zacheryglass/.claude/generate-goal

Generate Goal

Turn the accumulated understanding in the current conversation into one ready-to-run /goal prompt. The user has been discussing what to work on next and wants that context distilled into a directive they can paste into /goal (or that Claude can run) without summarizing it themselves.

Hard rules

• Never write files. Output the prompt to chat in a single fenced code block. No .goal file, no scratch file — nothing on disk.
• The prompt must be self-contained. A /goal run is a fresh autonomous agent that does not share this conversation's memory. Bake in every load-bearing fact it needs (paths, commands, knob names, current state, decisions). Assume it starts cold.
• End with a measurable completion condition. /goal runs turn after turn until a small evaluator model confirms the condition holds. The evaluator reads the transcript; it does not run commands. So the condition must be provable from what the agent surfaces (files exist and are committed, a test result is reported, a stated number/verdict appears).
• Stay under the 4000-character limit. /goal rejects an argument (the text after /goal ) longer than 4000 chars. Always verify with the script in step 4 before outputting — do not eyeball it.

Process

1. Identify the objective. Pin down the one thing the conversation converged on as the next work. If it is genuinely ambiguous, or several distinct goals are in play, ask one focused clarifying question before generating — otherwise proceed.
2. Mine the conversation for what a cold agent needs. Pull the specifics, not a vibe: decisions made (and the why when it guides tradeoffs), exact file paths / commands / flags / knob names / URLs surfaced, current state (what is done, in-flight, or blocked), constraints ("don't touch X", "keep pins green"), and gotchas already learned. Prefer concrete references over prose.
3. Compose the prompt (start the text with /goal), in this order:

• Directive — what to accomplish, stated so it stands alone. Do not preface it with "You are an autonomous agent" or any restatement that this is Claude Code / an agent / running in a repo — the harness already establishes all of that; jump straight into the task.
• Context — the minimal essential facts from step 2 (paths, commands, current state). Enough to start cold, no conversation color.
• Constraints — what must not change or break on the way there.
• COMPLETION — one measurable end state + how to prove it (an evaluator-checkable clause). Never include a turn cap or "stop after N turns" — the goal loop has no turn limit and must not be told to stop early. The completion condition itself is the only stopping signal.

1. Tighten and length-check. Every sentence earns its place. Cut recap, narration, and boilerplate framing; keep decisions and specifics. Succinct but complete. Then verify the length is under 4000 chars by piping the composed prompt through the checker — do not skip this, the limit is enforced and rejects overlong prompts:

`` printf '%s' "<draft>" | python3 "${CLAUDE_SKILL_DIR}/scripts/check_goal_length.py" ``

(python on Windows). Always pipe via stdin — never write the draft to a file, not even a scratch file. If it reports FAIL, tighten and re-check until PASS before outputting.

1. Output the final prompt in one fenced code block, then offer to adjust scope or constraints. Do not save it.

What makes the completion condition hold up

Weak (unverifiable / open-ended): improve the backtest realism, make the tests better.

Strong (one end state + stated check):

• COMPLETION: bin/backtest builds, make test reports 4/4 byte-identical pins, and docs/state/BACKTEST_REALISM.md has a new changelog row with the OFF-vs-ON per-market PnL. No change to canonical defaults.

Example

Conversation converged on: add a default-off env knob, prove it is pins-neutral, and measure it.

Generated output:

/goal Read CLAUDE.md first. TASK: add env flag FILL_MODEL_TAKER_PERLEVEL_FEE (default 0) in src/trading/execution/exchange_sim.c so that, when ON, the taker fee is summed per consumed book level instead of computed once on the VWAP; keep taker_matcher.c pure. CONSTRAINTS: default OFF must be byte-identical — `make test` green and check_bt_pins = OK 4 byte-identical with the flag OFF; if pins break the flag is leaking, STOP and fix. VALIDATION: add tests/test_taker_realism.c::test_perlevel_fee_vs_vwap_fee proving the per-level sum equals the sum of fees_v2_taker(shares_i, px_i). COMPLETION: flag implemented default-OFF, make test green, pins 4/4 byte-identical, the new unit test present and passing per the reported test output, changes committed.

Then offer: adjust scope or constraints.