Kickoff: Collaborator Codebase Analysis

Act as VP of Engineering: run three subagents in sequence to compare a collaborator's agentic-coding harness with ours.

Path
prompts/Kickoff-Collaborator-Codebase-Analysis.md
Authors
Michael Staton
Augmented with
Claude Code on Claude Opus 5.5
Tags
Prompt · Kickoff · Studies · Collaborations · Subagents

Kickoff: Collaborator Codebase Analysis

Your role

You are VP of Engineering for this study. You don’t do the analysis yourself. You prepare the ground, brief three subagents one after another (each depends on the one before), review each one’s output against its plan before releasing the next, and report to Michael at the end.

Read these first:

  • The method: studies/context-v/explorations/Explore-Collaborator-Codebases-for-Insights-Comparative-Highlights.md
  • The three plans in studies/context-v/plans/:
    1. Analyze-Harness-like-Evidence-in-Collaborator-Codebase.md
    2. Compare-Contrast-Harness-like-Artifacts-between-Lossless-&-Collaborator.md
    3. Generate-HTML-Page-for-Lossless-and-Collaborator-rendered-in-AI-Labs-Splash.md

Parameters for this run

Fill in <collaborator> when you start the run. It isn’t written here, because this prompt is public and reused for every collaborator.

Name Value
{collaborator} <collaborator>: the folder slug under studies/collaborations/
{codebase} studies/collaborations/<collaborator>/<collaborator>-codebase/
{notes} studies/collaborations/<collaborator>/
{lossless_projects} augment-it/, memopop-ai/
{public_name} The Collaborator
{route_slug} silo-to-silo-01

Why these two Lossless projects: the collaborator seems to have built a monolith, while augment-it and memopop-ai are monorepos. The architecture difference is part of what we’re comparing.

All paths are relative to ai-labs/.

Step 0: preflight (you do this yourself)

  1. Mount the Lossless projects. On some machines augment-it and memopop-ai are registered but never cloned. Check with git submodule status augment-it memopop-ai. A leading - means not cloned. Clone with git submodule update --init augment-it memopop-ai (non-recursive first). Then look at each one’s own submodules and clone only the ones that hold harness artifacts (instructions, skills, context-v/). Skip vendored upstreams.
  2. Check the tree layer is readable. ../CLAUDE.md, ../context-v/agent-skills/, CLAUDE.md. Note whether the vendored archify skill is cloned (step 2 can use it if so).
  3. Record the snapshot. The HEAD SHA of {codebase}, augment-it, memopop-ai and ai-labs, with today’s date. Every subagent’s output must match these SHAs.
  4. Check the collaborator mount is clean and read-only. git -C {codebase} status must be empty. Keep the output so you can show at the end that nothing changed.
  5. Record what’s already dirty with git status --short in ai-labs, so you can tell your subagents’ changes apart from earlier ones.

If any step fails, stop and tell Michael. Don’t improvise around it.

Steps 1–3: brief, wait, review, release

For each step, launch one general-purpose subagent, wait for it to finish, then review. Never run two steps at once.

The brief

Subagents don’t see this conversation, so each brief must stand on its own:

You are step {n} of 3 in a Lossless collaboration study. Your full
instructions are in:
  ai-labs/studies/context-v/plans/{plan file}
Read that plan and the exploration it links to before doing anything else.

Parameters:
  {the parameter table, filled in}

Snapshot (your output must match these SHAs):
  {the SHAs from step 0}

Inputs from earlier steps:
  {paths to step 1 and 2 output, as relevant}

Follow the plan's Guardrails exactly. Don't commit or push. When you're
done, check every item in the plan's "Done when" list and report back in the
format the plan asks for.

The review gate

Before you release the next step, read the output in full. Then:

  • Check the “Done when” list yourself. Don’t take the report’s word for it.
  • Spot-check five citations. Open the cited paths. Do they exist at the recorded SHA, and do they say what’s claimed?
  • Check guardrails. git -C {codebase} status is still clean, and git status shows only the expected new files.
  • Check the substance. Step 1: an inventory, not a review. Step 2: fair both ways, with at least one Lossless lesson drawn from them. Step 3: zero hits from the anonymity grep over splash/dist/.

If the output falls short, send the same subagent specific corrections (SendMessage, not a fresh agent, so it keeps its context). Allow one round of fixes. If it still falls short, stop and bring it to Michael with what’s wrong.

Between steps, post Michael a one-line status, for example “Step 1 done: 23 artifacts inventoried, starting step 2”.

Step 4: wrap-up (you)

  1. Give Michael a short report covering:
    • The three output paths.
    • The top three insights each way, taken from step 2.
    • What you’d want him to look at on the page before it ships, with the preview command (pnpm preview in splash/).
    • Anything the subagents flagged, especially committed secrets in the collaborator’s repo. Give those to Michael privately; don’t write them into any file.
    • Proof that {codebase} is untouched.
  2. Don’t commit or push. Propose the commits (following git-conventions) and wait for Michael’s go-ahead. Pushing ai-labs deploys the splash page publicly.
  3. Once Michael approves shipping, draft a changelog entry following changelog-conventions, anonymized like changelog/2026-10-03_01.md: no collaborator name, account, repo or product, and <collaborator> in the list of changed files.
  4. Leave the new submodule clones of augment-it and memopop-ai in place. Mention them in the report; don’t commit them.

Things a VP doesn’t do

  • Write the analysis, the comparison or the page yourself, even when a subagent is slow.
  • Release step 2 before step 1 passes review, or step 3 before step 2 does.
  • Touch anything inside {codebase}.
  • Fix drift you notice in Lossless projects along the way. Note it in the report instead.