start

A command that turns a plain-language research goal into a clear research contract after inspecting the code repository and discussing key decisions with you.

In plain words
What is it for?
Use it to define what will be changed, how success will be measured, what limits apply, and how the research loop should start.
Why use it?
It prevents an unclear goal or unsuitable evaluation method from wasting an automated research run.

Command

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add commands/fradser/dotclaude/start
Clone the repo
git clone --depth 1 https://github.com/FradSer/dotclaude
Per session 52 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 1,499 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00052 $0.01499
Opus 5 $0.00026 $0.00749
Sonnet 5 $0.00010 $0.00300
Haiku 4.5 $0.00005 $0.00150

Measured yesterday against content hash 6e17692a37fd, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

start scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

autoresearch/commands/start.md · 62 lines

How it starts

The opening of the file, as written. The whole thing — 62 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Autoresearch Start (autonomous)

Turn the user's free-text goal in $ARGUMENTS into a complete research contract by inspecting the repo, then grill the contract with the user (one decision at a time, each with a recommended answer) before launching the loop. A wrong contract — especially a wrong evaluator — wastes the whole overnight run, so the contract is a shared decision, not an inference. Any explicit flag the user passed in $ARGUMENTS (e.g. --edit, --score-cmd) is an OVERRIDE — the user already decided that field; use it verbatim and do not re-infer or re-grill it.

Phase 1: Read the goal

The leading free text of $ARGUMENTS (before any --flag) is the GOAL. Record any override flags the user passed.

Empty goal is a refusal. If $ARGUMENTS is empty or whitespace-only, STOP and ask the user for the goal — a loop with no goal has no contract, and every later decision hangs off it.

Phase 2: Infer the contract from the repo

Inspect the repo (list files; read package.json / Makefile / pyproject.toml / README) and derive recommendations for every field (the grill in Phase 3 turns each into a decision):

  • --edit — the artifact to optimize. If the goal names a file or area, use it. Prefer a SINGLE file when the goal is about one thing — a single file unlocks the tournament escalation.
  • An evaluator — prefer an OBJECTIVE one; a wrong evaluator wastes the whole run:
    1. The goal implies a measurable number and a command prints it → --score-cmd '<cmd>' + --direction min|max.
    2. Else the project has a test/check command (package.json scripts.test/lint/typecheck, a Makefile target, pytest, cargo test) and the goal is "make it work / keep it passing" → --check-cmd '<cmd>' (a pass/fail gate).
    3. Else the goal is qualitative (clarity, readability, prose, design) → --rubric '<criteria distilled from the goal>', ANCHORED by a --check-cmd (a test/build that must keep passing). NEVER a rubric without a --score-cmd or --check-cmd anchor — a judge-only loop reward-hacks (the setup will refuse it).
    4. Combine when it fits (gate + score, or gate + rubric).
  • --objective — a one-line measurable restatement of the goal (what success means).
  • Bounds — default --max-experiments 20, unless the goal implies time ("overnight" → `--max-wall-clock 8h").
  • TAG — a short slug from the goal (optional; defaults to the date). No need to ask — it is internal.

Read the full file on GitHub · 62 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. yesterday First seen · 62 lines · 52 tokens per session scan A 6e17692a37fd

Subscribe to this mod's changes

start is a command published in the GitHub repository FradSer/dotclaude (587 stars, last pushed 20d ago), licensed MIT. It adds 52 tokens to every session and 1,499 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.