A prompting coach for teams. Every prompt sent through Claude.ai, Claude Code,
VS Code or Copilot Chat is scored on five dimensions before it goes out. Weak
prompts get a short teaching loop; strong ones join the team's prompt library,
and the conventions people state are collected into a team wiki that is fed
back into the next person's context. A team's way of prompting compounds
without anyone writing docs.
A team-wide prompting coach. Scores every prompt sent through Claude Code, Claude.ai, VS Code or Copilot Chat on 5 dimensions in real time, coaches the weak ones, and feeds the strong ones into a team wiki that gets injected back into the next person's context.
Run the whole stack, no API key needed:Try it below
Built at PoliHack v19 (April 2026, BMW track "Applications that encourage AI
adoption") by @Bogzx,
@KunMihai1,
@bbeatricecretu and
@CosovanuGabi912. Open source (MIT) and
self-hosted: there is no hosted service to sign up for.
Trailhead is the codename used inside the code (package names, settings,
env vars); LearnLoop is the product name.
Score. A prompt is sent (Chrome extension) or a code task starts
(Claude Code / Copilot Chat via the MCP coach tool). The API scores it on
the five dimensions below and records the scores for the team's skill arc.
Coach. At 7/10 or more it goes through untouched. Below 7, the
extension holds the send and offers Improve / Send as-is / Edit;
nothing is sent automatically. In Claude Code the coach tool runs a short
teach → revise loop (at most 5 rounds) on the weakest dimension, using a
teammate's proven prompt as the example when there is one.
Promote. A prompt that clears a strict bar (unrounded mean ≥ 7, no
dimension below 5, and an independent re-score that agrees) joins the
team's prompt library, optionally after a teammate's review.
Remember. When someone states a team convention, wiki_save records
it; repeated mentions are deduplicated and become durable after three.
Reuse. The next prompt about the same part of the codebase is scored
and coached with that wiki node's conventions and library prompts in
context, so the rubric reads against how this team works.
The rubric, and how the score is measured
Every prompt is scored 0–10 on:
goal_clarity — the desired outcome is stated unambiguously
specificity — the change itself is named (files, functions, errors)
context_loading — the relevant code, docs or examples are referenced
constraint_articulation — what must not change; limits and invariants
output_specification — the shape the answer should take
overall is the rounded mean. The rubric text lives in
packages/scoring/src/score-prompt.mjs, and the score comes from one Gemini
call, so it varies from run to run. apps/api/eval/ measures how much on a
golden set of 30 prompts plus a held-out set of 16, reporting run-to-run
spread, how often the coaching and library gates flip, band hits, and whether
better prompts are ranked above worse ones.
What has been measured so far is the rule-based scorer
(packages/scoring/src/heuristic-score.mjs), a transparent baseline that a
model scorer should beat; the two have not been compared yet. CI checks it on
every change:
Prompt set
Overall in expected band
Pairs ranked the right way round
golden (30, visible while its rules were written)
29/30
152/152
held-out (16; per its author, written before the rules were frozen and not tuned on, but added in the same commit, so this can't be checked from history)
14/16
38/38
The bands were written in this repo, not by independent reviewers, and the
Gemini scorer has not been run on either set yet (it needs a key; one command
in apps/api/eval/README.md). Known misses and
the full method are documented there.
Try it
In your browser, 10 seconds: the Try it section of the landing
page runs the rule-based scorer on
whatever you type. Nothing leaves the page.
The whole stack, no key, about 2 minutes (Docker with Compose v2):
git clone https://github.com/Bogzx/LearnLoop && cd LearnLoop
TRAILHEAD_LLM=offline docker compose --profile demo up --build
# API on http://localhost:3000, Postgres schema applied, demo team seeded
curl -s -X POST localhost:3000/score -H 'content-type: application/json' \
-H 'X-Team-Token: trailhead_demo_acme_2026' \
-d '{"prompt":"fix the retry","user_id":"me"}'
Offline mode has no model: scoring uses the rule-based scorer, coaching uses
the static templates, and every response says so. --profile demo loads a
fictional team ("Acme Fintech") with a small wiki, four library prompts and
six days of synthetic activity, so the dashboard has something to show.
Then point the clients at it, all described in
SELFHOSTING.md:
# Dashboard on http://localhost:3001 (server-side env; the secret never reaches the browser)
TRAILHEAD_API_URL=http://localhost:3000 TRAILHEAD_TEAM_TOKEN=trailhead_demo_acme_2026 \
npm --workspace=apps/dashboard run dev
# Claude Code / Copilot Chat in one of your repos (the MCP package is not on npm,
# so the CLI runs from this clone; `init` registers the repo's team)
cd /path/to/your/repo
node /path/to/LearnLoop/apps/mcp-server/bin/cli.mjs init
Packages of the Chrome extension (.zip, load unpacked) and the VS Code
extension (.vsix) are built by every CI run (the extensions artifact) and
attached to GitHub releases
from v0.1.0; neither is on a store.
Hono + TypeScript on Node 22, raw pg. Multi-tenant: each team has a public id and a server-minted secret (stored hashed). Scoring, coaching, library, wiki, rich bootstrap, rate limits, Langfuse tracing.
MV3, on claude.ai. Score card under the composer, score badges on each message, Compare to team diff, Improve chat, outcome chips, wiki toasts, context picker. Every API call goes through the service worker.
STDIO server for Claude Code and Copilot Chat: coach, wiki_lookup, wiki_save, wiki_bootstrap, wiki_proven_prompts (+ ping), and a CLI: init, bootstrap, reset.
Postgres holds eight tables (packages/db/schema.sql): teams, wiki nodes,
learnings, library prompts, captures, skill observations, and the two
rich-bootstrap job tables.
Development
Node ≥ 22.6, npm workspaces.
npm ci
npm run typecheck # tsc --noEmit in every workspace
npm run lint # ESLint flat config
npm test # every workspace's tests, including the scorer baseline
npm run build # every workspace with a build script
# API integration tests against a throwaway Postgres (they wipe it):
docker run -d --rm --name trailhead-it -p 55432:5432 -e POSTGRES_USER=trailhead \
-e POSTGRES_PASSWORD=trailhead -e POSTGRES_DB=trailhead_it postgres:16-alpine
TRAILHEAD_IT_DATABASE_URL=postgresql://trailhead:[email protected]:55432/trailhead_it \
npm --workspace=apps/api run test:integration
CI runs all of that, plus npm audit, a Docker build with a
docker compose up smoke test in offline mode, and packaging of both
extensions. A v* tag attaches the extension packages to a draft release
(.github/workflows/release.yml).
Without Docker, the API runs against any Postgres:
psql "$DATABASE_URL" -f packages/db/schema.sql, optionally
node packages/db/seed.mjs, then npm run dev (reads .env).
Deployment
API: self-hosted. docker compose as above, or any Node 22 host
(railway.json is kept for Railway). Every environment variable is in
SELFHOSTING.md → All variables.
Dashboard: Vercel or any Next.js host. Set TRAILHEAD_API_URL and
TRAILHEAD_TEAM_TOKEN server-side; anyone who can open it can read that
team's data, so restrict access.
Landing page: static files on Vercel.
Extensions: from the release packages, loaded unpacked / installed from
the .vsix.
Specs
The project was specced before it was built; the specs are the record of
why:
A team-wide prompting coach. Scores every prompt sent through Claude Code, Claude.ai, VS Code or Copilot Chat on 5 dimensions in real time, coaches the weak ones, and feeds the strong ones into a team
A team-wide prompting coach. Scores every prompt sent through Claude Code, Claude.ai, VS Code or Copilot Chat on 5 dimensions in real time, coaches the weak ones, and feeds the strong ones into a team wiki that gets injected back into the next person's context.
ตรวจสอบล่าสุดเมื่อ 2026-10-28 A team-wide prompting coach. Scores every prompt sent through Claude Code, Claude.ai, VS Code or Copilot Chat on 5 dimensions in real time, coaches the weak ones, and feeds the strong ones into a team wiki that gets injected back into the next person's context.