agentleFS
boykushboykush/scraps/.claude/skills/dogfood/SKILL.md

dogfood

Skill45 starsChanged 54 days agoMITView on GitHub

Improve Scraps' MCP server and agent skills from how they were actually used. Reads the traces the owner's own deployment leaves in Jaeger, turns one finding into a change, and opens the PR without stopping to ask in between. Owner only - the traces sit behind the owner's login. Use this whenever the owner asks to dogfood, to analyse how the MCP server or a skill is being called, or to find the next MCP improvement from real usage (e.g. "/dogfood", "MCP の呼ばれ方を分析して改善して", "先週の trace から次の改善を").

81 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use dep-refs, feature-boundary-planner and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • axiom-desuSkill

    dep-refs

    Use when needs to inspect TypeScript export references in the z-code workspace, list exports from a file, verify whether an export is unused before deletion, investigate who imports a symbol during refactors, or combine pnpm knip unused-export results with pnpm dep:refs symbol-level reference tracing.

    Also in 6 projects that use this

  • axiom-desuSkill

    feature-boundary-planner

    Map a ZCode behavior change to current UI surfaces, state owners, protocol commands, persistence, and validation. Use for impact analysis, product-boundary planning, or an implementation handoff grounded in the checked-out source.

    Also in 6 projects that use this

  • lir-navcooSkill

    yuanbao

    Yuanbao (元宝) groups: @mention users, query info/members.

    Also in 18 projects that use this

  • Cartograf666Skill

    architecture-governance

    Bootstrap, document, enforce, and review a project's architecture through layer responsibilities, dependency direction, architecture manifests, legacy baselines, Change Envelopes, cross-layer approval policy, and deterministic architecture checks. Use when starting a project, auditing an existing codebase, planning cross-layer work, preventing architecture erosion, reviewing structural changes, or deciding whether implementation may expand into another layer.

    Also in 6 projects that use this

  • math-incSkill

    minecraft-modpack-server

    Set up a modded Minecraft server from a CurseForge/Modrinth server pack zip. Covers NeoForge/Forge install, Java version, JVM tuning, firewall, LAN config, backups, and launch scripts.

    Also in 5 projects that use this

  • math-incSkill

    pokemon-player

    Play Pokemon games autonomously via headless emulation. Starts a game server, reads structured game state from RAM, makes strategic decisions, and sends button inputs — all from the terminal.

    Also in 5 projects that use this

  • +1 more from math-inc

The file

We split this skill into the rules it gives an agent

4 rules. 0 are common, 1 is distinctive, and 3 are very specific to this project.

  • 0

    Common

    Most of the 81 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  • could be automatic: a setting could enforce it
  1. Very specific:

    Only the owner can run it.

    line 10
  2. Very specific:

    If the jaeger tools are missing or refuse the call, stop and say so.

    line 12
  3. Distinctive:

    When the traces cannot answer a question that matters, that gap is a finding in its own right.

    line 33
  4. Very specific:

    Follow the workflow in AGENTS.md: a branch from origin/main, TDD, Conventional Commits, one PR.

    line 58

“Common” is measured: kept by at least 6 in 10 of the 81 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Dogfood

The repository owner runs Scraps as a remote MCP server over their own wiki and uses it every day. Each request leaves a span. This skill reads those spans, finds where the tools and the way agents are told to use them fall short, and ships one improvement.

Only the owner can run it. The jaeger MCP server authenticates as them, and no one else gets past its login. This skill is also the only place in the repository that calls mcp__jaeger__*: outside it, leave those tools alone.

If the jaeger tools are missing or refuse the call, stop and say so. The package's README (apm_modules/boykush/ai-plugins/plugins/jaeger-remote-mcp/README.md) says how to sign in and when the cluster is down. Don't route around the login.

1. Read the traces

  • Find the service with the services tool rather than assuming its name: it is the one whose spans come from the scraps instrumentation scope.
  • Take the whole retention window unless the owner names another. A week of personal use is a small sample, so narrowing it further leaves little to read.
  • src/mcp/traced.rs owns what a span is named and which attributes it carries. Read it before interpreting them, since it changes as the telemetry improves.
  • A trace can hold other services' spans for the same request, such as a gateway in front. Count only the spans whose resource is the Scraps service, or each call is counted twice.

2. Find what the usage says

Set the calls beside what the server offers: src/mcp/server.rs for the tools, their descriptions and the server instructions, src/mcp/tools/ for each request schema.

  • Tools never called. A tool nobody reaches for is either unneeded for this kind of use or undiscoverable from where agents stand when they would want it.
  • Arguments that fight the schema. A default passed explicitly every time, an optional parameter never used, a value retried in another form.
  • Sequences. What follows what, and where a sequence ends early. A search with no read after it is a question the response could not answer.
  • Errors, and whether the message told the caller what to do next.
  • Latency that every tool shares, as opposed to one tool's own.

Set aside the calls that were not use: connectivity checks, and the calls this skill's own earlier runs caused. They are recognisable by their arguments and by arriving in a burst.

When the traces cannot answer a question that matters, that gap is a finding in its own right. A span that records the missing fact is a legitimate change to ship, and often the one to ship first.

3. Pick one change

Rank the findings by how much evidence stands behind each and how small the change is. Take the top one. Before writing code, state:

  • the evidence, as counts over a dated window
  • the change
  • what the traces should show afterwards if it worked

The last one is what makes the next run able to judge this one. A change with no observable expectation is a guess.

Where the change lands depends on what was wrong:

What was wrongWhere
A tool's behaviour, schema or responsesrc/mcp/tools/, src/mcp/json/
How agents are told to use the toolsDescriptions and instructions in src/mcp/server.rs
What the traces could not showsrc/mcp/traced.rs
How a skill drives Scrapsplugins/*/skills/

A changed tool also changes plugins/mcp-server/README.md, and a changed plugin skill needs its version bumped and its evals run (CONTRIBUTING.md).

4. Ship it

Follow the workflow in AGENTS.md: a branch from origin/main, TDD, Conventional Commits, one PR. Go from the finding to the open PR without asking for confirmation in between. The PR is where the owner reviews; don't merge it.

The PR body carries the evidence, the change and the expectation from step 3.

Keep what the owner searched for out of the repository. Span arguments are the owner's own queries, and this repository is public. Commits, PR text, tests and fixtures get counts and shapes ("26 of 28 searches passed the default explicitly"), never the query strings or scrap titles. Quoting them to the owner in the session is fine.

5. Close the loop

Report to the owner: the findings ranked, the one shipped with its PR, and the ones left for next time.

The next run starts by checking the previous one. Find the earlier dogfood PRs (gh pr list --state merged --search "dogfood in:body"), see whether each has reached the deployed version (the instrumentation scope's version on the spans), and compare the traces since then against the expectation it stated. An expectation that did not hold is the first finding of the new run.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.