cv-achievements

CV Achievements

Turn PR-activity JSON into a handful of CV achievement bullets (3–6 for most roles, up to 7 for a long multi-year tenure with many distinct themes, scaled to the role) for one job entry. Given a directory, this means the authored PRs plus one reviewed-voice collaboration bullet — one combined list, printed and saved to a timestamped results-<stamp>.md beside the input.

Input contract

The user may supply a path produced by fetch-prs.sh — either ONE JSON file or the <out>/<org>/<author>/ directory that holds prs-authored.json and/or prs-reviewed.json. Its shape:

{ "generated_at": "...", "mode": "authored-all|reviewed",
  "author": "...", "org": "...",
  "incomplete": false, "incomplete_reasons": [],
  "pr_count": 0, "prs": [ ... ] }

When given a path, this skill analyzes existing data without fetching. When no path is given, see Step 0 — it fetches the data for you via the bundled script.

If the path is a directory (or both files exist): process both files in one pass. prs-authored.json is the spine — ownership voice is the stronger CV claim, so it produces the bulk of the bullets. prs-reviewed.json then contributes exactly one distinct collaboration/gatekeeping bullet appended to the same list (see the reconciliation note in Step 3). Analyze each file in its own mode (Step 1) — never blend their themes into one bullet — but deliver a single combined list, not two separate sets the user has to stitch together.

If only one file is present, process that one and skip the other voice.

When incomplete is true, read incomplete_reasons and calibrate the warning to the reason — do not blanket-warn (see the table below). Always print the reasons so the user can judge.

Reason pattern	Impact on bullets	What to tell the user
file list truncated at 100 files / commit list truncated	Low — pr_count, titles, labels intact; theme detection unaffected. Only the few huge PRs understate breadth.	One line: those N PRs are large; breadth is slightly undercounted. Proceed.
result cap / search-window / query truncation (whole PRs missing)	High — the dataset is genuinely partial.	Tell the user up front, suggest a narrower re-fetch before trusting counts; don't present bullets as a full account.

Step 0 — Acquire data

This skill can now fetch its own data. The bundled fetch-prs.sh lives beside this file; reference it as "$CLAUDE_SKILL_DIR/fetch-prs.sh" so it resolves regardless of the current working directory or where the plugin is installed ($CLAUDE_SKILL_DIR is injected by the Claude Code harness at skill invocation time and points to the directory containing this SKILL.md file).

Path A — the user gave an explicit path (a JSON file or a <out>/<org>/<author>/ directory): use it as-is and skip fetching. This preserves the original input contract.

Path B — no path was given: acquire the data, then proceed.

1. Gather inputs:

   • --author (GitHub username) — required, ask if missing.
   • --org (GitHub org/owner) — required, ask if missing.
   • Mode defaults to BOTH datasets (omit --mode). Output dir defaults to .cv-data (relative to the current working directory).

2. Resolve and show the target directory: ./.cv-data/<org>/<author>/. Print the absolute path it resolves to (run pwd if needed) so the user sees exactly where data will be written — their workspace, never the plugin cache. Let them override the location before any write.

3. Probe existing data per file in that directory — check prs-authored.json and prs-reviewed.json independently:

   • Neither present → both will be fetched.
   • A file is present → read its generated_at and report the age (e.g. "authored: fetched 12 days ago"). Read its incomplete flag; if true, also report incomplete_reasons so the user can judge.
   • Only one present → offer to fetch just the missing one.

   Then ask the user how to proceed: reuse as-is / re-fetch (overwrite) / fetch only the missing dataset. Suggest a default from age and the incomplete flag — stale (e.g. weeks old) or incomplete data leans toward re-fetch; fresh complete data leans toward reuse.

4. Preflight — only if a fetch will actually run. Verify prerequisites and STOP with the exact fix if any is missing (never auto-install):

   command -v gh   >/dev/null || { echo "Install GitHub CLI: https://cli.github.com"; exit 1; }
   gh auth status  >/dev/null 2>&1 || { echo "Authenticate: run  gh auth login"; exit 1; }
   command -v jq   >/dev/null || { echo "Install jq: https://jqlang.github.io/jq/"; exit 1; }
   

   If any check fails, report the matching fix and stop — do not run the script.

5. Run the script for whichever datasets need fetching:

   # Fetch both (default):
   bash "$CLAUDE_SKILL_DIR/fetch-prs.sh" --author <a> --org <o> --out .cv-data
   
   # Fetch only the missing one:
   bash "$CLAUDE_SKILL_DIR/fetch-prs.sh" --author <a> --org <o> --mode authored-all --out .cv-data
   bash "$CLAUDE_SKILL_DIR/fetch-prs.sh" --author <a> --org <o> --mode reviewed     --out .cv-data
   

   Re-fetch is a clean overwrite — the script recreates each dataset file in full, so no merge handling is needed.

6. Hand the resulting ./.cv-data/<org>/<author>/ directory to Step 1 and continue exactly as for a user-supplied directory.

Step 1 — Read mode, pick voice

Read the top-level mode field:

mode	Meaning	Voice (verbs)
authored-all	Work the author led/built	Ownership: Designed, Led, Built, Architected, Migrated, Automated
reviewed	Work the author joined as a team member	Contribution: Contributed to, Participated in, Supported, Helped deliver

Step 2 — Analyze prs[]

Bodies are usually empty and titles carry the ticket-style summary (e.g. INF-3184 Add new image versions for metrics-server and cluster-autoscaler). So the title tells you what a PR did; files[] tells you which technology and which subsystem — the part a bullet needs to name a stack and group related work. Use them together, with files[] as the sharpening signal, not the lead.

2a — Detect the tech stack from file extensions and paths

A PR's file extensions are a reliable fingerprint of the technology, even when the title is terse. Map them to the stack you'd actually name on a CV:

Signal in files[].name	Technology to name
.tf, .hcl, .tftpl, .tf.json	Terraform / Terragrunt (IaC)
.ts + cdktf.json under a CDK/infra path	CDKTF (TypeScript IaC)
.yaml/.yml under playbooks/, roles/, inventory/	Ansible
.yaml/.yml with k8s kinds, charts/, Chart.yaml, values.yaml	Kubernetes / Helm
.github/workflows/*.yml	GitHub Actions CI/CD
.gitlab-ci.yml, .gitlab/	GitLab CI
*.sentinel, policies/, sentinel.hcl	Policy-as-code / governance
Dockerfile, *.dockerfile	Docker / container images
argocd/, applicationset, *.app.yaml	ArgoCD / GitOps
.go, go.mod, go.sum	Go (compiled services / CLI tooling)
.py under services/<name>/, adapters/, plugins/, with pyproject.toml/Pipfile	Python application / backend service
.py, .sh as standalone scripts (scripts/, tasks/, root)	Python / Bash tooling and automation
.vue, .jsx/.tsx, .scss/.css, .svelte (genuine authorship — see caution below)	Frontend (Vue / React / etc.)

Treat this as a starting map, not a closed list — these examples skew toward infra because that's one author's stack, but the method is stack-agnostic: read whatever the files plainly indicate (Go, Rust, Java, frontend, data pipelines, …) and name the technology the bulk of a cluster's files point to.

Two judgment calls the extensions alone won't make:

• Application vs. script. .py/.go under a named services/<name>/ (or adapters/, plugins/) directory, especially with build/manifest files, is a shipped service — a stronger CV claim than a one-off script in scripts/. The path tells you which; name it accordingly ("built a service" vs "wrote tooling").
• Authored vs. generated/synth output. Not every file in a PR was hand-written. Generated and vendored files inflate counts without proving skill: CDKTF emits cdktf.out/*.json (and *.tf.json) from a small .ts change; lockfiles (*.lock.hcl, package-lock.json, go.sum), a terraform fmt sweep, or a bulk .vue/.tsx/.scss rename all balloon file/line totals. Weigh the authored source — count the .ts/.go/.py/.tf a human wrote, not the emitted artifacts. A surge of generated or moved files in a single PR is the tell; confirm a real skill with the per-PR count of authored files across several PRs (the any() pattern below), not the raw file tally — one 30-file rename or a 500-file synth dump becomes a fake "major effort" otherwise.

2b — Cluster the work

• Primary cluster key: top-level path prefix, then repository. The directory a PR touches is usually the project boundary — ecr/images/ + ecr/charts/ is one "container image / ECR management" theme; playbooks/ is an Ansible-automation theme; .github/workflows/ is a CI theme — even across PRs with unrelated titles.
• Within a path cluster, group by recurring themes in title / labels.
• Also cluster by named subject across repos/paths. A single accomplishment often spans several locations: a service named foo might appear as app code in services/foo/, a charts/foo/ Helm chart, an argocd/foo manifest, and the deployments/foo Terraform that runs it — across two or three repos. When the same name recurs, that's one end-to-end "designed, built, and shipped X" theme, a strong CV bullet. Don't fragment it into a separate bullet per repo.
• Merge clusters that describe the same accomplishment; a bullet is a theme, not a PR.

When counting how many PRs fall in a cluster with jq, match at the PR level with any(...), not a bare .files[] predicate — the latter iterates files and counts a PR once per matching file, badly inflating large clusters. Use: [ .prs[] | select(any(.files[]?.name; test("^deployments/cloudflare"))) ] | length

Starter recon block — run these first to surface clusters before reading titles ($F = the JSON path):

# repos and tenure window
jq -r '[.prs[].repository]|group_by(.)|map({r:.[0],n:length})|sort_by(-.n)|.[][]' "$F"
jq -r '[.prs[].mergedAt // .prs[].createdAt]|min,max' "$F"
# PR-level histogram of top-level path prefixes (NOT file-level — see any() above)
jq -r '[.prs[]|[.files[]?.name|split("/")[0]]|unique[]]|group_by(.)|map({p:.[0],n:length})|sort_by(-.n)|.[][]' "$F"
# count PRs matching one cluster regex
jq -r '[.prs[]|select(any(.files[]?.name; test("REGEX")))]|length' "$F"

Note the inner unique in the histogram — it dedups paths within a PR so each PR contributes at most once per prefix (a file-level split would re-inflate).

2c — Weight clusters (what's CV-worthy)

• Breadth of meaningful change is the strongest signal: distinct non-trivial files across multiple repos/subsystems > one large auto-generated or vendored file.
• Then: PR count in the cluster, recency (mergedAt/createdAt), and presence of governance/quality artifacts (*.sentinel, CODEOWNERS, .pre-commit-config.*) which signal "established standards/compliance," a CV-worthy theme in itself.
• Treat raw additions+deletions as a weak proxy only — a 2000-line diff can be a single terraform fmt or a lockfile regen. Let files[] sanity-check linecounts.

2d — Discard noise

A PR is noise if files[] and title show no substantive change. Drop it when:

• Every file is config/lint/format/test scaffolding only — matches just .pre-commit-config.*, .yamllint, .yamlfmt, .editorconfig, *.md, or test_*.
• It's a dependency/version bump, a revert, or a typo/whitespace fix.
• The title flags it as non-work: DO NOT MERGE, Test, WIP, Fix mistake, Revert, bump.

2e — Grounding

For thin/empty body, infer the change from files[].name + commits[].message + diffstat — and state only what those plainly support. Name the stack and subsystem (files prove those); do not invent outcomes (cost %, SLA, audit results) that the data cannot show. If a cluster's impact is unclear, describe the work plainly.

Step 3 — Produce bullets

• Aim for 3–6 bullets, scaled to tenure and signal. A short stint or thin dataset should get fewer (2–3); a multi-year role carries 5–6. A long tenure (5+ years) with many genuinely distinct themes can run to 7 — but only when each extra bullet is its own real accomplishment, not padding. The rule is one-bullet-per-theme, not a target count: if merging two thin themes reads stronger, merge them; never split one accomplishment to hit a number.
• Match the existing index.md job-entry style: - bullets, action-verb first, past tense, impact-oriented.
• No PR numbers, no URLs, no repo names in the bullet text.
• Each bullet = one distinct theme. Merge overlapping clusters.
• Use the voice chosen in Step 1.
• Ground every claim in the data. State only impact you can support from titles, labels, files, commit messages, and diffstat. If a cluster's impact is unclear, describe the work plainly rather than inventing an outcome.

Reconciling authored and reviewed. When both modes exist for the same author, their themes usually mirror — the person who builds a subsystem also reviews changes to it. So the authored bullets carry the work, and the reviewed file contributes one bullet only: the distinct collaboration/mentorship/gatekeeping point that ownership voice can't make on its own (e.g. "core reviewer for the platform — vetted ~N pull requests across the X and Y codebases to uphold quality standards"). Ground it in what the reviewed data supports: review volume (pr_count) and repo/path breadth are solid; do not claim "reviewed N distinct contributors" unless the data carries a per-PR author field (it often doesn't — say so if asked). Never restate an authored accomplishment in reviewed voice; that double-counts one body of work. If the reviewed themes add no point beyond what authored already says, it's fine to skip the bullet rather than force a weak one.

Step 4 — Print and save

1. Print the candidate bullets, grouped under the target job entry (ask the user which job if ambiguous), so they can review them in the conversation.
2. Save the same combined list to a timestamped results-<stamp>.md inside the input directory, where <stamp> comes from date +%Y-%m-%d-%H%M run in the shell (e.g. .cv-data/<org>/<author>/results-2026-06-06-1825.md). The timestamp keeps each run as its own file rather than clobbering the last — handy for comparing successive passes. Save beside the source data, out of the site itself. Use a top heading naming the job entry, then the - bullets — authored bullets first, the single reviewed bullet last.
3. Stop there. This skill does NOT touch index.md and does NOT call any other skill. Inserting bullets into the actual CV is a separate, explicit step the user takes afterward (the update-cv skill, by hand, etc.).