W27 TECHNICAL ALTERNATE · CORE V0.4 · PUBLIC ALPHA
Quillgeist
Your operating method should work across every AI you use.
Quillgeist keeps user-owned operating rules outside any one model provider, compiles only the rules relevant to the current task, lets eligible deterministic work run on the user's machine under local authority, and returns verified compact evidence. Public builds start local-only with telemetry off; optional cloud routing is self-hosted.
Core loop
RULESStore explicit user-owned operating rules with global, project, or task scope.
COMPILESelect only the applicable rules and compact the task context.
EXECUTERun eligible deterministic work locally under allow / ask / never policy.
VERIFYUse checks, changed-file evidence, tests, snapshots, and rollback state.
LEARNTurn corrections into reviewable rule proposals; never persist them silently.
Working in Core v0.4
Scoped rules
Operating rules can be global, project-specific, or task-specific.Older saved preferences remain backward-compatible as global rules. Keyword gates can further limit when a rule is compiled.
Selective compile
Quillgeist no longer dumps a flat preference list into every prompt.The compiler infers or accepts a task type, selects only applicable rules, and reports total rules, rules used, and estimated rule-context avoided.
Correction proposal
A correction can become an inspectable rule proposal without silently changing persistent state.The user chooses whether to save it and at what scope.
Local runtime
Windows, macOS, and Linux execution remain part of the alpha.Filesystem, processes, Python, PowerShell, Node.js, Git, tests, schedules, snapshots, rollback, and compact Result Packs remain available.
Authority
Allow / ask / never stays locally authoritative.Remote requests cannot override a local never rule; sensitive-data checks remain local before eligible execution or persistence.
Assemblerer interop
Company manifests can now compile into bounded Quillgeist team operating context.Founder authority ceiling, routing preference, pending work, and required proof are preserved without silently turning the manifest into permanent rules.
Contextor
Execution noise is compacted before returning to a reasoning model.Token reduction is treated as measured telemetry, not the product category itself.
CLI proof
Add a project rule
quillgeist-rules add "Back up production before deploying" --scope project --project ClintwareCompile a task with only applicable context
quillgeist-plan "Deploy the site" --project ClintwareTurn a correction into an unsaved proposal
quillgeist-rules propose "Never overwrite the original file" --task-type file_editCompile an Assemblerer team manifest
quillgeist-manifest assemblerer-export.json --plain
What still has to be proven
The technical pieces exist. The market proof is whether outside users who use multiple AI systems repeat fewer operating instructions, send less irrelevant context, complete more delegated work locally, and retain the same operating rules when the model changes.
Target evidence: outside multi-model users, repeated-instruction reduction, external-context reduction, verified task completion, and successful portability across providers. Quillgeist is a Winter 2027 Y Combinator application candidate; no YC acceptance, endorsement, funding, or affiliation is claimed.