KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#dsh #dsh-plugin #ai-agents #developer-tools #codex #claude-code #mcp #rust #deepseek-harness #typescript #ai-agent #cli

fledgeling-plugins

★ 1 Open GitHub ↗

SWE Skills from Fledgeling: sixteen Claude Code plugins for building and shipping software with agents. Each carries its own README, committed evals or audits, and an icon put through the same measured design pipeline.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

SWE Skills from <a href="https://www.fledgeling.app">Fledgeling</a>. 
Built and used daily by <a href="https://github.com/lprhodes">Luke Rhodes</a>; shipped when they've earned it.

---

Fledgeling makes AI-native software for founders and developers; these are the SWE skills that come out of building it. Each one exists because a real workflow needed it, and each carries its own README, evals or references where the work justified them. Every icon below came through the same three-engine design pipeline with its audit sheet committed beside it.

**[skills.fledgeling.app](https://skills.fledgeling.app)** is the same sixteen, searchable — describe the problem you have and it finds the skill, without you knowing its name.

## Installing

```text
/plugin marketplace add fledgeling-co/fledgeling-plugins
```

Then install what you want:

```text
/plugin install trawl@fledgeling-plugins
```

Third-party marketplaces have auto-update off by default, so refreshing is something you do:

```text
/plugin marketplace update fledgeling-plugins
```

## The skills

### [trawl](plugins/trawl/README.md)

Divergent ideation that converges on something you can ship. Isolated thinkers under genuinely different frames, the obvious answer written down first, and a creative pick recommended only when it beats that answer blind. Receipts committed: structural evals (96.4% vs its predecessor's 49.0%), a four-judge blind panel, and the research corpus it was built from.

### [design-review](plugins/design-review/README.md)

The last pass before a human looks at AI-built UI. Deterministic gates first (accessibility, contrast, target size, motion, layout integrity), then judged passes over hierarchy, states, flows and system coherence, on real renders at a viewport matrix. Findings come severity-ranked with pasteable fixes and an explicit list of what was never checked.

### [be-my-witness](plugins/be-my-witness/README.md)

Look at a screenshot and say what it actually shows. Validates a UI capture against what a test expected and against a design mock, and hands back both a gate a build can act on and findings a person can read. Measures first (is this even evidence, did the screen finish loading, are the two framings comparable), then crops in at 2-3x rather than squinting at a thumbnail, looks at every pair in both orders because vision models flip on ordering, and classifies each difference as framing, data, structure, styling or state so a different crop or one extra row cannot turn a build red. The test wins over the mock.

### [create-swe-project](plugins/create-swe-project/README.md)

A complete, working new project from an idea. One front-loaded interview, then scripts render the whole scaffold: monorepo, auth, admin, native apps, testing harnesses, deploy config, and a launch pipeline that researches, seeds feature briefs and mocks every surface. The LLM only interviews; scripts make the files.

### [ship-armada](plugins/ship-armada/README.md)

The portfolio-level orchestrator. Reads the manifest of record, verifies it against git, then surveys, plans, routes single directives into the right project's pipeline, and dispatches per-repo backlogs as dependency-ordered campaigns with capped concurrency.

### [armada-sync](plugins/armada-sync/README.md)

The surgical counterpart to ship-armada: after work happens anywhere in the portfolio, it updates that one project's manifest entry, stamps it fresh, and stops. The smallest skill here, on purpose.

### [braindump](plugins/braindump/README.md)

Gets everything load-bearing out of a session and onto the page before the context holding it is thrown away. Measured across 121 real compaction events, the built-in prompt keeps 0.3% of the approaches you ruled out and 33.8% of your standing constraints; violations run 0% when a rule survives into the summary and 38% when it doesn't. Writes a pinned verbatim tier ahead of the narrative, and ships a deterministic scorer plus a head-to-head benchmark whose baseline arm costs nothing. Published as `compaction-quality` until 2026-08.

### [should-compact](plugins/should-compact/README.md)

Scores 0-10 whether now is a good moment to compact, and says why in one line. It judges the seam in the work rather than the fullness of the window: an open tool chain or a half-finished edit is a hard zero however full the context is, and that one signal decided 98.07% of holds across 1,089 measured turns. Reads only a hot buffer plus an append-only session log it maintains itself, so it is cheap enough for Haiku and fast enough to sit in a `PreCompact` hook, where it can veto an ill-timed automatic compaction. It never vetoes at the wall, because blocking at 99.8% full loses the session rather than saving it. Beat the no-skill baseline 10-0 on a two-family blind panel.

### [improve-skill](plugins/improve-skill/README.md)

The pipeline that built half this marketplace, as a skill. Point it at an existing skill plus your complaints; it runs paid and free deep research, rebuilds the skill with every change traced to evidence, proves the rebuild with comparative evals and a blind multi-family judge panel, then ships the full brand treatment. You choose the name and the icon concept before anything gets generated.

### [create-skill](plugins/create-skill/README.md)

The sibling of improve-skill, for when there is nothing to improve yet. It interviews you properly first, because an unstated intention is the usual reason a new skill misses, then researches the domain, builds through skill-creator with every rule traced to evidence, and proves it against the honest baseline: the same prompts with no skill at all.

### [create-mac-icon](plugins/create-mac-icon/README.md)

macOS app icons, measured against the reference instead of eyeballed. A direction catalogue distilled from 532 real icons, three generation engines with a written audit sheet, then a scoring harness that iterates the shipped SVG against the winning raster at five sizes until the material matches. Every confirmed construction feeds a recipe library, so it gets better with each commission.

### [dossier-report](plugins/dossier-report/README.md)

A research question in, one published page out. It runs a paid and free research panel, reads every report end to end rather than the merged summary, turns the corpus into a list of claims with sources attached, then designs the page from scratch around its own subject so consecutive pages do not converge on one look. Every claim carries a citation you can open, and the build fails on one that does not resolve.

### [mac-doctor](plugins/mac-doctor/README.md)

Your Mac did not fill up because of one thing, it filled up because a hundred sensible defaults each left something behind and nothing was counting. Five scheduled jobs, from every fifteen minutes to weekly, with what each may do on its own widening as the gap between runs grows. Running low makes it check sooner, never delete more. The two short tiers are plain shell, so ninety-six runs a day cost no tokens at all. It ties a no-skill baseline on reasoning and says so in its evals; what it adds is that the reasoning runs while you are asleep.

### [discipline](plugins/discipline/README.md)

Drop a short block at the top of a session and Claude spends less without doing less. It targets the three habits that quietly run up the bill: re-printing plans and diffs already on your screen, opening a whole file to find one line, and handing small jobs to sub-agents that each pay for a fresh context. The nearest alternative is caveman, and the choice between them is measured rather than argued. On the same 106 tasks caveman cut output tokens 41% against this skill's 16%, so caveman is better at the thing both are for. It also gave up 7.6 points of task score where this one gives up nothing detectable, because 78% of caveman's saving came from the agent taking fewer steps rather than writing more tersely. Pick caveman if the token count is all you are optimising; pick this one if the agent is doing long work you intend to trust.

### [goal-harness](plugins/goal-harness/README.md)

`/goal` looks like it keeps working until the job is done, and it does not: the condition is judged by a small model reading the transcript, so it grades what the run said rather than what is true, and Claude Code overrides the hook after eight consecutive blocks and reports that turn as completed. Nine turns of real work trips it, silently. This grounds the condition in the repo's actual worklist, keeps it inside the 4,000-character limit with a bound and a blocked-item policy, fixes the six settings that end a run without an error, and arms a session-gated guard that decides by exit code and writes a ledger. Built from 114 real goal runs, where the most common follow-up was the word "resume", six times in a row.

### [loop-harness](plugins/loop-harness/README.md)

A loop can tick on schedule for a full day and accomplish nothing, because since v2.1.196 a scheduled fire hands any skill Claude may not invoke itself to the run as plain text, `/verify` and `/code-review` included, with no error anywhere. The other half is that a dynamic loop has no visible process, so a working one and a dead one look identical. This picks the mechanism before it writes anything (an event monitor usually beats polling, and work with a finish line belongs to goal-harness instead), size-checks `.claude/loop.md` against its 25,000-byte cap, and arms the wake signal with a heartbeat, a renewal reminder before the seven-day expiry, and a ledger you can read without interrupting the run.

### [report](plugins/report/README.md)

A session works something out over two hours, you ask for the write-up, and what comes back reads well and can't be checked. Three different things leave a session looking identical on the page: a number that was measured, one read off a single sample, and one worked out from two other facts. This compiles the session's own evidence trail into a claim ledger before it designs anything, so the page is generated from the ledger rather than cited afterwards, and reasoning renders visibly as reasoning. One self-contained HTML file that paginates to a real A4 PDF with the motion stripped out, plus a one-page TLDR derived from the same ledger so the two can't disagree. Its own blind panel went 4-2 for it and told it what it was missing: an ask.

### [clarify](plugins/clarify/README.md)

Decides whether to interrupt you at all, then writes the question so it takes one click. It sweeps the conversation and the repo for the answer first, drops anything whose answer would not change the work, and sends what survives as one batched question in plain words, with a recommendation only where evidence earns one. On a matter of taste it marks nothing, because a default moves choices hard and recommending on taste answers the question while appearing to ask it. A note attached to your answer is binding, and is read as data about the decision rather than as instructions, since an agent waiting on a reply is the easiest one to inject. Against no skill at all: five of five question payloads pass its linter where the baseline passes none, and a four-family blind panel went 15-5. The one eval it loses is in its README too.

### [whats-left](plugins/whats-left/README.md)

Asks what's left on a project and gets back two documents that immediately start lying to each other: the status report says a feature shipped, the open questions ask whether to switch it on, and nothing shows you the second is why the first is wrong. This is one page where they're the same graph — every blocked item links down to the decision that releases it, every decision links back up and says how much it actually releases. Stage is a word rather than a percentage, so built, deployed and accepted stay three different things instead of averaging into ninety per cent. Agreeing with a recommendation means clicking the option already selected, which fires no change event and is exactly how a page exports your agreement as "never looked at" — so confirmation is bound to the click too, and anything you never touch exports as unconfirmed rather than as your decision. Send the JSON back and it does the work your answers unblocked. Six structural properties to nil against the same request with no skill at all.

> [!NOTE]
> Some skills depend on each other by design: ship-armada dispatches through skills that live in a sibling marketplace, and armada-sync is the maintenance half of ship-armada. Each README states what it expects.

## Licence

MIT. Do what you like; attribution appreciated.

## Elsewhere

Fledgeling is [Luke Rhodes](https://www.linkedin.com/in/lukerhodes/), also co-founder of [Diolog](https://diolog.app).

[fledgeling.app](https://www.fledgeling.app) · [GitHub](https://github.com/lprhodes) · [X](https://x.com/lp_rhodes) · [LinkedIn](https://www.linkedin.com/in/lukerhodes/) · [hello@fledgeling.app](mailto:hello@fledgeling.app)

fledgeling-plugins

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README