Back to search
trialrun
MarketplaceWorks with Claude CodeGitHub
Claude skill that coaches AI-literacy habits while you work with AI: check the output, keep the decision, protect the data
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Checked today
Badge for your README
Markdown
[](/listings/kostadindev--trialrun-skill)HTML
<a href="/listings/kostadindev--trialrun-skill"><img src="/api/listings/kostadindev--trialrun-skill/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
As a Claude Code plugin:
/plugin marketplace add kostadindev/trialrun-skill /plugin install trialrun@trialrun
Or copy the skill folder by hand:
git clone https://github.com/kostadindev/trialrun-skill mkdir -p ~/.claude/skills cp -R trialrun-skill/plugins/trialrun/skills/trialrun ~/.claude/skills/trialrun
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
.claude-plugin/marketplace.json · Manifest · 21 lines
- {
- "name": "trialrun",
- "owner": {
- "name": "TrialRun contributors"
- },
- "metadata": {
- "description": "Learn AI literacy while you work with AI, and run TrialRun assessments.",
- "version": "0.1.0"
- },
- "plugins": [
- {
- "name": "trialrun",
- "description": "Coaches three AI-literacy habits during real work (check the output, keep the decision, protect the data) and runs the TrialRun assessment.",
- "version": "0.1.0",
- "author": {
- "name": "TrialRun contributors"
- },
- "source": "./plugins/trialrun"
- }
- ]
- }
README.md · README · 114 lines
- # trialrun-skill
- [](https://github.com/kostadindev/trialrun-skill/actions/workflows/ci.yml)
- A Claude skill for learning **AI literacy while you work with AI**. Claude does your work and coaches three
- habits as you go:
- 1. **Check the output.** It says where each figure or fact came from and what to verify.
- 2. **Keep the decision.** When you hand over a judgement call, it gives the deciding facts and its
- recommendation, then ends with a `Your call:` line.
- 3. **Protect the data.** It notices personal details, credentials and confidential material in what you
- share, and tells you what to leave out next time.
- At most one short note per reply (about 25 words, never more than 30), only when a habit is at stake. Ask "how am I doing?" for a
- reflection with a `Habit to keep:` and a `Habit to build:`. It never repeats restricted data, and it **never
- plants errors in real work**.
- The skill also runs the **TrialRun assessment**: Claude plays Ledger, a scripted assistant that makes planted
- mistakes in an accounting task, then grades the attempt and writes a replay.
- ## Install
- As a Claude Code plugin:
- ```
- /plugin marketplace add kostadindev/trialrun-skill
- /plugin install trialrun@trialrun
- ```
- Or copy the skill folder by hand:
- ```bash
- git clone https://github.com/kostadindev/trialrun-skill
- mkdir -p ~/.claude/skills
- cp -R trialrun-skill/plugins/trialrun/skills/trialrun ~/.claude/skills/trialrun
- ```
- ## Use
- ```
- Coach me on AI literacy while we work. Here's our payroll export: what's the average salary?
- Coach me on AI literacy while we work. Should we capitalise this £18,000 roof repair?
- How am I doing on AI literacy so far?
- Run a short TrialRun drill and grade it.
- ```
- Example note: `Literacy note: only the salary column was needed. Next time, delete names, addresses and NI numbers before pasting.`
- ## With trialrun-mcp
- **Coach mode** works on its own. With the [trialrun-mcp](https://github.com/kostadindev/trialrun-mcp) gateway
- pointed at your folder, the skill tells Claude to read your files through it, so those reads are logged and
- restricted data is detected; it records each coaching moment in the gateway's habit log, and your reflection
- draws on the gateway's report, which keeps labels only, never raw values.
- The gateway sees only what goes through it. The skill instructs Claude not to use built-in file tools while the
- gateway is connected, but Claude Code doesn't enforce that. To guarantee every read is logged, also disallow
- Read, Glob and Grep in your Claude Code permissions. Text you paste into the chat is only checked if Claude calls
- the gateway's `check_text`.
- ```bash
- claude mcp add trialrun -e TRIALRUN_WORKSPACE=/path/to/folder -- npx -y github:kostadindev/trialrun-mcp
- ```
- **The assessment modes (drills, grading) require the gateway in scenario mode**, which serves the Fernvale
- Kitchens files. Without it the skill stops rather than invent files:
- ```bash
- claude mcp add trialrun -- npx -y github:kostadindev/trialrun-mcp
- ```
- ## Evaluation
- Real `claude -p` sessions with the skill and the gateway, scored by code (`evals/agent/`). Each case runs three
- times.
- | Run | Runs fully passing | Checks passing |
- |---|---|---|
- | 8 tuning cases, round 1 | 3/24 | 192/219 |
- | 8 tuning cases, round 3 | **24/24** | **219/219** |
- | 5 held-out cases, first run | 12/15 | 114/117 |
- The skill passed every held-out check; the three held-out failures were a gateway detection gap, since fixed.
- **These cases cover coach mode only.** Ledger's scripted traps, grading and scenario authoring have no live
- agent cases yet. Details, every check and every change to the checks: [`evals/agent/README.md`](evals/agent/README.md).
- ## Development
- ```bash
- npm install
- npm run check # typecheck, unit tests (a failing fixture for every case check), plugin validation; runs in CI
- npm run agent -- --gateway /abs/path/trialrun-mcp/dist/server.js # live agent runs (manual; needs Claude Code)
- npm run agent -- --gateway /abs/path/trialrun-mcp/dist/server.js --builtin-readers # readers enabled: does it still use the gateway?
- npm run agent -- --no-gateway # coach mode without the gateway
- ```
- | Path | What it holds |
- |---|---|
- | `plugins/trialrun/skills/trialrun/` | The skill: `SKILL.md` and `references/` |
- | `.claude-plugin/`, `plugins/trialrun/.claude-plugin/` | Marketplace and plugin manifests |
- | `evals/agent/` | Agent evaluation harness, cases, unit tests and results |
- | `scripts/` | Plugin and skill validation |
- ## Limitations
- - Coaching is judged by the model, so it can miss a moment or add one that isn't needed.
- - Answers to technical questions can be long.
- - Without the gateway, the skill sees only what's in the conversation.
- Part of TrialRun, built for Cambridge EduX 2026.
- ## Licence
- MIT
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.