Guided Engineering

Quick start

Ten minutes, one agent host, one small example. Tested with Claude Code on Windows; see Skills and agents for other hosts.

1. What you need

2. Get the example project

The example is a tiny repository with one package and two failing tests. Copy it out of the specification repository so that it is a project of its own:

git clone https://github.com/guided-engineering/OpenPrompt-Specification
git -C OpenPrompt-Specification checkout 9ed6800e5c7f32ce61073c8746fcedd4e23dde08
cp -r OpenPrompt-Specification/examples/project ops-demo
cd ops-demo

The commit checked out above is the state of the example before it was solved.

3. Install the Skills

In ops-demo:

npx skills add guided-engineering/skills -a claude-code -y

This copies ops-author and ops-execute into .claude/skills/. To install a fixed revision, name its commit. This one is the revision the local study measured; the current one differs from it by a single corrected sentence in ops-author:

npx skills add guided-engineering/skills#146c4b51fee8f1a6609b424fe6381c28c27c60e4 -a claude-code -y

No installer? Copy the two folders under skills/ of the Skills repository into .claude/skills/ by hand. The skills CLI sends anonymous install telemetry unless DISABLE_TELEMETRY=1 is set.

Now record the starting state, with the Skills in it, so that git status later shows only what the agent changed:

git init -q && git add -A && git commit -q -m "start"

The installer also writes a skills-lock.json; it is part of that commit.

4. Read the contract

The example ships with its request and with the OPS prompt written from it:

cat request.md
cat .guides/prompts/prompt.slug.clean-separators.yaml

To see the authoring step yourself, start your agent in ops-demo and ask:

Write an OPS prompt for the request in request.md. Do not implement it.

ops-author writes a prompt under .guides/prompts/, validates it, and tells you what it assumed. Compare it with the one that ships.

5. Plan, then execute

Ask for a plan first:

Plan .guides/prompts/prompt.slug.clean-separators.yaml. Plan only.

Then for the work:

Execute .guides/prompts/prompt.slug.clean-separators.yaml.

ops-execute validates the prompt, reproduces the two failing tests, changes packages/slug/src/slug.mjs, runs the tests again after the last edit, and writes a worklog entry under .guides/operation/worklog/.

6. Check the evidence yourself

Still in ops-demo:

node --test packages/slug/test/slug.test.mjs
git status --short

The worklog entry is a claim. The validator checks that it does not claim more than it shows:

node .claude/skills/ops-execute/scripts/ops-validate.mjs --all

Then read the diff. Whether the change is good is your decision.

In your own project

Nothing to set up beyond step 3. Describe a change the way you normally would and ask for it "as an OPS prompt"; review the YAML; ask for a plan; ask for execution; read the worklog entry with the diff. A project that wants more than the default agent adds .guides/agents/agents.yaml; see the contract.

What was verified on this page

The commands of steps 2, 3, 4 and 6 were replayed as written from the two public repositories on 2026-11-13, on Windows: the clone and the checkout, the copy, the installer (both Skills, the current version and then the measured revision), the starting commit, the two failing tests of the starting state, and the validator from the installed Skill.

The two requests of step 5 were not replayed in a fresh session for this page. The example in the repository was executed by the session that built it, and the same Skill files were exercised in the local study.