Skip to content

README.md

Latest commit

 

History

History
91 lines (65 loc) · 5.29 KB

README.md

File metadata and controls

91 lines (65 loc) · 5.29 KB

Example: a real /pre-loop run (Getting Automated content library)

This is a worked example of what the /pre-loop skill produces. It was run on a genuine Getting Automated task and grounded in a real context repo — the business-context repo that tells agents who Getting Automated is, what's authoritative, and how work gets reviewed. That's why the brief it wrote reads like the actual business and not a generic template.

The prompt that kicked it off:

"Keep my public content library valid — I keep finding guides with missing frontmatter and dead related-links."

The output: loop-brief.md — the full contract the loop runs on. Everything below is the run that produced it.

🧪 …and then we actually ran it → run-result.md

On Hunter's real library, the loop took the validator it built (validate_content.py) from 20 errors to 0 — fixing 14 structural issues (frontmatter only, bodies untouched) and refusing to invent the 67 facts it was missing (flagged for review instead). Real before/after, on a branch.

1. What /pre-loop read first (Step 1 — infer before asking)

Before asking a single question, the skill read the repos.

From the context repo (the business):

  • AGENTS.md / 00-start-here.md — Getting Automated is a production-automation business and a free public library; the position is "reliability, not template theater."
  • content-and-assets.md — the public content is the trust surface; the pillars and the lead magnet.
  • systems/source-of-truth.md — "content already published" is authoritative in content-and-assets.md + the s3-content/ folder; owner = Hunter Sneed.
  • company/claims-and-cautions.md — avoid legacy membership/community copy; use confidence labels.
  • workflows/agent-workflow.md — cite sources, review before output, don't invent facts.

From the website repo (the system being looped on):

  • frontend/src/api/contentService.ts, frontend/src/types/blog.ts — the frontmatter schema the site expects.
  • s3-content/guides/*.md, s3-content/videos/*.md — the real, current frontmatter shape.
  • frontend/package.json — the build command (the downstream check).

What it inferred: stack = Python scripts + markdown content + a TS frontend; convention = YAML frontmatter per the frontend types; owner = Hunter; nothing publishes without his review; the production-first voice must not drift.

2. What it asked (the wizard — only the gaps)

It didn't re-ask anything it could infer from the repos. It surfaced five quick choices:

Question Answer
Good fit to loop? Loop it — a real verifier exists once we scaffold a validator
Scope? Frontmatter validity + relatedContent links + index sync. Not prose, claims, or positioning
Run on a branch or in place? A branch (loop/content-integrity); nothing deploys
What needs human approval? Hunter reviews the PR before any merge/deploy; pause on anything touching a claim or price
Verifier — existing or scaffold one? Scaffold one — no content validator exists today

3. Fitness check (Step 5 — when not to loop)

✅ Good loop. It cleared every gate:

  • Real verifier — a content validator we can build and run, so "done" is checkable independent of the agent.
  • Clear source of truth — the frontend schema + the files; ownership is Hunter.
  • Reversible — all git, on a branch, no auto-publish.
  • Reviewed — Hunter approves before anything ships.

If there had been no checkable "done" — say, "make the guides more engaging" — the skill would have stopped here and told you to define a checkable proxy first, instead of writing a brief.

4. What it wrote

  • loop-brief.md — the full contract: architecture foundations, scope/non-goals, the verifier stack, the anti-gaming guards, safe execution, and the health signals.
  • A schema doc (CONTENT_SCHEMA.md, in the website repo) — the inferred frontmatter schema, so the validator and the loop share one source of truth.

5. The kickoff it handed back

Not just an end-state — the full seed, pointed at the verifier stack and guarded against gaming:

/goal "scripts/validate_content.py exits 0 (all s3-content frontmatter valid, every
relatedContent id resolves, index.json matches the files on disk) and the frontend
build passes — fixes touch only frontmatter and relatedContent, never the validator,
the schema, or any post-frontmatter prose"

Plus: read the context repo's content-and-assets.md + source-of-truth.md and the frontend content types first; mirror the frontmatter of a known-good guide (your-ai-agent-needs-a-context-repo.md); work only under s3-content/, scripts/, and CI.

Recurring version: once it's green, the same brief runs as a Routine (Claude Code) or an Automation (Codex) on every content change — so new guides are validated before they can publish, not after the fact.

Want to see the blank version it fills in? → ../../loop-brief-template.md. A second filled example (adding rate limiting to an API) → ../../loop-brief-example.md.