portfolio-review
Skill2 starsChanged 30 days agoMITView on GitHub
Audit publishable portfolio copy, content completeness, claims, links, canonicals, redirects, images, social metadata, accessibility, and rendered presentation. Use for every public content addition or meaningful edit before it is considered ready for review or publication.
15 other projects use a version of this.
Adopted by
Most recently taken up by
- PatrickSUDO/fadacai-portfoliotheir own version, first seen Nov 13, 2026
- CSPulse/customer-success-skillstheir own version, first seen Nov 8, 2026
- rgourley/quant-garagetheir own version, first seen Nov 1, 2026
- Shubhamkumartrichy/claude-trading-skills-indiatheir own version, first seen Nov 1, 2026
- JasmineNy/codex-skills-portfoliotheir own version, first seen Oct 20, 2026
- monellonasti/portfolio-pipelinetheir own version, first seen Oct 17, 2026
Other versions15 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 15
- digitalinnovationone/dio-agent★ 2.8ksame name
- PatrickSUDO/fadacai-portfolio★ 145same name
- Shubhamkumartrichy/claude-trading-skills-india★ 64same name
- nimadorostkar/Claude-Skills-collection★ 37same name
- Mozurok/fhorja.dev★ 10same name
- rgourley/quant-garage★ 7same name
- CSPulse/customer-success-skills★ 2same name
- shashu10/personal-finance-skills★ 2same name
- tusharagg1/aifolimizer★ 2same name
- yeshelloAB/investment-analysis-skills★ 1same name
5 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository9 other files it gives its agents.
- agent-scripts/AGENTS.mdAGENTS.md
- AGENTS.mdAGENTS.md
- frontend-responsive-designSkill
- google-maps-a2uiSkill
- infographic-agentSkill
- portfolio-contentSkill
- portfolio-designSkill
- portfolio-presentingSkill
- skill-improvement-loopSkill
The file
We split this skill into the rules it gives an agent
12 rules. 0 are common, 10 are distinctive, and 2 are very specific to this project.
0
Common
Most of the 14 edited versions we read keep these.
10
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 8
Treat review as a bounded evidence loop, not a vibe check.
- Distinctive: line 14
Run git status --short before work and inspect the focused diff.
- Very specific: line 17
Use docs/PORTFOLIO_EVIDENCE_LEDGER.md for existing claims.
- Distinctive: line 22
Check every number, date range, attribution, role, causal statement, and current product fact against a primary source, checked-in …
- Distinctive: line 28
Verify that the opening states the thesis, pays off the title/social promise immediately, each section advances it, and the ending tells …
- Distinctive: line 35
Run the portfolio build to validate internal links, assets, duplicate slugs, canonical rules, and aliases.
- Distinctive: line 43
Keep one canonical owner. Do not publish duplicate local and external canonicals.
- Distinctive: line 48
Prefer real screenshots and artifacts.
- Distinctive: line 49
Keep prompt scaffolding and generation context out of the rendered canvas.
- Very specific: line 52
Ensure social preview images are optimized .jpg files, compressed (e.g., JPEG quality 70-80), and ideally under 200KB so preview bots and …
- Distinctive: line 53
Write literal, asset-specific alt text.
- Distinctive: line 72
Run the narrowest relevant checks, then the complete content path when practical:
“Common” is measured: kept by at least 6 in 10 of the 14 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Portfolio review
Treat review as a bounded evidence loop, not a vibe check. Read the portfolio writing, design, and content skills before reviewing their surfaces.
Before making or reviewing a material change, define success across three dimensions: the reader-visible outcome, fidelity to the writing/design rules, and the complexity or weight added to achieve it. Grade the result, not the implementation path. Use deterministic checks for structural requirements, a separate qualitative review for taste, and negative cases for things that must stay absent. Turn a real failure or user correction into the smallest durable regression check.
1. Inventory the change
- Run
git status --shortbefore work and inspect the focused diff. - List every changed public claim, link, canonical or alias, image, alt string, and metadata field.
- Identify the intended reader action, the misconception or tension used by the title/social copy, and the one idea each visual must communicate.
- Use
docs/PORTFOLIO_EVIDENCE_LEDGER.mdfor existing claims. Add durable new evidence there when a material claim will recur.
2. Audit copy and claims
- Lead with the outcome and concrete work. Cut throat-clearing, generic summaries, repeated conclusions, and unsupported superlatives.
- Check every number, date range, attribution, role, causal statement, and current product fact against a primary source, checked-in artifact, or Ryan-approved internal evidence.
- When copy cites an eval or benchmark, inspect the execution path instead of accepting the repository link as evidence. Confirm that the runner invokes the model or system, train and held-out cases are disjoint, scores are computed from retained outputs, and the public configuration or logs support the claimed run. Constants, fixtures, commented-out trainers, and stubbed calls may document an intended result, but they never support a measured one.
- Apply the metric taste rule: real numbers are fine for public/verifiable stats, prior-company results, and aged or long-public current-employer work; recent internal current-employer usage or growth figures (Google Maps Platform users, API engagement) must use qualitative, understated framing, never a precise percentage.
- Distinguish observed correlation from causation. Put unrelated evidence in separately labeled groups.
- Credit team work accurately. Default to "Our team built… I led the strategy," not solo credit, but do not dilute genuinely individual work. Remove or qualify anything the evidence does not support.
- Generalize third-party tools: name first-party surfaces (AI Studio), never enumerate competitor AI products. Flag employer product-marketing tone that reads as selling the platform or looking for work elsewhere.
- Verify that the opening states the thesis, pays off the title/social promise immediately, each section advances it, and the ending tells the reader what to do.
- For technical posts, verify that prose introduces each code sample and explains the behavior or tradeoff after it. Flag ornamental samples, unexplained blocks, and long runs of code without a human narrative.
- Search the changed prose for banned voice patterns such as passive self-credit, hype adjectives, resume bullets, and em dashes. Also flag three or more adjacent sentences that repeat the same subject-verb structure, especially glossary definitions and metric sequences. Read the copy aloud for varied rhythm and conversational flow because style cannot be proven by a regex, and compare the rhythm against the hand-written calibration reference linked in the writing skill.
- Review formatting as part of voice. A sequence of report table, glossary bullets, diagnostic JSON, pseudo-rubric, numbered checklist, and survey of outside examples is a blocker when one lived artifact could carry the argument. After the claims pass, do a separate de-scaffolding read that keeps the evidence but restores a human narrative.
3. Audit links and URL ownership
- Run the portfolio build to validate internal links, assets, duplicate slugs, canonical rules, and aliases.
- Open material external links or fetch them with a real GET. Confirm the destination supports the adjacent claim; do not treat a successful status alone as evidence.
- A failed third-party browser request is not proof of a CSP violation in this environment. Reproduce the same request on a page without CSP, compare it with a direct GET, and inspect the served header and violated directive before changing an allowlist. If the no-CSP control also fails, report browser reachability as unavailable instead of weakening the policy.
- For a renamed page, verify the new canonical and social URLs in rendered HTML and test every old path, with and without a trailing slash, for an HTTP 308 to the clean root-relative target. Confirm query strings survive.
- Keep one canonical owner. Do not publish duplicate local and external canonicals.
4. Audit images and metadata
- Require every hosted essay to have three distinct assets: a 1200x675 thesis header, a 1200x627 social preview, and at least one 1200x675 inline mechanism or evidence image.
- Prefer real screenshots and artifacts. Generated visuals must show a mechanism or evidence, never invent metrics, UI, logos, or product behavior.
- Keep prompt scaffolding and generation context out of the rendered canvas. Source labels, review dates, tool metadata, and process instructions live in the generation record unless the article is specifically presenting them as evidence.
- Iterate with the lower-cost image model; render the approved final prompt with the quality model. Save exact prompts, model IDs, aspect ratio, size, thinking setting, and post-processing in
docs/. - Inspect file signatures with
file; never trust an extension. Verify physical dimensions and generated HTMLwidth/heightattributes. - Ensure social preview images are optimized
.jpgfiles, compressed (e.g., JPEG quality 70-80), and ideally under 200KB so preview bots and answer engines can ingest them instantly. - Write literal, asset-specific alt text. Do not reuse the header description for the social or inline image.
- Inspect the rendered page at desktop and narrow mobile widths. Check legibility, crop, padding, hierarchy, dark/light behavior when relevant, and whether the image earns its space. For diagrams, verify that each visual carries one idea, retained labels remain readable at article width, boxes have visible internal padding, and no connector, arrowhead, icon, or border collides with text. Remove microcopy before shrinking it.
5. Run the maker/checker loop
For every publishable content change, use at least one independent read-only reviewer after the maker pass. For an essay or multi-surface page, split review into these lanes when agents are available:
- Copy and claim integrity.
- Links, canonical ownership, redirects, and metadata.
- Visual taste, image truthfulness, accessibility, desktop, and mobile rendering.
Give reviewers the raw diff, rendered page, and assets. Do not give them the intended verdict. Ask for actionable blockers, not rewritten taste preferences.
After each review round: classify findings, make one focused correction pass, rerun deterministic checks, and request a fresh read-only review. Stop after at most three rounds. Stop earlier when all lanes are clean. If a material claim, source, or design decision remains unresolved, return NEEDS_HUMAN; do not average reviewer opinions or declare victory.
Use deterministic tools and lower-cost models for inventory and mechanical checks, a balanced model for edits, and the strongest justified model only for final cross-surface synthesis or unresolved ambiguity. For local voice, cadence, and citation checks, run npm run voice:review -- <file> (or ./scripts/gemma-local.sh review <file>) for Ryan's voice standards, npx hemmingway review <file> for anti-slop copyediting, and npm run eval:citations -- --file <file> to check active first-person framing, rhythm, and citation integrity on Metal before final sign-off. The maker is never the only grader.
6. Required verification
Run the narrowest relevant checks, then the complete content path when practical:
npm run check:content # mechanical writing and content rules; run this first
npm run voice:review -- <file> # local voice and cadence critique (Dense 31B)
npx hemmingway review <file> # local anti-slop copyediting critique (clean stdout for piping)
npm run eval:citations -- --file <file> # verify academic & external citation links
cd portfolio && npm test && npm run build
cd ../gateway && npm test
cd .. && node scripts/build-local.mjs && node scripts/smoke.mjscheck:content decides only what a regex can decide: em-dashes, banned phrases,
hype adjectives, the essay asset contract, alt-text distinctness, tag
vocabulary, and phrases repeated across entries. A clean run means the review
lanes below can spend their attention on claims, evidence, rhythm, and taste. It
is not a substitute for any of them, and a warning it raises is a prompt to look,
not an instruction to obey.
Also verify changed image signatures and dimensions, inspect desktop/mobile browser captures, and test affected redirects with HTTP requests. Report every command and result. Do not claim browser, link, or image review without captured evidence.
Finish only when the diff is focused, the worktree contains no secrets or accidental generated files, documentation is updated, independent review is clean, and deterministic checks pass or their exact limitation is disclosed.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

