amp-prompts

Tooling that extracts the system prompts baked into the @sourcegraph/amp CLI bundle and writes human-readable Markdown files under prompts/ and skills/.

The Amp bundle ships as a single minified dist/main.js. Each prompt is a template literal that often takes feature-flag arguments (e.g. enableOracle, enableDiagnostics) and references other minified identifiers like R4, V6, AA. This repo:

1. Pins @sourcegraph/amp as a normal dependency so Dependabot can open a PR each time a new bundle is published.
2. Parses the installed bundle with @babel/parser, locates every arrow function or template literal whose body looks like a prompt, and evaluates it inside a node:vm sandbox.
3. The sandbox's globals are a Proxy whose backing map is built from every name = "value" assignment in the bundle, so a reference to R4 resolves to the string "finder", V6 to "Read", and so on. Any identifier we can't resolve is filled in with a permissive proxy that coerces to "" and is callable, so methods like dirs.map(...) don't throw.
4. Each prompt-builder function is invoked with all boolean flags forced to true, so optional sections (oracle, diagnostics, check mode) are always included.

Quick start

bun install
bun run extract

The script writes:

• prompts/<name>.md — one file per prompt. The filename is derived from Amp's own basePromptType switch when the bundle exposes one, then from a curated substring lookup table (see KNOWN_PROMPTS in extract-amp-prompts.mjs), then from heuristics that read the first "You are …" line. This keeps public Amp names from the Owner's Manual and models page (smart, deep, rush, review, search, oracle, librarian) when a prompt can be tied back to Amp's mode/subagent wiring. Trivial prompts (single sentence, no real content) are dropped, as are exact-text duplicates.
• subagents/<name>.md — one file per extracted subagent prompt. These are specialized secondary-agent surfaces such as review, search, oracle, and librarian.
• skills/<name>.md — one file per extracted skill prompt. Skill-like prompts are separated from ordinary prompts so the generated catalog is easier to scan.
• the generated catalog below — an index linking to every generated file, with the bundle version it was extracted from.

Catalog

Source: node_modules/@sourcegraph/amp/dist/main.js Package: @sourcegraph/amp@0.0.1780827515-ga4daec

Notes:

• Extracted by parsing the bundle with @babel/parser, locating prompt-producing arrow functions and template literals, then evaluating each in a node:vm sandbox.
• Free identifiers (e.g. R4, V6) are resolved through a Proxy whose backing map was built from name = "value" assignments in the bundle.
• Boolean feature-flag parameters (oracle/diagnostics/check-mode etc.) are forced true so all optional sections are included.
• Main and utility prompts are written under prompts/; subagent prompts under subagents/; skill prompts under skills/.

Prompts

• prompt — line 744
• subagent-summary — line 1527
• thread-reader — line 2473
• ai-assistant — line 2588

Subagents

• review — line 1730

Skills

• code-review-skill — line 1040

Bumping Amp

@sourcegraph/amp is pinned exactly in package.json. The workflow at .github/workflows/update-amp.yml runs daily, can also be triggered manually, asks the npm registry for the latest published version with bun pm view, updates the exact pin with bun add, runs bun run extract, and opens a PR labelled amp-bump.

The workflow at .github/workflows/regenerate-prompts.yml also listens for amp-bump PRs that touch package.json / bun.lock, runs bun run extract, and pushes any regenerated prompt files back onto the same PR branch. This keeps manual amp-bump PRs and workflow-created amp-bump PRs consistent. It also enables auto-merge for amp-bump PRs opened by github-actions[bot] or hearnadam.

Layout

extract-amp-prompts.mjs       # the extractor
package.json                  # pins @sourcegraph/amp
.github/workflows/update-amp.yml
.github/workflows/regenerate-prompts.yml
prompts/                      # generated; refreshed by the workflow
subagents/                    # generated subagent prompts
skills/                       # generated skill prompts
README.md                     # includes the generated catalog

Caveats

• Heuristic naming will sometimes pick a wordy name for a brand-new prompt Amp ships. When that happens, add a substring entry to KNOWN_PROMPTS in the extractor and re-run.
• A prompt whose body is shorter than ~200 characters is dropped on the assumption that it's a trivial guard message rather than a real prompt. Adjust MIN_BODY_CHARS if that turns out to filter too aggressively.
• Free identifiers that map to functions (rather than strings) fall back to the permissive proxy. If a future Amp release hits a code path that actually needs the real function, the extractor will silently render "" where that call would have produced text.