Evidence-backed career records · v1.0
Recall reconstructs everything you've demonstrably built — commits, PRs, tickets, metrics — into one evidence model, then spends it on the whole job switch: brag doc, role fit, tailored applications, interview prep, exports. Local-first. Yours.
Engineering retrospective · Maya Ellison
How it works
Scripts do the mechanics — digesting history, merging careers, rendering pages — and the model does only judgment. Sources write evidence.json; renderers read it. Adding either side never touches the other.
One schema-validated contract per repo, merged into a career store that survives clones, dedupes forks, and can exclude or anonymize any employer. Plain JSON on your disk, chmod 600.
Community adapters are one-file PRs. The conformance kit (conform.js) validates any adapter's output against the schema — required fields, legal citation types, honesty rules — and lists every violation in one run.
The job switch, end to end
Every mode runs on the same evidence model. Nothing is retyped, nothing is invented; if a job posting asks for something the record doesn't show, Recall says gap — never a bluff.
Digest → judgment → an interactive, self-contained page and its Markdown mirror. Filed into the career store automatically.
Ranks the 3–5 roles the evidence actually supports today — each with proof to lead with and the one honest gap.
Parses the JD once (cached by content hash), matches strong / partial / gap, and renders bullets, cover letter, and prep — every citation validated.
An append-only tracker: drafted → applied → screen → interview → offer. Illegal jumps rejected; history never rewritten.
Asks the questions this posting will ask, critiques your answers against the record, and tightens the STAR wording until it lands.
Scope, impact, leverage — built only from the highest-confidence measured evidence, every number traced to its source.
A cron-safe refresh keeps evidence up to date for free; the delta render answers "what have I shipped since my last review".
Preview first — every email and link host in the page listed for review — then publish the exact reviewed bytes to an endpoint you choose.
The output
Maya is a synthetic persona, but every pixel is the real renderer: live charts with per-mark tooltips, a contribution heatmap, counted skills with quarterly sparklines, measured signals, and six STAR stories — all from one evidence.json.
Install
Recall runs as a Claude Code skill and reuses the git and MCP connections you already have. No API keys, no accounts, nothing to build.
Installs the skill and its scripts, then ask Claude /recall.
# installs to ~/.claude/skills npx @premdevai/recall # then, in any repo /recall
Install straight from the repo's marketplace manifest.
claude plugin marketplace add \ premdevai/recall claude plugin install recall
Cron keeps the evidence fresh — no model, no tokens.
# weekly, pure script
0 9 * * 1 cd ~/repo && \
node ~/.claude/skills/\
recall/refresh.js
Why it holds up
Every claim links to a commit, PR, ticket, or metric. Skill depth is real merged-PR and commit counts — never a made-up 0–100 score. Requirements without adjacent evidence are called gaps, not stretched.
Self-reported work is marked in every format and never counted as verified — the renderer, the career store, the exporters, and the application kit all enforce it independently of anyone's good intentions.
Local-first: plain JSON on your disk, chmod 600, no sync. The only outbound write in the entire codebase is the opt-in share — and it never fires without a previewed, explicit confirm.
Clarity, confidence, and proof behind every line — reconstructed from the work you've already done.