phd-ghostwriter-base

Reusable template for a single-collaborator paper-writing scaffold. Clone this folder once per paper.

What this is

A Claude Code roleplay folder that turns the assistant into a senior-researcher writer for one paper. The writer operates in two modes:

• Story mode — reads experiment data, proposes ≥2 framings, pushes back on weak claims, refuses to draft until a story is committed.
• Drafting mode — patient drafter, takes the first-draft pen, waits for line-level feedback, revises against specific corrections.

The mode is signalled by your trigger words (story / 讲故事 → story mode; draft / 写 → drafting mode). Default is drafting.

There is no separate supervisor folder. This scaffold absorbs the story-construction role.

What is preserved from the heavier two-folder pattern

• MANIFESTO + drift modes — 6 invariants and 6 named drift modes that the writer re-anchors to after auto-compaction.
• Claim-trace anti-hallucination ledger — every number in the draft has a row pointing at a source file in inputs/. Numbers from prior drafts are not terminal sources (R-PD).
• Voice samples + feedback log — voice-first drafting. The writer re-reads samples and the accumulated feedback log before each section.
• Append-only session-replay system — decisions.md / journal.md / sessions/ capture decisions, condensed entries, and full minutes.
• Standup opening protocol — every session opens with a 5-line reorientation + status walk.

How to use

1. Clone this folder for your paper:

   cp -r ~/phd-ghostwriter-base ~/phd-ghostwriter-<paper-nickname>
   cd ~/phd-ghostwriter-<paper-nickname>

2. Run the INIT.md checklist (fill placeholders, populate inputs/, set up artifacts/voice-samples/).
3. Start the session: claude.
4. First message: a story-mode trigger if no story is committed yet, otherwise a drafting trigger.

File map

• MANIFESTO.md — 6 invariants, 6 drift modes (highest layer)
• CLAUDE.md — auto-load protocol, two-mode triggers, post-compact recovery, standup
• persona.md — writer identity, 7 voice markers, story-mode + drafting-mode behavior
• rules.md — M1–M3 (mode), R1–R9 (drafting), R-S1–R-S3 (story), R-IN/R-PD (inputs), R13–R17 (memory), R22 (craft skill)
• project.md — section table, story status, deadline (templated)
• decisions.md — append-only writing decisions log
• journal.md — append-only condensed session entries
• todo.md — current open items
• INIT.md — one-shot setup checklist
• sessions/ — detailed per-session minutes
• artifacts/:
  • drafts/ — current LaTeX draft (writer writes here)
  • voice-samples/ — symlinks to your prior writing
  • claim-trace.md — claim → source ledger (anti-hallucination)
  • feedback-log.md — accumulated voice/style rules
  • story-candidates.md — append-only log of proposed and selected framings
• inputs/ — what the writer reads to learn about the work
  • experiments/ — structured results (csv/json/md tables)
  • notes/ — freeform markdown notes
  • pasted/ — transcripts of pasted numbers/logs
  • external-refs.md — pointers to external repos / notebooks
  • prior-drafts/ — existing paper drafts as input
  • README.md — input format guide

Triggers

• story / pitch / frame / framing / 讲故事 / 定方向 / 这篇怎么写 → enter story mode
• draft <section> / 写 <section> / revise <section> / redraft <section> → enter drafting mode
• feedback: ... → line-level review of last draft
• 结束今天 / end session / update memory / save memory → memory update + minutes

Origin

This base was extracted from an earlier two-folder pattern, where a supervisor folder handled story/scope/push-back and a ghost folder handled drafting only. The base is a single-folder collapse of that pattern: the writer absorbs the story-construction role and uses mode discipline instead of folder discipline to keep "push back" and "draft patiently" from blending.