Method
How an entry is written, checked and kept current
The Field Manual is a library of procedures, not prompts. A prompt is at most one line inside a step; the entry is everything around it that makes the result trustworthy.
The information model
- Outcome
- What you have at the end. If it cannot be stated in one sentence, the entry is too broad.
- Use it when / Skip it when
- The fit test. Skip-when lines are required, because knowing when not to use a method is part of the method.
- Before you start
- Prerequisites: data, access, permissions and people you need first.
- Steps and checkpoints
- Three to nine steps. Every step ends with a checkpoint you can verify before moving on.
- Works in
- Environments with full or partial support and a note. Environments that are not listed were not tested.
- Effort
- Setup minutes (once) and minutes per run. Estimates from the reviewer's run, not guarantees.
- What can go wrong
- Safety notes, each tagged with a risk type: accuracy, privacy, security, cost, rights or effect on people.
- Review record
- State, version, last review date and revision history. Review due is computed against the catalogue date.
Review states
- Reviewed
- Steps were re-run end to end at the recorded review.
- Needs review
- Published, but something it depends on changes quickly; re-run before relying on it.
- Draft
- Not published. Drafts are validated but kept out of the site.
An entry is marked review due when its last review is more than 180 days before the catalogue date (1 Mar 2026, a fixed demo date). The build computes this; nobody sets it by hand, and the visitor's clock is never used.
Environments
Environments are kinds of tool, not brands, so entries stay true as products change.
- Chat assistant (Chat)
- Model API and scripts (API)
- Coding agent in a repository (Agent)
- Spreadsheet (Sheet)
- Local model on your machine (Local)
- No-code automation (Flow)
What demo means here
Every entry in this build has "demo": true and the build refuses entries without it. The procedures are written carefully and are meant to be sound, but they were written to prove the format: they are not an exhaustive catalogue, not endorsed by any vendor, and their review dates and version histories are sample records rather than a real editorial log. Worked examples use fictional inputs.
Adding an entry
- Create a draft:
python3 tools/new_skill.py my-skill-id "Title of the skill". It writesdata/skills/my-skill-id.jsonwith the next free code and the statedraft. - Fill in every field. Run
python3 build.py --check; it lists every problem with the file and field. - Run the steps end to end yourself. Then set the state to
reviewed, addreviewed_onand a history entry. - Run
python3 build.pyandpython3 tools/audit.pybefore publishing.
What works now, and what does not yet
Works in this build
- Search and filters by task, level, environment and area, kept in the page address.
- A page per entry and per task that works without JavaScript.
- Validation that refuses incomplete or inconsistent entries.
- A JSON export of the published catalogue at api/skills.json.
Needs more than a static build
- Real editorial review by named reviewers, with dates that mean something.
- Reader feedback on whether a step worked, which needs a backend and a privacy review.
- Versioned entries per tool release, which needs a maintained compatibility source.