6 Skills for AI and Psychological Safety

What these skills are

A concern put into words on one side, a decision tested before it lands on the other. Someone knows something needs raising and cannot find the wording for it. Or they have the wording and are not sure the route is safe. Somewhere else in the same firm, a plan is about to be announced by a person who will hear about the problem with it afterwards. The pack sits on both sides of that: two skills for the person with something to say, two for the decision that is still open, and two for a concern that has already arrived.

AI helps prepare the work. People choose what to share, hear each other, check the facts, and make the decisions.

On the speaking side, concern-clarifier keeps the observation, the interpretation, the possible consequence and the actual request apart, writes the opening in your language and level of formality, and helps you choose a route — a direct conversation, a written note, an intermediary, or an established confidential channel. It preserves the strongest legitimate point instead of softening it into an apology, and it will not promise you anonymity that a workplace tool cannot deliver. speak-up-rehearsal then plays the other person one turn at a time, pushes back on the substance rather than on you, and hands back a revised opening.

On the deciding side, ai-decision-stress-test separates what you know from what you are assuming, names the objections that would matter most and the evidence that would support or weaken each, and turns the unresolved ones into neutral questions for the people who will implement the thing. invite-team-challenge designs the consultation that follows — written and private routes included, and what can still change said out loud. respond-to-ai-raised-concerns takes a concern that arrived through an AI-assisted or unattributed document and assesses the argument on its merits, declining the part where you go looking for the author. close-the-concern-loop writes what changed, what did not, and why, with each commitment owned, timed, and marked proposed until agreed.

Mechanically, each skill is one folder with a SKILL.md file, and each one also ships as a portable prompt you can paste into a plain Claude chat — the whole workflow in one file, no connector needed. PROJECT-INSTRUCTIONS.md routes a described situation to the right one of the six, so you can describe the problem rather than pick from a menu. The concern-and-response note, four fictional practice examples and the evidence notes travel with the pack.

Download all 6 skills. One zip: six ready-to-install skill zips, readable SKILL.md files, six portable prompts for a plain Claude chat, the concern-and-response note and the evidence notes. Free, no signup. The same six skills are open on GitHub: polar-bear-org/claude-skills.

The 6 skills you get

Two for the person with something to say, two for the decision that is still open, and two for a concern that has already landed. The numbers are the pack's own.

If something needs saying

2. Put a Concern Into Words

Use when: You know something needs raising and the wording will not come, it crosses into someone else's discipline, or you need it in another language

Output: A short concern statement in your own voice and level of formality, one concrete question or request, and a suggested route — observation, interpretation, possible consequence and ask kept apart, and no promise of anonymity a workplace tool cannot keep

3. Rehearse Speaking Up

Use when: You want to practise questioning someone with more authority before you do it for real

Output: A role-play played one turn at a time, feedback on observable choices rather than on your courage, and a revised opening with a clarifying question and a way to pause the conversation — or just the script, if that is what you wanted

If you are the one deciding

1. Test a Decision Before Announcing It

Use when: You are about to put out a plan, a policy, a review-process change or an AI rollout and want it challenged first

Output: A challenge brief: the proposal, the strongest case for it, the most consequential plausible objections with the evidence that would support or weaken each, and neutral questions for the people who will implement it — every objection labelled a hypothesis, never employee testimony

4. Make Room for Challenge

Use when: You want input while the decision is still open, across differences in authority, confidence, language and expertise

Output: A draft invitation, spoken, written and private ways to take part, neutral questions sent in advance, and a response commitment for the owner to confirm — with what can still change said plainly, and explanation labelled as explanation when it cannot

If a concern has already landed

5. Respond When AI Carries the Concern

Use when: A concern arrives through an AI-assisted argument or an unattributed document and you have to answer it

Output: An issue-and-evidence note that keeps the claim separate from how it arrived, a response draft saying what will be checked, and a follow-up step with an owner and a date to confirm. Ask it who wrote the thing and it declines that part and carries on

6. Show What Happened Next

Use when: People raised something, you have decided, and they are waiting to hear what came of it

Output: A response-and-action note: what changed, what did not and why, each commitment with an owner, a timing and an observable sign of completion, marked proposed until someone agrees — plus a route for correction and a proportionate check-in

No sequence, no assessment, and a clear edge

There is no compulsory sequence here and nothing to score. You choose one situation, open the matching skill, and describe the work issue in aliases with only the detail you need to share. If you are working from project instructions, the routing picks the smallest useful workflow from what you describe, asks one question when the choice is genuinely unclear, and respects a skill you name yourself. It will not run all six at you.

Two of them do pair up naturally. A colleague who has used concern-clarifier to get the wording right can take that opening into speak-up-rehearsal before the real conversation. A leader who has run ai-decision-stress-test has the questions that invite-team-challenge then builds a consultation around — and if the requirement is already fixed, the invitation has to say so and stick to the implementation choices that genuinely remain open.

The pack is informed by a Harvard Business Review article on AI and psychological safety, by Jayshree Seth and Amy C. Edmondson, and by Edmondson's 1999 study of learning behaviour in work teams. Those are a framework and practitioner observations, not four established effects of using AI. The evidence notes that ship with the pack set out what the sources say, what is Polar Bear's own design choice, and what nobody has shown: we have not tested whether using these skills improves psychological safety or workplace outcomes, and the pack is not a validated intervention or an assessment.

So it stops where your judgement starts, and it is blunt about where that is. AI-generated objections are hypotheses to take to people, not evidence of what your team thinks. Someone using AI to write does not tell you they are afraid. A rehearsal cannot tell you how the other person will react. And the skills produce drafts: they do not send the message, read private logs, keep employee records, rate anyone, or make an employment decision. Where a concern involves retaliation or serious misconduct, all six point at the appropriate human process instead.

Setup guide

1. Download the pack. One zip: an install folder with six ready-to-upload skill zips, a skills folder with the same six skills as readable files, six portable prompts plus a combined ALL-SKILLS.md, the project instructions, the concern-and-response note, four fictional practice examples and the evidence notes.
2. Install your skills. In Claude Code, add the marketplace and install ai-psychological-safety-pack, and all six load at once. In Claude, go to Customize, then Skills, and upload one zip per skill from the install folder — not the whole pack as though it were a single skill. Feature availability and the interface labels depend on your plan and workspace.
3. Or paste the portable prompt into whatever you already use. In a plain Claude chat, paste one portable/[skill-name].md file into a new chat and describe the situation: each prompt carries the whole workflow, with no connector or external reference needed. Where your assistant supports project sources, add portable/ALL-SKILLS.md as a source and PROJECT-INSTRUCTIONS.md for routing, and paste the individual prompt instead if retrieval is unreliable. Check your workspace's sharing settings before you add work information.
4. Describe the situation, in aliases. Pick one situation; there is no compulsory sequence and no assessment. Describe the work issue using aliases and only the information you need to share. Then check the output, adapt the wording, and decide what human conversation or action comes next — that part is still yours.

Where to start

The quality bar

Every skill in the pack holds the same standard, the one we hold when we do this work ourselves:

• AI-generated objections are hypotheses to check with evidence and people — never employee testimony, statistics, probabilities, or a consensus the skill made up
• Using AI to draft, translate, or work faster does not establish that someone is afraid, and silence does not diagnose a team's culture
• Nothing here promises anonymity: workplace tools and records can retain identities, prompts, and access logs, and the skills say so rather than reassure you
• Ask any of the six to identify an anonymous author from prompt logs, writing style, or metadata and it declines that part, then carries on with the substantive response
• No scores: not on courage, candour, personality, psychological safety, or employability, and nothing produced in the work is used to assess a participant
• The person keeps their concern: polished wording never erases their uncertainty, their urgency, or their voice, and they decide whether, when, and with whom to share it
• A consultation that cannot change the outcome is labelled explanation or implementation input, rather than advertised as influence people do not have
• Suggested owners and dates stay marked as proposals until someone agrees to them, and a commitment is not complete because the note was drafted
• The skills prepare drafts and practice conversations: they do not send messages, inspect private logs, keep employee records, or make employment decisions
• A serious allegation, or a credible fear of retaliation, goes to the appropriate human process — not into a group exercise or a public debate

Who made this

Polar Bear is a people ops consultancy for human-size teams (20 to 200 people). Built by ex-McKinsey founders with a dream to make AI work for People, not instead of them. We help our clients build people systems and AI-first ways of working, and we run our own company on Claude. This pack is the free, self-serve version of how we work.

This pack prepares the conversation and then gets out of the way. When you want that to be how the whole company works, AI carrying the drafting and the rehearsal so people do the listening, the checking and the deciding, across strategy, delivery, and everyday work, that's what we build with clients.

Meet Alexey. Wondering whether people tell you what they actually think? Bring a decision you are about to announce, or a concern you are not sure how to raise. Find 30 min together · Alexey on LinkedIn

Install in Claude Code

Two lines, and every skill in the pack loads at once.

/plugin marketplace add polar-bear-org/claude-skills
/plugin install ai-psychological-safety-pack@polar-bear-skills

Using Claude on the web instead? Download the zip and upload each skill from its install folder under Customize → Skills.