launch-brief-publisher · my-agent/ overview

{
  "meta": { "founder": "demo-user",
            "problem": "2h/launch turning brief.md
              into internal + investors + twitter" },
  "agent": { "name": "launch-brief-publisher",
             "model": "claude-opus-4-8",
             "system_summary": "reads ./brief.md;
               writes 3 tone-tuned files to
               /mnt/session/outputs/" },
  "outcome": { "rubric_criteria": [ ... 5 rows ],
               "max_iterations": 3 },
  "tools":   { "toolset":
                 "agent_toolset_20260401",
               "enabled": [read,write,edit,glob,grep] }
  ...
}

{
  "name": "launch-brief-publisher",
  "model": "claude-opus-4-8",
  "environment_id": "env_demo_launch_brief_publisher",
  "system": "You are launch-brief-publisher.
    Read ./brief.md and produce three tone-tuned
    files in /mnt/session/outputs/ (internal.md,
    investors.md, twitter.md) quoting the launch
    date, headline number, and CTA verbatim…",
  "tools": [
    { "type": "agent_toolset_20260401" }
  ],
  "mcp_servers": [],
  "resources": [{
    "type": "memory_store",
    "name": "voice-samples",
    "access": "read_write"
  }],
  "skills": []
}

# Outcome — launch-brief-publisher

Definition of done. Per-run grader (embedded
as rubric.content on every session) AND the
held-back regression check for case-02/03.

Every criterion is binary — either it lands
or it doesn't. Split anything that starts
feeling scalar.

1. All three files exist in
   /mnt/session/outputs/
2. Each file quotes brief.md's launch date,
   headline number, CTA verbatim
3. internal.md reads like Slack — ≥1 emoji
4. investors.md is one paragraph, ends with
   an explicit ask
5. twitter.md is a 5–9 tweet thread; tweet 1
   hook, last tweet CTA link
6. No claim without a source pointer or
   an explicit "assumption:" tag

max_iterations: 3 (only knob to stop grind)

You are launch-brief-publisher. Read
./brief.md and produce three release notes
in /mnt/session/outputs/ that quote the
launch date, headline number, and CTA from
brief.md verbatim.

The three files:
  - internal.md   Slack, ≥1 emoji
  - investors.md  1 paragraph, milestone-first
  - twitter.md    5–9 tweet thread

Rules:
  - Never invent dates, dollar amounts,
    headcount, or customers not in brief.md
  - Never post or send — output is markdown
  - Every claim: source line pointing back
    to brief.md OR "assumption:" prefix

Grade against outcome.md. Eval case for
this run lives at ./evals/case-01/

<link rel="stylesheet" href="overview.css">

Live schema of the app:
 • top bar (statepill, model slug,
   agent/session IDs as Console deep links)
 • trigger → worker in env frame → output
   (🤖 agent · 🛠️ tools · 🔌 connectors ·
    🧠 memory · 📄 skills · 🎯 outcome)
 • sessions run-log (fed by
   evals/results-v<N>.json) + evals table
 • v1/v2/v3 version rail

The launch-your-agent skill Edits slots in
place after each iteration — no full
re-Write of this file after Phase 1.

# v1 — Post internal.md into #launches
  How: Slack MCP + vault; toolset
       always_ask; grader adds
       "post confirmed" row.

# v2 — Tone QA pass before returning
  How: author tone-checker custom skill,
       add "tone-verdict = pass" row.

# v3 — Trigger from brief.md commit push
  How: GitHub webhook → POST
       /v1/deployments/<id>/run with
       brief text as initial_events.

# always — re-run held-back evals before
  promoting the deployment.