Ghostwrite
Write one bounded artifact in the voice of the human or team responsible for it. The skill turns supplied facts or validated research and a clear objective into useful public or internal prose without exposing the machinery that produced the evidence or inventing facts to make the draft feel complete.
This is the reusable authorship primitive behind briefs, release notes, trust reports, maintainer updates, newsletters, and social drafts. It can also package an accepted draft for a channel and prepare a provider-neutral handoff. It never publishes, posts, sends, or claims provider acknowledgement.
Writing standard
- Lead with the reader's problem, decision, or next action—not the evidence collection process.
- Turn evidence into concrete claims and examples; do not dump receipts, issue threads, graph traces, or machine packets into the body.
- Match the project's vocabulary and supplied writing context instead of defaulting to generic AI, launch, preview, or adoption language.
- Write as the responsible maintainer or operator. The surfaced artifact should not call itself agent-generated or narrate how a model worked.
- Prefer one sharp, useful artifact over several thin sections.
- Narrow the claim or return
needs_more_evidencerather than filling a gap with plausible prose.
Runners and chain boundaries
draft composes admit, one writing turn and verify. admit binds a ready
research packet or the exact supplied facts by its native digest and optionally admits
digest-bound brand-voice and taste-profile context. Only admitted claims and
source digests may enter the draft. verify checks every material claim and
every context binding before releasing it, and withholds anything it cannot
bind. A skill that owns its own writing judgment composes admit and verify
around its own writing turn, so its draft keeps the same guarantees.
package converts a validated draft bound to its evidence into a deterministic
channel payload, bound to its digest, in not_sent state; a draft without an
evidence binding, or with a claim citing a source it is not bound to, is held.
handoff binds that payload's digest to the exact target and records whether
delivery crosses outside the team, whether it then needs the delivering lane's
approval, and the next actor. Neither runner grants external authority.
Delivery belongs to message-send, a publication skill, or the relevant provider
connector.
Inputs and result
objective,audience, andchanneldefine what the artifact must achieve.- Supply either
evidence_pack(a ready validated research packet) orfacts(the factual statements actually supplied for this task). Never supply both. Facts do not require a research run or a model-authored packet/digest. Native admission binds their exact bytes; they remain operator-supplied assertions, not independent provider verification or approval to publish. - Optional writing-context packets apply voice or taste. A packet is admitted
only when its declared
applicabilitynames the requestedaudienceexactly and lists the requestedchannelamong its brandchannelsor tastesurfaces. A packet outside that scope, or one that declares none, returnsneeds_more_evidenceso the right context can be supplied. Each admitted packet's exact digest, the audience and channel it was admitted for, and only the rules actually used are recorded. - Apply the admitted context's structured
profileandapplicability, not just its evidence-rule summaries. Brand vocabulary, cadence, claim limits and channel adjustments matter; taste principles, examples and dislikes matter. verifytakes the evidence contextadmitreturned and the candidate draft written from it. The admitted context, not the candidate, is the authority for source digests and context rules.packagetakes arunx/ghostwrite.draft.v1packet;handofftakes therunx/ghostwrite.publish.v1packetpackagereturned, its channel, the exact target, andboundary_kind(internalorexternal).
The draft result contains the content brief, reader-facing body with claim
references, review checklist, distribution notes, and writing-context bindings.
Packaging returns the channel payload, its evidence binding and digest, and
not_sent state. Handoff returns the target, boundary and next actor, without
delivery evidence.
Stop conditions
- Withhold the draft when material claims lack admitted evidence.
- Reject unknown source or writing-context digests and any context rule that was not present in the supplied packet.
- Withhold the draft when a supplied writing context does not apply to the requested audience and channel.
- Do not let voice guidance turn an unsupported factual claim into a safe one.
- Do not mark a package published, delivered, acknowledged, or approved merely because its local formatting succeeded.
- Keep confidential source content and secrets out of public bodies and handoff metadata.
Example
A research packet supports three claims about a release. A brand-voice packet
asks for direct, evidence-led language. draft can write a concise release note
using those claims and record the voice packet digest. package can create the
newsletter payload. handoff can point to the email provider lane. None of those
steps means the newsletter was approved or sent.
Agent task contract
ghostwrite-draft
Write only from the supplied evidence context. Return the content brief, draft,
review checklist, distribution notes, and exact context bindings. Copy the brief's
objective, audience and channel exactly from your inputs; rewording any of them
withholds the draft. Every material claim must cite admitted source digests. When writing contexts are present,
record each exact packet digest and only rules actually applied. Return
needs_more_evidence when the requested artifact cannot be supported. Do not
add publication state, provider claims, or facts absent from the evidence.
Respect each context's applicability, structured profile and stop conditions.
Profile examples guide style; they are not evidence for new factual claims.
Writing from supplied facts
For a short update or explanation, provide the actual factual statements as facts plus objective, audience and channel. Admission selects this source and binds it natively. The writing turn uses its evidence_context.evidence_log and source_digests exactly as for researched evidence. Do not invent a research packet, source URL or external verification to make the request fit a protocol. The finished draft's distribution_notes.source_assurance is bound by the verifier to operator_supplied_facts or validated_research_packet, rather than chosen by the writer. A binding check establishes what supports a claim; it is not a new independent check that the supplied fact is true.
Composing skills can pass facts into admit and keep their own one writing turn between admit and verify. Researched callers retain the original packet and its exact digest. Existing evidence_pack callers need no conversion, and package and handoff keep the same input/output contracts and unsent delivery posture.
Context applicability judgment
Audience and channel/surface descriptions in a writing profile are prose, not provider authority IDs. Admission preserves their exact source text and target scope. In the existing writing turn, judge each admitted context against the actual task, including its boundaries, expiry conditions and stop conditions. Equivalent wording does not require rebuilding the profile; a genuinely different or uncertain use must be withheld. Missing declared source scope still blocks admission.
Each context_binding.applied entry must contain the admitted packet_digest,
its selected applied_rules, and scope_review: {decision, reason}. Decision is
applies, out_of_scope or uncertain. Explain the actual scope relationship
and any limits. This is recorded agent judgment, not independent native proof
that the prose follows a style. A missing, uncertain or negative review withholds
the draft and preserves the reason in review_checklist. No extra agent turn
or provider permission is added.