Skip to content

Add decision-oriented usage guide#34

Open
DhruvParikh1 wants to merge 4 commits into
rjcorwin:mainfrom
DhruvParikh1:docs/usage-guide
Open

Add decision-oriented usage guide#34
DhruvParikh1 wants to merge 4 commits into
rjcorwin:mainfrom
DhruvParikh1:docs/usage-guide

Conversation

@DhruvParikh1

Copy link
Copy Markdown
Contributor

Summary

  • Adds GUIDE.md — a task-oriented guide that helps users pick the right cook command for their situation
  • Structured around escalation levels, a decision flowchart, head-to-head comparisons (xN vs review, vN vs vs, pick vs merge vs compare), real-world recipes, and an operator-order explainer
  • Complements the existing reference docs (SPEC.md, website) by addressing "which command should I use?" rather than "what does each command do?"

Test plan

  • Read through GUIDE.md and verify all examples are accurate against SPEC.md
  • Confirm operator-order diagrams match actual parser behavior
  • Check that recipes cover the most common user scenarios

🤖 Generated with Claude Code

@rjcorwin

Copy link
Copy Markdown
Owner

Suggested changes for GUIDE.md

  1. Mention ralph's default max tasks is 100 — users should know what happens when they omit N
  2. Review cost in the head-to-head table — "1 call (if perfect) to 3N calls" is misleading; it's up to 6 calls (2 per iteration × 3 iterations), not 9
  3. Mention race N as longhand for vN — at least once, so users who encounter it aren't confused
  4. Consider adding a note that cook --help exists for a quick flag reference

@rjcorwin rjcorwin left a comment

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nice addition! See changes requested. Also plz rebase to drop the unrelated commits. Thank you!

Comment thread GUIDE.md Outdated

Runs the entire review loop (work → review → gate → iterate) 3 times. Each round starts fresh work, goes through its own quality gate, and must reach DONE before the next round begins. This gives 3 independent review-gated passes.

**Use this when:** You want multiple rounds of quality-gated work — for instance, a first round for basic implementation, a second round for edge cases, and a third for polish. Each round builds on the last but goes through its own quality check.

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Interesting use case, but these operators don't fit that pattern (yet). For example, sounds like we'd want to define a different review/gate prompt for each xN iteration, but in its current form it would use the same prompts for each iteration. This use case might not need an addition to cook, instead chaining them with sequential commands. ie. `cook "add dark mode" review && cook "improve accessibility of dark mode" review

Comment thread GUIDE.md Outdated
| Mode | When to use |
|---|---|
| `--sandbox agent` (default) | Most of the time. Each agent uses its own built-in sandboxing. No setup required. |
| `--sandbox docker` | When you need network isolation, a reproducible environment, or you're using OpenCode (which requires Docker). Also useful in CI/CD where you want guaranteed clean environments. |

Copy link
Copy Markdown
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Some more reasons to use docker sandbox:

  1. I find it useful to not have to define every command that the agent should and shouldn't run like we have to with agent sandbox solutions.
  2. Your environment variables are not exposed here, like they are in Claude's agent sandbox.

GUIDE.md helps users pick the right command for their situation,
structured around escalation levels, decision flowcharts, head-to-head
comparisons, and real-world recipes rather than reference docs.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@rjcorwin

Copy link
Copy Markdown
Owner

An aside, this guide seems like good context for the cook skill. Adding it into the scale could help steer the agent towards the right configuration. This could also be the basis for a /cook-help skill that instructs the agent to ask the user a series of questions to determine the right config for the job.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants