Skip to content

Latest commit

 

History

History
41 lines (24 loc) · 3.4 KB

WORKFLOW.md

File metadata and controls

41 lines (24 loc) · 3.4 KB

Some of these docs were written with Claude Code in mind, e.g. they reference tasks and subagents.

But for the most part they should work fairly well in other contexts (e.g. Cursor).

My workflow for starting a new epic is usually something like:

  • Switch to best model - I usually use Claude Opus 4 where I want the most brains, e.g. upfront thinking & planning (though I'm not certain it's really that much smarter than Sonnet)

    • Claude Code: /model opus
    • Cursor: Switch your model to Claude Sonnet 4 in the model selector

  • We want to build X. Here's some background, desired features, concerns, etc. Be in @instructions/SOUNDING_BOARD_MODE.md

  • Discuss. This step takes the longest, answering the model's questions, considering various options & tradeoffs, etc.

  • If there's a new software library or specialist topic involved, I might say "Follow instructions in @instructions/WRITE_DEEP_DIVE_AS_DOC.md for topic X. That way, I'll have a new docs/SOFTWARE_LIBRARY_X.md that we can continually refer back to, containing up-to-date snippets and best practices from the web.

  • Create a new planning doc for this, following instructions in @instructions/WRITE_PLANNING_DOC.md. Read that, check I'm happy with it, discuss/manually edit as needed. This is the key step. Because it has all the context from the deep dive and our conversation, the planning document is usually pretty rich.

  • I occasionally Run @instructions/CRITIQUE_OF_PLANNING_DOC.md in Cursor with o3, and then feed that critique back to Claude to see if it wants to update its plan. (In practice, I mostly just rely on Claude, and only rope in o3 if we're doing something really tricky, or if we get struck.)

  • Clear context - Clear the context window, adding a nice summary of what has been discussed before

    • Claude Code: /compact
    • Cursor: Start a new chat (there's no equivalent to /compact in Cursor, but fortunately you can just reference the planning doc)

  • Switch to implementation model - I might switch over to Sonnet if I think the implementation part is straightforward. (Even with the more expensive Anthropic Max Plan, I hit the rate limits for Opus sometimes).

    • Claude Code: /model sonnet
    • Cursor: Switch your model to Claude Sonnet 4 (or Gemini 2.5 is great too) in the model selector

  • Run @instructions/DO_PLANNING_DOC.md for [planning doc]. Make a cup of tea. I have the Claude permissions mostly in YOLO mode, but it can't commit. The model will do a single stage (with lots of sub-actions), and then stop.

  • It'll pause at the end of the stage, often waiting for approval on a commit message. Read the summary, do some manual testing, perhaps also Run @instructions/DEBRIEF_PROGRESS.md.

  • Continue iteration - Clear context as above, then:

    • Do next stage of planning doc, as per @instructions/DO_PLANNING_DOC.md

  • Housekeeping - Every so often:

    • Run @instructions/UPDATE_HOUSEKEEPING_DOCUMENTATION.md.

    • Run @instructions/UPDATE_CLAUDE_INSTRUCTIONS.md. I think it's probably important that CLAUDE.md (or some equivalent Cursor rules) includes important stuff, e.g. a summary of instructions/CODING_PRINCIPLES.md, project-specific coding guidelines, and reference/DOCUMENTATION_ORGANISATION.md). Then the prompts can be very short, and you can trust that the agent will find the right bit of the code reliably and without wasting too much context.