Agents
Each file in this directory is a versioned prompt for one role in the F(AI²)R writing pipeline. Treat them as source code: edit, diff, review, commit. Do not improvise new prompts in chat without landing them here first.
Role catalogue
| File | Role | Pass |
|---|---|---|
orchestrator.md |
Routes tasks across the other agents and the human author. | both |
scientific-writer.md |
Drafts manuscript prose at section granularity. | author |
source-analyzer.md |
Reads cited material, extracts claims and quotes. | author |
fair-aligner.md |
Checks artefacts against the FAIR principles. | auditor |
provenance-curator.md |
Maintains doc/provenance.ttl; refuses fabricated triples. |
auditor |
layout-scrutinizer.md |
Reviews LaTeX layout, figure placement, typography. | auditor |
readability-reviewer.md |
Reviews prose for clarity, flow, and novelty signal. | auditor |
illustration.md |
Plans and specifies figures (TikZ / SVG / matplotlib). | author |
condenser.md |
Page-budget enforcer. Fires when the compiled PDF exceeds the configured budget (default 10 pages of body). | auditor |
presentation.md |
Slide-deck custodian. Owns slides/pitch-5min.md and slides/conference-30min.md; ensures slide claims are downstream of the paper text. |
author |
research-protocol.md |
Defines literature-search and verification protocols. | both |
Primary artifacts — consistency invariant (binds every agent)
The manuscript (paper/), the PROV-O graph (doc/provenance.ttl),
and the logbook (doc/logbook.md) are primary artifacts and must be
consistent and up to date at all times. No agent may finish a task
that touches one of them without ensuring the other two reflect the
change. Each prompt below restates this rule with a tailored "what I
must keep in sync" clause.
Conventions every prompt follows
- Role, scope, non-goals. Every prompt opens with what the agent is and what it will not do.
- Inputs and outputs are explicit. A prompt names the files it reads and the files it is allowed to write.
- Primary-artifact consistency. Every output bundles (a) the manuscript edit, (b) the proposed PROV-O triples, and (c) the logbook line. If the agent does not produce one of these directly, it names who must, before the session ends.
- Verification ladder. Claims emerge tagged
unverified,needs-research,ai-confirmed,human-confirmed, orsource-vendored. - Refusal is fine. Agents must say "I cannot verify this" rather than invent a citation.
- Provenance hand-off. Every output ends with the triples it expects
provenance-curatorto add todoc/provenance.ttl.