Skip to content

_bootstrap_skill.md

Latest commit

 

History

History
150 lines (107 loc) · 6.16 KB

_bootstrap_skill.md

File metadata and controls

150 lines (107 loc) · 6.16 KB
name _bootstrap_skill
description Meta-skill the agent runs when a user requests a lensing capability that isn't covered by an existing skill. Walks through confirming scope, identifying source repos, reading the relevant API, drafting the new skill in the the workspace house style, adding any missing wiki content, and registering the skill in the index. Use only when a genuinely new skill is needed — don't invoke for variants that an existing skill already handles.

Bootstrapping a new the workspace skill

When the user wants something this workspace can't already do, your job is to create a new skill that can do it, and to teach the user the underlying API along the way. This file is the protocol.

The output of running this skill is:

  • A new skills/al_<name>.md file authored against skills/_style.md.
  • A symlink at .claude/skills/al_<name>.md pointing to ../../skills/al_<name>.md.
  • An updated skills/README.md entry.
  • Optionally, one or more new wiki pages the new skill links into.
  • A working .py script in scripts/ that demonstrates the new skill on the user's data.

Step 1 — confirm scope with the user

Before reading any code, restate the requested capability and confirm it with the user. Be explicit about what the skill will and won't do. A skill that tries to cover three different lensing tasks is worse than three small skills. If the request spans more than one task, propose a split and ask which to write first.

Also confirm:

  • The data type (imaging, interferometer, point-source, multi-wavelength).
  • The user's depth (newcomer / API-newcomer / returning) — this changes the framing.
  • Whether they have an existing dataset to validate the skill on, or want a synthetic one (in which case the new skill may chain into al_simulate_dataset).

Step 2 — read the house style

Read _style.md end-to-end. The new skill must match its conventions: Orient → Ask → Branch → Combine arc, four properties (science first, encourage reading, conversational, composes), python-first, source citations as <Project>:<path>.

Step 3 — identify source repos

Open ../sources.yaml. Decide which projects the new skill needs to touch. Common patterns:

  • New profile / model object → PyAutoGalaxy + PyAutoLens.
  • New non-linear search / inference feature → PyAutoFit.
  • New data structure (grid, mask, geometry) → PyAutoArray.
  • New configuration option → PyAutoConf + the package whose config you're touching.

If the new skill needs a repo not in sources.yaml, add it there first (with a name, import, role, and git URL) and tell the user what you added.

Step 4 — read inside the source repos

You need real API knowledge, not guesses. Either:

python -c "import <package>, pathlib, inspect; print(pathlib.Path(inspect.getfile(<package>)).parent)"

…and read from the installed location, or clone the repo from its git URL into ./sources/<project>/ (gitignored) and read from there:

mkdir -p sources && git clone <git_url_from_sources_yaml> sources/<project>

Read the relevant module(s). Identify:

  • The class(es) the new skill will instantiate or subclass.
  • Their constructor signatures and key methods.
  • Existing usage patterns in <Project>/test_<package>/ (tests) or autolens_workspace/scripts/ (tutorials).

Take notes; quote signatures in the new skill where they matter.

Step 5 — locate or draft the wiki page

For every concept the new skill teaches, identify the matching wiki page. If one exists, the skill will link to it (don't restate the wiki). If it doesn't exist, draft the wiki page in the same change.

A new wiki page has:

  • Frontmatter listing the sources (project + paths + pinned commit).
  • A short overview of the concept.
  • A table or list of the API surface (classes / functions / configs).
  • Links to ReadTheDocs or other primary references.

See wiki/README.md for the wiki page template.

Step 6 — draft the skill

Save the new skill at skills/al_<task>.md. Follow _style.md:

  • Frontmatter name: matches the filename (without extension).
  • Frontmatter description: is one paragraph; a future agent must be able to decide whether to activate this skill from the description alone.
  • Body uses Orient → Ask → Branch → Combine.
  • Python recipes are full, runnable snippets — not pseudocode.
  • Every code-aware concept cites <Project>:<path> and links to the wiki page.

Step 7 — register the skill

ln -s ../../skills/al_<task>.md .claude/skills/al_<task>.md

Add a one-line entry under the appropriate category in skills/README.md.

Step 8 — verify

Write the Python script the new skill recommends — for the user's actual data, in scripts/. Run it with:

PYAUTO_TEST_MODE=1 NUMBA_CACHE_DIR=/tmp/numba_cache MPLCONFIGDIR=/tmp/matplotlib \
  python scripts/<script>.py

If it doesn't run, fix the skill. Don't ship a skill whose recipe doesn't execute.

When it runs, tell the user:

  • What the new skill is called and where it lives.
  • The path to the generated script.
  • What wiki pages were added or updated.
  • One or two natural next-step skills the user can chain into.

When NOT to bootstrap a new skill

  • The request is a small variant of an existing skill — extend the existing skill instead.
  • The request is essentially "run this existing tutorial script for me" — point the user at the existing script in autolens_workspace and the relevant wiki page, but don't wrap it in a new skill.
  • The request requires source code we genuinely cannot read (closed-source dep, dead link). Surface this to the user and don't fake it.

Agent procedural checklist (slim, for your sanity)

  1. Confirm scope with the user, propose a name.
  2. Read _style.md.
  3. Resolve source repos from sources.yaml; clone or import as needed.
  4. Read the relevant source. Take signature notes.
  5. Read or draft the matching wiki page(s).
  6. Write skills/al_<task>.md.
  7. Add symlink under .claude/skills/; add entry to skills/README.md.
  8. Generate the user-specific script in scripts/. Run it with PYAUTO_TEST_MODE=1.
  9. Report to the user: skill name, script path, new wiki pages, suggested chain.