/ai-apply

The front-door decision skill: given a task or workflow, walk Gera through the AI Applicability Framework one question at a time and land an honest adopt / augment / skip verdict. It answers "to what extent can AI be leveraged here?" before any build starts.

This is the judgment half of deciding what to build. It sits one step upstream of /build, /skill-builder, /level-up, and the rest of the builder family — those build the thing; this decides whether the thing is worth building, and in what shape.

The question sequence, the "what a good answer looks like" notes, and the scoring rubric are the AI Applicability Framework — read references/ai-applicability-framework.md (the generic version + scoring section) before running. That page is the source of truth; this skill is the interview that walks it and files the result.

When to use

• Gera invokes /ai-apply or says "should I use AI for this", "does AI fit here", "is this worth automating", "where does AI help in X", "assess this for AI", "AI applicability check".
• He's about to build/automate something and you (or he) aren't sure AI is the right call, or aren't sure how deep to go.
• Distinct from neighbors: /grill-me extracts knowledge Gera already holds; the builders' Discovery Interviews are tied to building ONE artifact; this skill DECIDES whether to build at all. If he's already decided to build and just wants it built, skip straight to /build.

If he didn't ask and there's no live decision on the table, don't manufacture work.

The boundary (read this before the data question)

This skill is for personal and public-data tasks only. The moment a task touches SoftServe Internal+ material or customer IP/code, the assessment itself must not route any of that content through the multi-brain (Codex / Gemini / local brains) or through Servy. Question 5 is the explicit gate; if it blocks, the verdict is SKIP-or-RE-ROUTE and you say so plainly. See references/softserve-ai-usage-policy.md.

Setup — do this BEFORE the first question

1. Pick the task + slug. From $ARGUMENTS if given, else ask one line: "What task or workflow are we assessing?" Make a kebab slug.
2. Get today's date: date +%F.
3. Read the framework: references/ai-applicability-framework.md — the generic 6-question version and the adopt/augment/skip rubric. Also skim the ## Knowledge base + ## Multi-LLM orchestration sections of CLAUDE.md and context/priorities.md so your recommended answers reflect his real stack and current priorities.
4. Create the assessment file at knowledge/ai-applicability/{YYYY-MM-DD}-{task-slug}.md (create the knowledge/ai-applicability/ folder if it doesn't exist). Seed it with the template below: title, date, the task in one line, empty per-question log, empty verdict.
5. Tell Gera where you're saving, in one line. Then ask Q1.

The checkpoint rule (non-negotiable)

After EVERY answer, BEFORE the next question:

• Append the question's captured answer + the decision it drove to the assessment file (use Edit). Keep his wording where it matters.
• Update the running notes if a later answer changes an earlier one.
• Only then ask the next question.

Never batch. The file, not your context window, is the source of truth — if the session drops, the file already holds everything decided so far. (Same discipline as /grill-me.)

Interview method — one question at a time, with a recommended answer

Walk the six questions in order (each upstream answer shapes the ones below it). For each question:

1. Ask it in plain words.
2. Offer your recommended answer — your best inference from his context (CLAUDE.md, the wiki, the repo, the task description) — so Gera can confirm, correct, or redirect instead of starting from blank. This is the grill-me move: never make him fill a void you could pre-fill.
3. Capture his actual answer + the decision it drives, checkpoint, then move on.

Explore instead of asking where you can — if a step is answerable from a file or the repo, read it rather than asking. Flag and move on if only a stakeholder knows (e.g. a manager, a client). Match references/voice.md — warm, plain, low jargon, first person, no em dashes. This is a conversation, not an interrogation.

The six questions (full notes in the framework page)

1. What is the task, and could it just stop existing? — drives ELIMINATE vs CONTINUE. (If eliminate, you may be done — verdict SKIP.)
2. Is this a real constraint, or just annoying? — drives PRIORITIZE vs DEFER.
3. Can you explain the steps clearly enough to hand to a person? — capture the five elements (Trigger, Data Sources, Transformations, Decision Points, Destination); drives READY-TO-SCOPE vs MAP-FIRST.
4. What share can AI realistically do, and where does a human stay? — capture the 60/30/10-style split + the starting autonomy level (L0–L4); drives the AUGMENT-vs-AUTOMATE shape.
5. Does it touch confidential/customer data — and is the channel allowed? — the hard gate; drives PROCEED vs RE-ROUTE vs BLOCK.
6. What number does this move, and how will you roll it out safely? — capture one KPI + a phased rollout; drives GREEN-LIGHT vs PARK.

Early-exit allowed: if Q1 lands on eliminate, or Q5 hard-blocks with no re-route, you can short-circuit to the verdict — still checkpoint the answers you have and record why you stopped.

Score the verdict

After the six answers, apply the rubric from the framework page (a guide, not arithmetic — one hard NO can sink a strong case):

• SKIP if any of: Q1 says eliminate, Q5 blocks with no legal re-route, or Q6 finds no number it moves. SKIP is a real, valuable outcome — name it without apology.
• AUGMENT (AI-assisted, human in the loop — the 30% lane) when the process is legible (Q3), Q4 keeps a meaningful human-review share, and you'd start LOW (L1–L2). The safe default for most worthwhile tasks.
• ADOPT / AUTOMATE (toward hands-off — the 60% lane) only when ALL hold: real constraint (Q2), fully legible process (Q3), large no-human-touch share (Q4), in-bounds channel (Q5), a concrete KPI (Q6), and a phased rollout. Don't jump straight to L4.

When it lands between AUGMENT and ADOPT, take AUGMENT and let real-usage data promote it later. Default to the lowest autonomy level that works; prefer boring/deterministic over clever.

At the end — write the verdict, recap, hand off

1. Finalize the assessment file: fill the ## Verdict block — the call (ADOPT / AUGMENT / SKIP), the starting autonomy level, the KPI, the data-channel ruling, and a two-line "why this call."
2. Recap to Gera in chat, terse: the verdict, the one reason that drove it, and the suggested next step.
3. Hand off by verdict:
   • ADOPT / AUGMENT → suggest the right builder via Servy's triage (script / multi-brain pre-checks, then Skill vs Agent vs Routine), and offer to kick off /build or the specific builder. Pass the captured process map along — it's the start of the Discovery Interview.
   • SKIP → done. Optionally note it so the same idea doesn't keep resurfacing.
4. If the task should graduate into the wiki as reusable knowledge (a process Gera will reuse), offer that per references/wiki-protocol.md — don't auto-write without his yes.

Assessment file template

# AI Applicability: {Task}
Date: {date} · Task: {one line}

## Per-question log
### Q1 — Could it just stop existing?
- Recommended: {your inferred answer}
- Captured: {his answer}
- Decision: {ELIMINATE | CONTINUE}

### Q2 — Real constraint or just annoying?
- Recommended / Captured / Decision: {...}

### Q3 — Can you explain the steps to a person?
- Process map: Trigger / Data Sources / Transformations / Decision Points / Destination
- Decision: {READY-TO-SCOPE | MAP-FIRST}

### Q4 — What share can AI do, where does a human stay?
- Split: {fully-auto % / AI-assisted % / manual %}
- Starting autonomy level: {L0..L4}

### Q5 — Confidential/customer data, allowed channel?
- Data class: {public | personal | SoftServe Internal+ | customer IP}
- Decision: {PROCEED | RE-ROUTE | BLOCK}

### Q6 — What number does it move, rollout?
- KPI: {metric in a bucket}
- Rollout: {phased plan}

## Verdict
- Call: {ADOPT | AUGMENT | SKIP}
- Starting autonomy level: {L0..L4}
- KPI: {...}
- Data channel: {ruling}
- Why this call: {two lines}
- Next step: {builder / route / done}

Notes / guardrails

• One question at a time. Always. Each carries a recommended answer so Gera can confirm or redirect.
• Never skip a checkpoint. Append to the file after every answer.
• SKIP is a win, not a failure. Don't talk Gera into a build the framework says to skip.
• The data gate (Q5) is hard. Personal/public-data only; never route SoftServe Internal+ or customer IP through Servy or the multi-brain. When in doubt, BLOCK and route to an approved tool.
• Don't invent answers. If only a stakeholder knows, flag it with an owner and move on.
• This skill is interactive — it stays in the live session (do not fork to an isolated agent).
• It's a decision skill, not a builder. When the verdict is ADOPT/AUGMENT, hand off to the right builder; don't start building inside this skill.