coaching-colleagues-on-ai

Guides informal AI coaching conversations with Loxihart colleagues — from first questions to skeptic conversations to helping power users go deeper. Activates when a colleague asks about AI tools, whe

SKILL.md

GitHub

Colleague AI Coaching

Handles informal AI coaching conversations at Loxihart. Reads the opener, calibrates to the person, gives one concrete thing, and follows up.


Evals

Should trigger:

  1. "A coworker asked me how they'd even get started using Claude — what do I tell them?"
  2. "One of my teammates thinks AI is useless and doesn't see the point — how do I talk to them about it?"
  3. "A peer tried Claude, got a bad result, and asked me why it didn't work"

Should NOT trigger:

  1. "Help me get better at using Claude for my own work" — that's improving your own usage, not coaching another person; this skill doesn't fit (no specific alternative — just out of scope)
  2. "Write a leadership brief about AI adoption for [PERSON]" — formal leadership communication → briefing-loxihart-leadership

Decision Tree

What's the opener?
  ├─ "I want to try this / how do I start?" → Curious Beginner (Section 2)
  ├─ "I tried it but the results weren't useful" → Disillusioned (Section 3)
  ├─ "I'm not sure AI is right for me / my work" → Skeptic (Section 4)
  ├─ "I'm already using it — I want to do more" → Power User (Section 5)
  └─ "I'm worried about my job / AI replacing people" → Existential (Section 6)

If the opener is ambiguous, ask one question: "Have you tried it yet, or are you just starting to think about it?"


Section 1: Core Principles

Full conversation patterns in references/conversation-patterns.md. Adoption signal logging in references/adoption-signals.md.

Ask before explaining. "What have you tried so far?" surfaces the actual blocker faster than any tutorial. Most colleagues who ask for help have already tried something — find out what before restarting from zero.

Give one thing, not a curriculum. The most common coaching mistake is overwhelming the person with options. Identify the single highest-value thing they could try based on their actual work tasks. One thing. Not three. Not a list.

Acknowledge before redirecting skeptics. A colleague who is worried about AI is not ready for a demonstration. They need acknowledgment first. "That's a real concern a lot of people have" is not patronizing — it's the prerequisite for any productive next step.

Follow up once. Three days after a session, send a single Slack message: "How did [the thing you showed them] go?" One follow-up. Not a check-in series. The follow-up signals you care and creates accountability without pressure.

Log the signal. After every coaching conversation, note in your adoption log what the person tried, what their blocker was, and whether they're using AI now. This is the qualitative data that the survey misses. See references/adoption-signals.md.


Section 2: Curious Beginner

Someone who wants to try but hasn't started, or has just started and wants guidance.

Opener pattern:

  1. "What kind of work tasks take up the most of your time right now?"
  2. Map their answer to a single AI use case for that task
  3. Show it live — open Claude, type the prompt, show the output
  4. Let them try one variation while you're there
  5. "Try this on a real task this week and let me know how it goes"

The one-thing selection guide:

  • Lots of writing / documentation → show prompt-for-draft pattern
  • Lots of research / reading → show summarization pattern
  • Repetitive tickets or requests → show template generation
  • Troubleshooting → show "explain this error" pattern
  • Uncertainty about format or tone → show rewrite/reformatting

What not to do: Don't open with "so, AI is really changing everything..." — start with their task, not the technology.


Section 3: Disillusioned

Someone who tried Claude (or another AI tool) and got bad results.

Opener pattern:

  1. "What did you try and what happened?"
  2. Diagnose the gap: bad prompt, wrong tool, unrealistic expectation, or real limitation?
  3. Show the fix for the specific thing that failed — not a general tutorial
  4. "The output you got makes sense given how it was prompted — here's the adjustment."
  5. Let them re-run their actual failed task with the fix applied

Common gaps and fixes:

What happened Likely cause Fix
Output was too generic Prompt had no context Add role, context, and constraints
Output was wrong/hallucinated Treated it as a search engine "Claude reasons, it doesn't look things up — verify factual claims"
Output was too long / too short No length instruction Add "in 3 bullet points" or "in one paragraph"
Didn't do what I asked Ambiguous ask Rewrite ask as a command with explicit output format
Got worse results than expected Expectations were calibrated to demos "Demos show best-case — real use requires iteration"

What not to do: Don't minimize their frustration ("oh it works great once you learn it") — diagnose what actually went wrong.


Section 4: Skeptic

Someone uncertain whether AI is appropriate for their work, or doubtful it will be useful.

Opener pattern:

  1. "What specifically concerns you about using it?" — don't assume the concern; ask for it
  2. Reflect the concern back accurately before responding to it
  3. Find the smallest possible demonstration relevant to their actual concern
  4. "What would it take for you to think it was worth trying on one task?"
  5. Agree on that one low-stakes task. Don't push further.

Common concerns and responses:

Concern Response approach
"I might make mistakes / publish something wrong" Agree that verification is required. Show how to use it as a draft, not a final output.
"It doesn't know our systems or processes" "That's right — it's a thinking and writing tool, not a database. Show me a task where thinking or writing is the bottleneck."
"I don't have time to learn a new tool" "What's the one task that takes the longest and requires the least judgment? That's where 20 minutes pays off immediately."
"It's not accurate enough" "For what specifically? Let's find one task where accuracy is verifiable — so you can check it and decide."

What not to do: Don't argue with the concern. Don't demonstrate something impressive unrelated to their actual work. Don't push for immediate adoption.


Section 5: Power User

Someone already using AI tools who wants to go deeper.

Opener pattern:

  1. "What's the highest-friction thing in your workflow right now — even if you're not sure AI can help?"
  2. Look for: repetitive tasks with structure (automation candidates), multi-step tasks (chaining candidates), knowledge retrieval tasks (context-injection candidates)
  3. Show one pattern upgrade: system prompt, multi-turn strategy, or tool use pattern appropriate to their current level
  4. Share one resource (a specific skill, a specific prompt pattern) rather than a general "here's more to explore"

Don't overwhelm with options. A power user's risk is breadth without depth — they know a lot of mediocre approaches instead of a few excellent ones. Help them go deeper on one thing they're already doing.


Section 6: Existential Concern

Someone worried about AI replacing their job. This is a real concern and cannot be bypassed with a demonstration.

This conversation is not a coaching session — it is a human conversation first.

  1. Don't start with "well, actually AI creates jobs too..."
  2. Acknowledge the concern directly: "That's not a paranoid thing to worry about — a lot of roles are changing."
  3. Ask: "What specifically are you worried about — is it your current role, or something longer-term?"
  4. Based on their answer: either discuss what's actually changing in their specific domain, or (if they're open to it) explore how learning AI tools could strengthen their position specifically
  5. Do not push toward adoption. This conversation has succeeded if they feel heard.

If they're open to exploration after the above: treat them as a curious beginner and run Section 2.


Self-Verification

Before any coaching session:

  • I know what kind of conversation this is (beginner / disillusioned / skeptic / power user / existential)
  • I have one specific thing prepared to show, not a curriculum
  • I will ask what they've tried before demonstrating anything
  • I have the follow-up reminder set for 3 days out

After any coaching session:

  • Logged the adoption signal (see references/adoption-signals.md)

CLAIMS

  • Giving one concrete thing to try is more effective than a comprehensive overview — it reduces decision paralysis and creates a clear success-or-failure signal the person can report back on
  • Asking what they already tried before explaining anything surfaces the actual blocker faster than a tutorial approach
  • Logging coaching conversation signals creates the qualitative data that quantitative adoption surveys consistently miss

COUNTEREXAMPLE

  • The "one thing" approach fails when the underlying concern is existential (job threat) rather than practical — that conversation requires acknowledgment before any task demonstration is appropriate

CONTRADICTIONS

  • "Don't give a tutorial" conflicts with situations where someone genuinely needs step-by-step onboarding — read the opener; some people need the structured walkthrough, especially for first-ever use