mirror of
https://github.com/prdlk/cv.git
synced 2026-08-02 17:31:41 +00:00
96 lines
4.4 KiB
Markdown
96 lines
4.4 KiB
Markdown
---
|
|
description: Read every URL referenced in docs/experience and docs/projects and enrich entries with verifiable material
|
|
---
|
|
|
|
# Discover Experience
|
|
|
|
Enrich the canonical resume sources under `docs/` by mining the URLs each entry
|
|
already references — repos, live sites, press, docs, package registries, talks.
|
|
The goal is to capture verifiable facts (metrics, scope, adoption signals,
|
|
tech-stack evidence) that the YAML doesn't state yet.
|
|
|
|
Scope: $ARGUMENTS — if empty, process all of `docs/experience/` and
|
|
`docs/projects/`. If a directory, glob, or entry name is given, restrict to it.
|
|
|
|
## Phase 0 — Inventory
|
|
|
|
1. `search` for URL-bearing fields across the scope: `url`, `repo`, `links`,
|
|
`demo`, `press`, `website`, `github`, or any bare `https://` value.
|
|
2. Build a per-entry worklist: entry → [urls]. Entries with zero URLs get
|
|
flagged in the final report as "no external evidence available" — do not
|
|
invent sources for them and do not web_search around them unless the entry
|
|
names a public artifact (a company, a published app, a talk).
|
|
3. Deduplicate URLs shared across entries; fetch once, attribute to all.
|
|
|
|
## Phase 1 — Fetch (fan out)
|
|
|
|
Use `read` directly on every URL — GitHub repos, PRs, arxiv, registries, and
|
|
docs pages all come back as structured markdown with anchors intact. For GitHub
|
|
repos, prefer the scheme forms where useful (`issue://`, `pr://`) and read the
|
|
README, release notes, and repo metadata (stars, language breakdown, last
|
|
commit) rather than crawling the tree.
|
|
|
|
If the scope spans more than ~5 entries, fan out with `task`: one subagent per
|
|
entry, each yielding a schema-shaped result:
|
|
|
|
```
|
|
entry_id, urls_read[],
|
|
facts[]: { claim, source_url, anchor, confidence: high|medium|low },
|
|
contradictions[]: { yaml_says, source_says, source_url },
|
|
dead_links[]
|
|
```
|
|
|
|
Per-fact rules for extraction:
|
|
- **Verifiable only.** A fact must be literally present at the source — star
|
|
counts, npm downloads, contributors, changelog entries, press quotes, listed
|
|
features, launch dates, named integrations. No inference from vibes.
|
|
- Prefer facts a resume can use: scale, adoption, recency, named technologies,
|
|
third-party validation (press, citations, forks).
|
|
- Marketing copy on my own sites is weak evidence — mark it `confidence: low`
|
|
and prefer registry/repo/press corroboration.
|
|
- Note anything at a URL that contradicts the current YAML (dates, stack,
|
|
scope, status). Contradictions are findings, not silent fixes.
|
|
|
|
## Phase 2 — Enrich
|
|
|
|
For each entry, merge subagent findings and propose a YAML diff:
|
|
|
|
- Append to (or create) an `evidence:` list: `{ fact, source, retrieved: <date> }`
|
|
- Strengthen existing bullets ONLY when a fetched fact directly supports the
|
|
upgrade (e.g., bullet says "open-source CLI" and the repo shows 1.2k stars →
|
|
bullet may carry the number, with the evidence pointer)
|
|
- Add missing structural fields the source confirms: `stack`, `status`
|
|
(active/archived per repo activity), `launched`, `press`
|
|
- Record `dead_links:` so I can fix or drop them
|
|
- NEVER touch employers, titles, dates, or claims of ownership based on
|
|
fetched content — external sources verify artifacts, not my role in them.
|
|
If a source implies a role discrepancy, list it under contradictions instead.
|
|
|
|
Show me the diff per entry via `ask` (apply / skip / edit). Apply approved
|
|
diffs with `edit`. Do not batch-apply without approval.
|
|
|
|
## Phase 3 — Gap search (optional, ask first)
|
|
|
|
After URL mining, ask whether to run `web_search` for coverage the YAML doesn't
|
|
link: press mentions, conference talks, podcast appearances, aggregator
|
|
listings tied to entry names (company names, product names, my name). Same
|
|
extraction rules; anything found enters as `confidence: medium` at best until
|
|
I confirm it's actually about me/my work — common-name collisions are real.
|
|
|
|
## Phase 4 — Report
|
|
|
|
- Entries enriched vs. skipped, fact count by confidence level
|
|
- All contradictions found (these matter more than the enrichments)
|
|
- Dead links list
|
|
- Entries with no external evidence — candidates for the interview command
|
|
instead, since their content can only improve through my own answers
|
|
|
|
## Hard rules
|
|
|
|
- Every added fact carries `source` + `retrieved` — unattributed facts are
|
|
banned from docs/.
|
|
- Enrichment never rewrites bullet prose style; it adds facts and evidence.
|
|
Wording polish belongs to the resume-optimize pass.
|
|
- If a fetch fails, record it and move on. Never substitute memory of what a
|
|
page "probably says."
|