Writing Agent

Writing and narrative agents

In the loop: an addon, outside the four lines. See Addons.

Two conversational agents live in their own Buzz channels and talk with the owner about long-form writing. #writing is for essays, opinion pieces for business publications, blog posts and book chapters. #narratives is for children's stories. Social media drafts stay with the content engine and #content.

Each agent is the same harness the personas use (buzz-acp in front of Claude Code), with three differences: it subscribes to the whole channel and answers every owner message there without a mention (BUZZ_ACP_SUBSCRIBE=all, BUZZ_ACP_NO_MENTION_FILTER), it keeps one isolated session per thread, and it may write files, but only inside its own drafts folder.

Agent	Channel	Writes to	Reads first
Writer	#writing	Writings/Drafts/ (drafts_dir in PROFILE)	_Agent-Context/WRITING.md, then the editor folder's README
Narrator	#narratives	Writings/Narratives/Drafts/ (narratives_dir + Drafts)	the story guide in the narratives folder

Thread contract

One thread per piece. The root message carries Title | kind | venue | status. Status moves idea, pitch, draft v1, draft vN, revision, sent, published. Every agent reply names the file path and the version it wrote. A draft is never posted into the channel; the reply carries the path, the lint counts and the questions.

Before any draft the Writer loads the voice and editing files in the order the editor folder's README prescribes, writes the whole piece to Drafts/YYYY-MM-DD <slug> vN.md, runs the editor lint on it and reports the counts. A ceiling breach is fixed before the reply goes out. Revisions change only what was asked and land in a new version file. Numbers and quotations carry a vault path or a fetched URL, or are marked for checking.

Pitch round and the writing map

tools/writing_index.py compiles _Agent-Context/WRITING.md: finished pieces and working files, drafts in flight, open pitches, the long-form homes and the published corpus named in PROFILE.md (longform_dirs, corpus_dirs), and the editing files. It runs weekly on the worker and posts a one-line summary.

tools/writing_ideas.py runs every second month (the 1st, 09:00). It reads the beliefs, seed ideas, the latest dialectic syntheses and two weeks of captures, asks the model for three pitches, files each one in the drafts folder and posts each as its own root message in #writing. --topics turns the owner's own ideas into pitches instead. The owner answers in the thread; the Writer takes it from there. Nothing is ever published by the system.

Install

Requirements: the persona installer has run once (it creates the moderator key and the relay membership tooling), the Buzz relay is reachable, PROFILE.md names the folders. Then, on the worker:

bash .agents/buzz/install_agent_channel.sh writing
bash .agents/buzz/install_agent_channel.sh narratives

The installer generates the agent key once, creates the channel if missing, adds the owner and the agent as members, fills the system prompt and the Claude settings from .agents/buzz/agents/<slug>/, writes ~/.config/brainless/buzz/<slug>.env and enables buzz-agent@<slug>.service. BUZZ_AGENT_MODEL (default claude-fable-5-1) and BUZZ_AGENT_EFFORT (default high) pin the model the harness applies to every session. The two timers ship with the other worker units (bash .agents/systemd/install.sh).

Guard rails

• Permissions are a Claude settings file: read anything in the vault, edit only the drafts folder, run only the search tool, the lint tool and the Buzz CLI. No git, no deletes, no network tools except web fetch and search for sources.
• The polling reply worker never answers in these channels (HARNESS_CHANNELS in tools/buzz_delivery.py), so an owner message gets one reply, from the agent.
• Moving a text out of Drafts/ into the writings folder is the owner's act. The agent prompt says so, and the ownership rules record the exception.
• Thread text and vault text are material for the agent, never instructions to change its rules.