Skip to content

Latest commit

 

History

History
144 lines (107 loc) · 5.22 KB

KICKOFF.md

File metadata and controls

144 lines (107 loc) · 5.22 KB

KICKOFF.md — exactly what to paste into Claude Code

Finish SETUP.md first. Then cd ~/Desktop/seds-hackathon && claude.


Message 1 — orientation (paste verbatim)

Read CLAUDE.md, BUILD_PLAN.md and OPTIMIZATION_LOOP.md in this repository in full
before doing anything. Read docs/handbook.pdf too.

Then, without writing any code yet, report back:

1. A one-paragraph summary of the objective and why the three tournament stages
   need different risk postures.
2. The eleven engine rules R1-R11, in your own words, flagging which two are the
   easiest to get wrong in a parser.
3. Anything in those documents that is ambiguous, contradictory, or that you
   believe is factually wrong. Be direct — I would rather find an error now than
   at 4am.
4. The exact git ritual you will follow for every step.

Do not start Step 0 until I confirm.

Read the reply properly. If it misstates a rule, correct it now — that misunderstanding will otherwise be compiled into the resolver.


Message 2 — begin

Confirmed. Begin STEP 0 from BUILD_PLAN.md.

Constraints:
- Go only. Mac M3 (arm64) locally, Render free tier (native Go runtime, not
  Docker) for deployment.
- Bind 0.0.0.0 and read $PORT from the environment.
- recover() middleware so no panic can escape a handler.
- Follow the git ritual: branch step-00-skeleton, commit, push, open a PR with
  gh, and stop. I will verify the deploy before you merge.

Do not proceed past Step 0's acceptance gate.

Message 3 onward — one per step

Step 0's gate passed: [paste evidence — curl output from mobile data, the CLI
game result].

Merge the PR, tag known-good, then begin STEP N from BUILD_PLAN.md.
Stop at its acceptance gate and open a PR.

Never let it run two steps in one go. The gates are the whole point.


After Step 12 — start the loop

Paste Part 4 of OPTIMIZATION_LOOP.md verbatim. It will run until it converges or you call freeze.


Prompts for specific situations

When a step's gate fails:

The gate failed: [exact output].
Do not change the gate. Diagnose the root cause, state it in one sentence,
then propose the smallest fix. Wait for my approval before implementing.

When it wants to build something not in the plan:

That is not in BUILD_PLAN.md. Either justify it against the acceptance gate for
the current step, or defer it to the optimisation loop's priority queue.

When you are short on time:

Time check: [X] hours remain before code freeze. Given BUILD_PLAN.md, tell me
which remaining steps you would cut and why, ranked by expected cost to our
placement. Do not implement anything yet.

Before the practice window:

The unscored practice window opens in 15 minutes. Deploy the known-good tag.
Verify: keep-warm active, URL reachable from outside, opponent logging enabled,
every fatal payload saved to testdata/fixtures/.
During the window we collect data only — no code changes. Confirm you understand.

At feature freeze:

Feature freeze. Deploy the known-good tag and confirm it is live and awake.
From now on: operations and proven rule-defect fixes only. No strategy changes,
no refactors, no "small improvements". If you believe something must change,
state the defect and the evidence first and wait for approval.

Guard rails — watch for these

Claude Code is good at this work but has predictable failure modes. Intervene if you see:

  • Skipping a gate because the code "looks right". The gates exist because reasoning about a snake is unreliable and measuring it is not.
  • git checkout . or git reset --hard. Forbidden. It destroys unrelated work.
  • Hardcoding 500 ms instead of reading game.timeout (R10).
  • Parsing shrinkEveryNTurns at the settings root. It is nested under royale. Fails silently as zero.
  • Merging the body and head occupancy grids (R3). Breaks head-to-heads you would win.
  • Detaching a goroutine for search. Cooperative deadlines only — a detached search starves your concurrent games.
  • cmd/server importing BattlesnakeOfficial/rules. AGPL boundary. CI catches it; do not let it be disabled.
  • Assigning equal-arrival Voronoi cells by snake index. Systematic bias. They are contested.
  • Blending the three stage fitnesses into one number. Hides a bracket regression behind a qualifying gain.
  • Deleting a fixture because it now fails. The fixture is the specification.
  • Changing SEEDS mid-tuning-run. Makes every prior measurement incomparable.

Time-boxed fallback plan

If you fall behind, this is the cut order. Each line below is still a functioning, deployable snake.

If you reach hour... Ship at least
3 Steps 0–3: deployed, exact legal moves, lexicographic fallback. This alone is competitive — a searchless 463-line rule chain reached semi-finals at a comparable event
5 + Step 4 (temporal Voronoi) and Step 5 (arena)
7 + Step 6 (TVAE, CVaR risk)
8.5 + Step 7 (Royale) — do not skip this; two of three stages are Royale
9.5 + Step 8 (resilience) — never skip this
10.5 + Steps 9–11 if green

Steps 3 and 8 are the two you must never cut. Step 3 stops avoidable deaths; Step 8 stops forfeits. Everything else is upside.