1
Nudge
Looks at the last hour and picks the strongest coaching moment for a skill on the path. It writes one nudge with a prompt that is ready to paste.
Skill Coach · for developers in AI-era teams
Team Developer works out who should grow into which role and which skills they need. Skill Coach is the other half. It is a personal coach that runs on the employee's own machine. It notices moments in daily work where the new way of working would apply, and suggests a small experiment right then. Nothing leaves the laptop, and the coach never reports to a manager.
The principle
The best exercise is not a course. It's the next real ticket, done a new way.
The coach suggests experiments of 10 to 30 minutes, tied to what is on screen right now. It makes one change at a time: at most one nudge per hour and one focus skill per day. A skill counts as practised when there is something to show for it, such as a reviewed agent diff, a spec or an automated test.
Derived from the learning path and yesterday's activity.
When a matching situation appears on screen, with a ready-to-paste prompt.
Where the time went, what was tried, and one workflow change for tomorrow.
A day with David
David is the fullstack developer from the Team Developer example. His learning path points to AI-augmented Product Engineer. Below is a scripted replay day, classified by a local model. Select a moment to see what the coach sees, why it counts as a coaching moment, and which playbook advice it draws on.
Scripted replay day · fictional person · shipped with the prototypeAlso on screen that day: a 1Password window with a password. Excluded apps are filtered out before anything reaches the model.
Not every moment becomes a nudge. The coach sends at most one per hour and won't repeat a skill within 45 minutes. The evening review picks up the strongest moments of the day.
What it does
The coach only comments on skills that are on the person's learning path. It doesn't judge everything on screen. It looks for the moments where a skill on the path could be practised.
Looks at the last hour and picks the strongest coaching moment for a skill on the path. It writes one nudge with a prompt that is ready to paste.
Checks every 15 minutes and shows a desktop notification when there is something worth trying. It sends at most one per hour.
An evening review: where the time went, the AI-assisted share, progress per skill, at most three workflow changes for tomorrow and one focus skill.
Answers questions from the playbooks, the learning path and the last two hours of work. For example: “How do I review Claude Code's CRM change before pushing?”
Practised hours against the learning hours on the path, evidence per skill, and how helpful each type of nudge was rated.
Records proof of a skill and exports it in the format Team Developer understands. The export happens only when the employee chooses to share it.
Privacy by architecture
A coach that watches the screen only works if people can trust it. That's why every part of Skill Coach runs on the employee's own machine, and the employee decides what is shared.
On-screen text (OCR and accessibility text) and audio transcripts, recorded locally.
Classifies work sessions and writes nudges and reviews. It never calls a cloud API.
A small vector index of the skill playbooks used to retrieve advice.
Nudges, feedback, observed practice, evidence and daily reviews.
Evidence of a skill and progress, exported in the format Team Developer uses. There is no report to a manager and no team view. That is by design.
Private apps are dropped before any processing.
Windows whose title contains certain keywords are dropped as well.
Email addresses, tokens and passwords are removed from the text before the model sees it. Both lists can be configured.
Where the coaching comes from
The coach gets its goals from Team Developer and its advice from playbooks for each skill. You can extend the playbooks with your own material.
Team Developer's analysis provides each person's suggested future role and the skills to learn. Skills are ranked by the assigned hours per month that depend on them, with learning and coaching hours for each. The coach only works with skills on this path.
For David this adds up to 88 learning hours, about 5.5 months at 16 learning hours per month.
| Skill to learn | Dependent h/month | Learning |
|---|---|---|
| AI pair programming with Claude Codehigh | 52.4 h | 16 h |
| Reviewing AI-generated codehigh | 44.0 h | 16 h |
| Agent orchestrationmedium | 16.1 h | 24 h |
| Spec-driven developmentmedium | 16.1 h | 16 h |
| Context engineeringmedium | 8.4 h | 16 h |
Plus 8 coaching hours per skill. Snapshot of David's path, shipped with the prototype.
Each playbook describes what good practice looks like, micro-exercises inside real work, workflow changes, anti-patterns, and what counts as evidence for the learning path.
Excerpt · Reviewing AI-generated code
How it works
A nudge fires because an explainable rule matched, for example “agent edits accepted wholesale and pushed”. The model only writes the text, using the playbooks it retrieves. This keeps interruptions predictable, and it makes the feedback for each rule meaningful.
OCR, accessibility text and audio transcripts from screenpipe.
Excluded apps and windows are dropped. Emails, tokens and passwords are redacted.
Activity in the same app with no gap over 5 minutes. Sessions under 2 minutes count as noise.
The local LLM maps each session to a task on the path, or to communication, meeting or research. Rules then flag coaching moments.
Playbook sections for the moment's skill.
The local LLM writes the nudge or the review. The employee rates it: helpful, not helpful, or tried it.
Agent change not reviewed · spec still open · routine work done by hand · documentation written by hand · requirements clarified in a meeting.
Minutes count as practice when an AI tool is visibly in use on a task that needs the skill. Hand-typed work doesn't count, and accepting an agent's edits wholesale doesn't count as reviewing them. People can add practice themselves as well.
Each nudge is rated helpful, not helpful or tried it. The ratings for each trigger show which rules work and which interrupt for nothing.
Honest by design
Skill Coach is a prototype. These are the points to know before running it with real people.
The triggers are a handful of heuristic rules. Real screen data will need more and better rules, or an LLM judge on top of them.
OCR text is noisy and incomplete. The classifier sees 450 characters per session.
Practice is approximated by AI-tool use on a path task. It says nothing about quality. Evidence such as a reviewed diff, a spec or an automated test is the stronger signal.
Observed hours per task don't flow back into Team Developer's catalog yet. The stored sessions, practice and evidence are the data for it.
No sync and no team view, by design.
A works council agreement, a DPIA and a clear opt-in are required before any use with real employees.
In a free first conversation we'll talk about your team, the skills that matter now, and whether a local coach fits your setup and your works council process.
Get in touch
No sales call. You speak directly with the founder about your team and whether Skill Coach fits your situation.
FAQ