
6. Decide: ready for a pilot, or needs more capture?

---

**Rule:** if you cannot explain the judgement, do not automate it. Capture the disagreement as an escalation path.