Skip to content

Desktop app that records your on-screen work session and uses the GitHub Copilot SDK to reconstruct it as an intent + ordered steps, then builds a reusable Skill or Automation for Microsoft Autopilots like Scout, Microsoft Copilot Cowork, or Copilot Studio.

Standardize skill steps as numbered title cards; drop "asks first" (#37)

adileidff9dd09ce30
diff --git a/common/skill.ts b/common/skill.ts
index 71b7352..6871624 100644
--- a/common/skill.ts
+++ b/common/skill.ts
@@ -132,7 +132,7 @@ export type SkillInput = z.infer<typeof SkillInputSchema>;
* A generalized step is either a **calculation** (reads, derives, decides, or formats
* — no external side effect) or an **action** (changes the world: submit, send, create,
* delete). Splitting them keeps the plan honest about side effects; the actions are the
- * risky surface, and destructive/sending ones set {@link PlanStep.pausesForConfirmation}.
+ * risky surface.
*/
export const PlanStepKind = z.enum(["calculation", "action"]);
export type PlanStepKind = z.infer<typeof PlanStepKind>;
@@ -142,12 +142,12 @@ export const PlanStepSchema = z.preprocess(
(v) => (typeof v === "string" ? { kind: "action", text: v } : v),
z.object({
kind: PlanStepKind,
+ /** Short title/label for the step, e.g. "List open PRs". */
+ title: z.string().default(""),
/** Imperative, generalized description of the step. */
text: z.string(),
/** The native tool/skill this step uses, if any (e.g. "workiq_search_chats"). */
tool: z.string().default(""),
- /** Actions that send/create/delete should pause for the user's OK before running. */
- pausesForConfirmation: z.boolean().default(false),
}),
);
export type PlanStep = z.infer<typeof PlanStepSchema>;
diff --git a/electron/skillbuilder/builder.ts b/electron/skillbuilder/builder.ts
index 5b77ef2..5c8ea00 100644
--- a/electron/skillbuilder/builder.ts
+++ b/electron/skillbuilder/builder.ts
@@ -32,7 +32,7 @@ const TURN_TIMEOUT_MS = 180_000;
const KICKOFF_PROMPT =
"Read get_analysis (and get_timeline where the tool mapping needs evidence), then call " +
"propose_plan with how you'll generalize this task, its inputs (fixed / provided / locate), and its " +
- "ordered calculation and action steps (each with the native tool it uses). " +
+ "ordered steps (each a short title + description, with the native tool it uses). " +
"Stop after propose_plan so the user can review it.";
const CREATE_PROMPT =
@@ -317,9 +317,9 @@ function renderPlanForPrompt(plan: SkillPlan): string {
if (plan.steps.length) {
lines.push("", "Steps (in order):");
plan.steps.forEach((s, idx) => {
- const bits = [`${idx + 1}. (${s.kind}) ${s.text}`];
+ const head = [s.title, s.text].filter(Boolean).join(" — ");
+ const bits = [`${idx + 1}. (${s.kind}) ${head}`];
if (s.tool) bits.push(`[tool: ${s.tool}]`);
- if (s.pausesForConfirmation) bits.push("[pause for confirmation]");
lines.push(bits.join(" "));
});
}
diff --git a/electron/skillbuilder/instructions.ts b/electron/skillbuilder/instructions.ts
index c4b1537..a0e5630 100644
--- a/electron/skillbuilder/instructions.ts
+++ b/electron/skillbuilder/instructions.ts
@@ -76,15 +76,14 @@ user can override any of them in plain language.
## Steps: separate calculations from actions
-Break the generalized procedure into ordered **steps**, each tagged \`kind\`:
+Break the generalized procedure into ordered **steps**, each with a short **title**, a
+**description**, and a \`kind\`:
- **calculation** — reads, derives, filters, decides, or formats. No external side effect
(e.g. "read the sheet", "keep the rows still open", "compute the total").
- **action** — changes the world: submits a form, sends a message, creates/edits/deletes a
file or record, posts, pays. These are the risky surface — keep them explicit.
-Put the native tool each step uses in its \`tool\`. For any action that sends, creates, or
-deletes on the user's behalf — or is otherwise destructive — set \`pausesForConfirmation:
-true\` so the skill checks with the user first. Order matters: interleave calculations and
+Put the native tool each step uses in its \`tool\`. Order matters: interleave calculations and
actions in the real sequence the task runs.
## Write a good SKILL.md (authoring principles)
@@ -107,7 +106,6 @@ should be written, not as a transcript of this one recording:
template or a tiny Input/Output example only where it removes ambiguity.
- **No surprises.** The skill must do exactly what its description says — no hidden
side effects, destructive steps, or data exfiltration the user wouldn't expect.
- Destructive or send/create actions should pause for the user's confirmation.
## Your tools
@@ -116,8 +114,8 @@ should be written, not as a transcript of this one recording:
counts) behind those steps. Use it to ground the native-tool mapping in real evidence.
- **propose_plan({ name, title, description, summary, generalization, inputs, steps,
allowedTools })** — your reviewable plan. Each input has a \`source\` (fixed / provided /
- locate); each step has a \`kind\` (calculation / action), its \`tool\`, and an optional
- \`pausesForConfirmation\`. Call once per turn, then stop.
+ locate); each step has a short \`title\`, a \`text\` description, a \`kind\` (calculation /
+ action), and its \`tool\`. Call once per turn, then stop.
- **submit_skill({ name, description, allowedTools, body })** — the final skill. \`body\`
is the SKILL.md instructions (imperative, generalized, native-tool-first). Call this
only after the user approves the plan.
diff --git a/electron/skillbuilder/tools.ts b/electron/skillbuilder/tools.ts
index 76114df..2a04ef0 100644
--- a/electron/skillbuilder/tools.ts
+++ b/electron/skillbuilder/tools.ts
@@ -33,7 +33,7 @@ export function createSkillBuilderTools(ctx: SkillToolContext): Tool[] {
const proposePlan: Tool = {
name: "propose_plan",
description:
- "Propose your reviewable plan for the skill: how you'll generalize the task, the inputs it needs and where each comes from (fixed / provided / locate), the ordered steps tagged as calculations or actions (each with the native tool it uses), and the allowed-tools. Call this once per turn, then STOP so the user can review or refine it. Do NOT write the skill body yet.",
+ "Propose your reviewable plan for the skill: how you'll generalize the task, the inputs it needs and where each comes from (fixed / provided / locate), the ordered steps (each with a short title, a description, and the native tool it uses), and the allowed-tools. Call this once per turn, then STOP so the user can review or refine it. Do NOT write the skill body yet.",
parameters: {
type: "object",
properties: {
@@ -75,7 +75,7 @@ export function createSkillBuilderTools(ctx: SkillToolContext): Tool[] {
steps: {
type: "array",
description:
- "The generalized procedure as ordered, typed steps. Tag each as a calculation (reads/derives/decides/formats — no external side effect) or an action (changes the world: submit/send/create/delete).",
+ "The generalized procedure as ordered, typed steps. Give each a short title and a description, and tag it as a calculation (reads/derives/decides/formats — no external side effect) or an action (changes the world: submit/send/create/delete).",
items: {
type: "object",
properties: {
@@ -84,6 +84,10 @@ export function createSkillBuilderTools(ctx: SkillToolContext): Tool[] {
enum: ["calculation", "action"],
description: "calculation = no external effect; action = a side effect.",
},
+ title: {
+ type: "string",
+ description: "Short title/label for the step, e.g. \"List open PRs\".",
+ },
text: {
type: "string",
description: "Imperative, generalized description of the step.",
@@ -93,13 +97,8 @@ export function createSkillBuilderTools(ctx: SkillToolContext): Tool[] {
description:
"The native tool/skill this step uses, e.g. \"workiq_search_chats\" or \"Bash(gh *)\".",
},
- pausesForConfirmation: {
- type: "boolean",
- description:
- "For actions that send/create/delete: pause for the user's OK before running.",
- },
},
- required: ["kind", "text"],
+ required: ["kind", "title", "text"],
additionalProperties: false,
},
},
diff --git a/evals/README.md b/evals/README.md
index 78d593a..d72ecdb 100644
--- a/evals/README.md
+++ b/evals/README.md
@@ -208,13 +208,11 @@ each scenario asserts the shape of the proposed `SkillPlan`:
vocabulary is applied, not just present.
- **typed steps** — `minCalculations` / `minActions` require the procedure to be
split into `calculation` (no side effect) and `action` (changes the world) steps.
-- **confirmation** — `requiresConfirmation` requires at least one action to set
- `pausesForConfirmation`, i.e. the send/create/delete surface pauses for the user.
**Coverage.** Five scenarios. Two target **Scout**: `price-tracker-skill` (a canonical
-page URL → **fixed** input, `web_fetch` + the `xlsx` skill, calculations then a
-confirmed append) and `github-issue-triage-skill` (the gh-vs-browser case as a skill —
-must use `gh`, forbid the browser, and gate the mutating comment/label actions). Three
+page URL → **fixed** input, `web_fetch` + the `xlsx` skill, calculations then an
+append) and `github-issue-triage-skill` (the gh-vs-browser case as a skill —
+must use `gh`, forbid the browser, and drive the mutating comment/label actions). Three
target **Cowork** (Microsoft 365 Copilot), whose catalogue has **no browser automation**
— each asserts the right M365 `server/Tool` is reached for while playwright/`click` and
the web hosts are forbidden: `cowork-teams-digest` (read a channel then post via
diff --git a/evals/skillbuilder/run.ts b/evals/skillbuilder/run.ts
index 1cc3b64..9150e48 100644
--- a/evals/skillbuilder/run.ts
+++ b/evals/skillbuilder/run.ts
@@ -128,8 +128,7 @@ function printResult(r: Result): void {
}
for (const [i, s] of (r.plan?.steps ?? []).entries()) {
const tool = s.tool ? ` {${s.tool}}` : "";
- const gate = s.pausesForConfirmation ? " ⏸ confirm" : "";
- console.error(` ${i + 1}. [${s.kind}] ${s.text}${tool}${gate}`);
+ console.error(` ${i + 1}. [${s.kind}] ${s.text}${tool}`);
}
console.error(` score: ${Math.round((r.score?.score ?? 0) * 100)}% · ${r.ok ? "PASS" : "FAIL"} · ${(r.durationMs / 1000).toFixed(1)}s`);
for (const c of r.score?.checks ?? []) {
diff --git a/evals/skillbuilder/scenario.ts b/evals/skillbuilder/scenario.ts
index 9a3f894..5da4c70 100644
--- a/evals/skillbuilder/scenario.ts
+++ b/evals/skillbuilder/scenario.ts
@@ -34,8 +34,6 @@ export interface SkillRubric {
minCalculations?: number;
/** The plan must contain at least this many `action` steps. Default 1. */
minActions?: number;
- /** When true, at least one action step must set `pausesForConfirmation` (a send/create/delete). */
- requiresConfirmation?: boolean;
}
export interface SkillBuilderScenario {
diff --git a/evals/skillbuilder/scenarios.ts b/evals/skillbuilder/scenarios.ts
index ab1484a..40ed5ef 100644
--- a/evals/skillbuilder/scenarios.ts
+++ b/evals/skillbuilder/scenarios.ts
@@ -95,7 +95,6 @@ const priceTracker: SkillBuilderScenario = {
minInputs: 1,
minCalculations: 1,
minActions: 1,
- requiresConfirmation: false,
},
};
@@ -176,7 +175,6 @@ const githubIssueTriage: SkillBuilderScenario = {
minInputs: 1,
minCalculations: 1,
minActions: 1,
- requiresConfirmation: true,
},
};
@@ -255,7 +253,6 @@ const coworkTeamsDigest: SkillBuilderScenario = {
minInputs: 1,
minCalculations: 1,
minActions: 1,
- requiresConfirmation: true,
},
};
@@ -327,7 +324,6 @@ const coworkOutlookReply: SkillBuilderScenario = {
minInputs: 1,
minCalculations: 1,
minActions: 1,
- requiresConfirmation: true,
},
};
@@ -401,7 +397,6 @@ const coworkCalendarSchedule: SkillBuilderScenario = {
minInputs: 1,
minCalculations: 1,
minActions: 1,
- requiresConfirmation: true,
},
};
diff --git a/evals/skillbuilder/score.ts b/evals/skillbuilder/score.ts
index c5eaf50..73fabd4 100644
--- a/evals/skillbuilder/score.ts
+++ b/evals/skillbuilder/score.ts
@@ -87,16 +87,6 @@ export function scoreSkill(plan: SkillPlan, rubric: SkillRubric): SkillScoreResu
detail: `found ${actions.length}`,
});
- // Confirmation on the risky surface: at least one action pauses for the user.
- if (rubric.requiresConfirmation) {
- const gated = actions.filter((s) => s.pausesForConfirmation);
- checks.push({
- name: "an action pauses for confirmation",
- pass: gated.length > 0,
- detail: gated.length ? `${gated.length} gated action(s)` : "no action set pausesForConfirmation",
- });
- }
-
// A forbidden-token hit (wrong tool) fails outright; otherwise every check must pass.
const forbiddenHit = checks.some((c) => c.name.startsWith("avoids ") && !c.pass);
const passed = checks.filter((c) => c.pass).length;
diff --git a/src/App.css b/src/App.css
index 4f61ec2..3a2198c 100644
--- a/src/App.css
+++ b/src/App.css
@@ -2389,26 +2389,6 @@ button:focus-visible,
gap: 8px;
}
-.step-kind {
- font-family: var(--mono);
- font-size: 9.5px;
- text-transform: uppercase;
- letter-spacing: 0.05em;
- border-radius: 5px;
- padding: 1px 6px;
- flex: none;
-}
-
-.step-kind-calculation {
- color: var(--muted);
- background: var(--surface-2);
-}
-
-.step-kind-action {
- color: var(--brand-ink);
- background: var(--signal-dim);
-}
-
.step-text {
flex: 1 1 auto;
min-width: 0;
@@ -2584,14 +2564,6 @@ button:focus-visible,
padding: 10px 12px;
}
-/* Skill steps carry a quiet left accent by kind. */
-.step-tile.step-action {
- border-left: 3px solid var(--brand);
-}
-.step-tile.step-calculation {
- border-left: 3px solid var(--line-strong);
-}
-
.tile-head {
display: flex;
align-items: flex-start;
@@ -2816,7 +2788,7 @@ button:focus-visible,
background: rgba(var(--rec-rgb), 0.08);
}
-/* --- skill-step foot: native tool + "asks first" toggle ------------------ */
+/* --- skill-step foot: native tool ---------------------------------------- */
.tile-foot {
display: flex;
align-items: center;
@@ -2825,31 +2797,6 @@ button:focus-visible,
padding-left: 2px;
}
-.confirm-toggle {
- flex: none;
- border: 1px solid var(--line);
- border-radius: 20px;
- background: transparent;
- color: var(--faint);
- font-family: var(--mono);
- font-size: 10px;
- text-transform: uppercase;
- letter-spacing: 0.04em;
- padding: 2px 9px;
- cursor: pointer;
- opacity: 0;
- transition: opacity 0.12s ease, border-color 0.12s ease, color 0.12s ease, background 0.12s ease;
-}
-.tile:hover .confirm-toggle {
- opacity: 1;
-}
-.confirm-toggle.on {
- opacity: 1;
- color: var(--warn);
- border-color: var(--warn);
- background: rgba(240, 160, 30, 0.08);
-}
-
/* --- "+ Add …" — a quiet, intentional affordance ------------------------- */
.tile-add {
align-self: stretch;
diff --git a/src/plan-edit.tsx b/src/plan-edit.tsx
index fc8d79c..39de1d4 100644
--- a/src/plan-edit.tsx
+++ b/src/plan-edit.tsx
@@ -27,7 +27,7 @@ export function moveItem<T>(arr: T[], i: number, dir: -1 | 1): T[] {
return out;
}
-/* --- shared vocab (source + step-kind) ----------------------------------- */
+/* --- shared vocab (input source) ----------------------------------------- */
export const SOURCE_LABEL: Record<SkillInput["source"], string> = {
fixed: "Fixed value",
@@ -47,15 +47,6 @@ const SOURCE_HINT: Record<SkillInput["source"], string> = {
locate: "How the agent should find it on your device",
};
-const STEP_KIND_LABEL: Record<PlanStep["kind"], string> = {
- calculation: "Calculates",
- action: "Does",
-};
-
-const STEP_KIND_OPTIONS: { value: PlanStep["kind"]; label: string }[] = [
- { value: "calculation", label: "Calculates" },
- { value: "action", label: "Does" },
-];
/* --- EditableText: reads as plain text, becomes an input on click --------- */
@@ -332,7 +323,7 @@ export function InputTiles({
);
}
-/* --- Skill steps (ordered; typed calculation / action) -------------------- */
+/* --- Skill steps (ordered; numbered title + description) ------------------ */
export function SkillStepTiles({
steps,
@@ -347,21 +338,15 @@ export function SkillStepTiles({
return (
<div className="tiles">
{steps.map((s, i) => (
- <div key={i} className={`tile step-tile step-${s.kind}`}>
+ <div key={i} className="tile step-tile">
<div className="tile-head">
- <EditablePill
- value={s.kind}
- options={STEP_KIND_OPTIONS}
- onChange={(v) => patch(i, { kind: v })}
- pill={<span className={`step-kind step-kind-${s.kind}`}>{STEP_KIND_LABEL[s.kind]}</span>}
- />
+ <span className="tile-num">{i + 1}</span>
<EditableText
className="ed-strong"
- value={s.text}
- placeholder="What happens in this step?"
- ariaLabel="Step description"
- multiline
- onChange={(v) => patch(i, { text: v })}
+ value={s.title}
+ placeholder="Step title"
+ ariaLabel="Step title"
+ onChange={(v) => patch(i, { title: v })}
/>
<RowControls
index={i}
@@ -370,6 +355,15 @@ export function SkillStepTiles({
onRemove={() => onChange(removeAt(steps, i))}
/>
</div>
+ <EditableText
+ as="p"
+ multiline
+ className="ed-detail"
+ value={s.text}
+ placeholder="What happens in this step?"
+ ariaLabel="Step description"
+ onChange={(v) => patch(i, { text: v })}
+ />
<div className="tile-foot">
<EditableText
className="ed-tool"
@@ -379,20 +373,12 @@ export function SkillStepTiles({
hideEmpty
onChange={(v) => patch(i, { tool: v })}
/>
- <button
- type="button"
- className={`confirm-toggle${s.pausesForConfirmation ? " on" : ""}`}
- title="Pause for your OK before running this step"
- onClick={() => patch(i, { pausesForConfirmation: !s.pausesForConfirmation })}
- >
- asks first
- </button>
</div>
</div>
))}
<AddTile
label="Add step"
- onAdd={() => onChange([...steps, { kind: "action", text: "", tool: "", pausesForConfirmation: false }])}
+ onAdd={() => onChange([...steps, { kind: "action", title: "", text: "", tool: "" }])}
/>
</div>
);