DraftLoop turns candidate-owned sources into a job-specific CV with local-first grounding, independent AI critique, and final human control.
It is designed for candidates who want useful drafting assistance without giving up traceability or control of their source material. Claims remain connected to candidate-provided evidence, provider and model identities are visible, and the candidate decides what to approve and export.
Current maturity: DraftLoop
v0.8.0-alpha.1is a released Usable CV MVP alpha. Its bounded capabilities are integrated and deterministically validated, but no representative consented outcome has been recorded. The roadmap and current status record its evidence and remaining gaps. This project is not production-ready.
The workflow keeps the job description and candidate sources in a local workspace, then applies an evidence-grounded evaluator–optimizer workflow: an author generates, an independent critic evaluates against a rubric, and accepted feedback drives bounded revision. The default cross-company pairing is Anthropic as author and OpenAI as critic. ADR 0003 records DraftLoop's adaptation and controls.
flowchart LR
subgraph Local["Local workspace"]
Inputs["Approved job requirements<br/>+ candidate sources"]
Gate["Visible provider-transmission<br/>approval"]
Author["Grounded author"]
Critic["Independent evaluator / critic<br/>+ bounded revision"]
Human["Human review<br/>and approval"]
Export["Local CV export"]
Inputs --> Gate --> Author --> Critic --> Human --> Export
end
Anthropic["Anthropic<br/>author"]
OpenAI["OpenAI<br/>critic"]
Author <-->|"approved context"| Anthropic
Critic <-->|"approved draft and evidence"| OpenAI
The loop is an assistant, not an authority. DraftLoop does not independently verify a career, contact past employers, replace interviews, or turn a candidate's source material into permission to invent facts.
Download the newest release compatible with your platform
and its SHA256SUMS file. Desktop packages are distributed as platform-specific
ZIP archives, including Windows x64. The releases page is authoritative for the
current platform set and release limitations.
To verify one download, replace <downloaded-archive>.zip below with the ZIP
you selected and compare that single digest with its matching line in
SHA256SUMS:
# Linux
sha256sum "./<downloaded-archive>.zip"
# macOS
shasum -a 256 "./<downloaded-archive>.zip"On Windows, run (Get-FileHash .\your-download.zip -Algorithm SHA256).Hash in
PowerShell and compare that output with the matching SHA256SUMS entry. Extract
the matching ZIP and launch the desktop executable from the extracted folder.
These are v0.8 alpha packages, not signed installers or
dependable real-application tooling; no representative consented outcome has
been recorded, and signing and automatic updates remain ahead. The CLI is a
separate, source-only interface and has no standalone installer.
Use Node.js 24.5.0 and pnpm 10.18.3:
pnpm install --frozen-lockfileStart the desktop shell and choose Try demo workspace to exercise the deterministic fixture workflow:
pnpm --filter @draft-loop/desktop startFixture mode is offline and uses no provider spend. To inspect the source-only CLI:
pnpm --filter @draft-loop/cli start --helpEvery command group — opportunity briefs, candidate profiles, writing policies, and candidate knowledge bases — is documented with runnable examples in the CLI and desktop operations reference.
Live use requires an explicit provider-transmission approval in the workspace, configured provider credentials, and may incur provider cost. Keep real candidate material out of the repository.
For the normal quality gate, run:
pnpm format:check
pnpm lint
pnpm typecheck
pnpm test
pnpm validate| Area | Technology or boundary |
|---|---|
| Runtime | TypeScript, Node.js 24.5.0, pnpm 10.18.3 |
| User interfaces | React 19, Vite, Electron 43; source-only Commander CLI |
| Product core | Framework-free domain contracts, Zod schemas, orchestrator ports |
| Providers | Explicit Anthropic and OpenAI SDK adapters |
| Local data and output | SQLite via Drizzle ORM; Markdown, PDF, and DOCX exports |
| Quality | Biome, ESLint, Markdownlint, Vitest, GitHub Actions |
flowchart LR
UI["CLI / Desktop"] --> Core["Shared application and core<br/>evaluator–optimizer workflow"]
Core --> Local["Local SQLite<br/>run history + exports"]
Core --> Adapters["Provider adapters"]
Adapters --> Anthropic["Approved Anthropic<br/>author route"]
Adapters --> OpenAI["Approved OpenAI<br/>critic route"]
CKB["Portable CKB component<br/>(basic controls + binding)"] --> Core
The portable Candidate Knowledge Base (CKB) component can store approved local source versions. CLI and desktop adapters can create, open, list, and inspect stores; workspaces can also bind explicit store/base snapshots with drift checks. Retrieval does not yet consume those snapshots, so the existing workspace evidence path remains authoritative for application runs.
- Source material, run history, and exports are local by default.
- A provider receives only context covered by an explicit user approval; the workspace shows the provider, model, transmission scope, and retention choice.
- Independent review is a product constraint: the default author and critic use different provider companies, and their identities are recorded.
- DraftLoop prepares local artifacts. It does not submit applications, publish documents, send messages, or perform uncontrolled web research on a user's behalf.
- Roadmap and current status · Release history · Stage evidence
- Architecture · Architecture decision records
- CLI and desktop operations reference
- Privacy and evaluation · Threat model
- Contributing · Releasing
Human approval is mandatory before an artifact is exported. DraftLoop can help prepare a CV, but the candidate remains responsible for factual review, final approval, and every action outside the local workspace.