Skip to content

Repository files navigation

ecogo-workflow

A Claude Code plugin that gives software developers a self-correcting knowledge base. It reads your RFCs, design docs, PR descriptions, and code. It answers your questions with citations. It fixes itself when the code changes. It learns your conventions over time.

You talk to it in plain English. It does the rest.

ecogo save https://my-rfc.example.com         ← feed it an RFC
ecogo how does our auth flow work?             ← get an answer with citations
ecogo the deployment runbook looks wrong       ← it re-grounds against your code
ecogo design a rate limiter                    ← walks you through a proper design
ecogo build me an auth service from scratch    ← design → plan → implementation

Why a developer should care

Every engineering team has the same problem: the knowledge rots faster than it's written. Slack threads disappear. PR descriptions are one-shot. The two senior engineers who know why v1 was abandoned eventually leave. New engineers spend weeks "asking around".

ecogo-workflow is a disciplined habit that compounds. You feed it the sources you already write (RFCs, design docs, meeting notes, ADRs, code comments) and it:

Situation What it does
You write a new RFC You ingest it; the plugin cross-links it to everything related
You join a new team You ask in plain English and get grounded answers pointing into the real code
The code changed and docs didn't correct re-grounds the affected page automatically
Same bug keeps biting you learn notices the pattern and drafts a preventive rule
You're starting a new feature brainstorming walks you from one-line idea to a written design spec before any code
You have a plan ready to execute subagent-driven-development runs it task-by-task with two-stage review

It runs in Claude Code. Your knowledge lives in Obsidian (so it's portable — plain markdown files on your own disk, not locked in a vendor's database).


What you'll need before installing

Three tools. All free. All work on Mac and Windows.

  1. Claude Code — Anthropic's CLI / desktop app. Get it from claude.com/claude-code.
  2. Obsidian — the editor where your knowledge will live. Get it from obsidian.md.
  3. Git — for tracking changes to your knowledge base. Most devs already have this; if not, get it from git-scm.com.

You also need Node.js 18+ — needed because the plugin auto-installs a search tool on first use. Get it from nodejs.org if you don't have it.


Step-by-step: Mac setup

1. Install Claude Code

Open Terminal (⌘+Space, type "Terminal", press Enter) and run:

npm install -g @anthropic-ai/claude-code

Then sign in:

claude

Follow the prompts to sign in with your Anthropic account.

2. Install Obsidian

Download the Obsidian DMG from obsidian.md and drag it to Applications.

Open Obsidian and create a new vault:

  • Click Create on the welcome screen
  • Name it ObsidianVault
  • Choose your home folder as the location (so the final path is ~/ObsidianVault)
  • Click Create

Inside the vault, create a folder called 03-Resources:

  • Right-click in the left sidebar → New folder → type 03-Resources

That's the folder where your knowledge bases will live, one subfolder per topic.

3. Configure git (if you haven't already)

In Terminal:

git config --global user.name "Your Name"
git config --global user.email "you@example.com"

4. Install the ecogo-workflow plugin

In Claude Code (the claude CLI or the desktop app), run:

/plugin marketplace add alisajil/ecogo-workflow
/plugin install ecogo-workflow@ecogo-workflow

Then restart your Claude Code session to pick up the new commands.

5. Verify it works

Start a fresh session in Claude Code and type:

/ecogo

If you see a friendly "Base is idle. Nothing urgent." with a list of suggested next actions — you're set. Skip to Your first 15 minutes below.


Step-by-step: Windows setup

1. Install Node.js

Download and run the Windows installer from nodejs.org. Pick the LTS version. Accept defaults.

Open PowerShell (press Windows key, type "PowerShell", press Enter) and verify:

node --version

You should see v18.x.x or higher.

2. Install Claude Code

In PowerShell:

npm install -g @anthropic-ai/claude-code

Then sign in:

claude

Follow the prompts to sign in with your Anthropic account.

3. Install Obsidian

Download the Obsidian installer from obsidian.md and run it. Accept defaults.

Open Obsidian and create a new vault:

  • Click Create on the welcome screen
  • Name it ObsidianVault
  • Choose your user folder as the location — e.g. C:\Users\<YourName>\ObsidianVault
  • Click Create

Inside the vault, create a folder called 03-Resources:

  • Right-click in the left sidebar → New folder → type 03-Resources

4. Install Git (if you don't have it)

Download from git-scm.com. Run the installer. Accept defaults.

In PowerShell:

git config --global user.name "Your Name"
git config --global user.email "you@example.com"

5. Install the ecogo-workflow plugin

In Claude Code, run:

/plugin marketplace add alisajil/ecogo-workflow
/plugin install ecogo-workflow@ecogo-workflow

Then restart your Claude Code session.

6. Verify it works

Start a fresh session and type:

/ecogo

If it responds with "Base is idle" — you're good.


Your first 15 minutes

A guided walkthrough. Replace auth-platform with whatever topic you care about.

Step 1 — Create your first knowledge base

In Claude Code:

/ecogo start a new knowledge base called auth-platform

This creates ~/ObsidianVault/03-Resources/auth-platform/ with folders for raw sources, wiki pages, reports, and a schema file. Open that folder in Obsidian if you want to see the structure.

Step 2 — Feed it a source

Any URL or file. An RFC, a design doc, a Notion page, a GitHub README, a meeting transcript. Let's use a web source:

/ecogo save https://raw.githubusercontent.com/golang/go/master/doc/README.md

This downloads the content to raw/articles/ — it's not processed yet.

Step 3 — Compile

/ecogo-workflow:ecogo compile

The plugin reads the raw source, writes a summary page, extracts the interesting concepts into their own pages, cross-links everything, and captures rationale (the why) where it finds it. This takes a minute or two.

Step 4 — Ask a question

/ecogo tell me about the Go documentation structure

You get back an answer with [[wikilinks]] — those are clickable in Obsidian and trace every claim back to the source you ingested. Every non-trivial factual claim is grounded.

Step 5 — Check it

/ecogo audit the engineering docs

This runs the lint check — looks for broken links, orphan pages, missing sections. For a tiny base this should pass cleanly; you'll see results accumulate as the base grows.

Step 6 — Do the smart thing

/ecogo

With no arguments, the plugin looks at the current state and picks the most useful action: compile if there's uncompiled raw material, lint if it's been a while, distill observations if the subsystem has new ones to roll up, etc. It announces what it's doing before doing it.


Common workflows for developers

"I just wrote an RFC and I want it integrated"

/ecogo save ~/work/rfcs/2026-04-auth-rotation.md
/ecogo-workflow:ecogo compile
/ecogo

"I'm starting a new feature and want a proper design first"

/ecogo design an RBAC layer for the admin portal

This invokes the brainstorming skill. It asks clarifying questions one at a time, proposes 2-3 approaches with tradeoffs, writes a design spec, and hands it off to the plan skill. It won't let you skip to code until you approve the design — this is by design.

"I have a written plan and want to execute it"

/ecogo implement the plan at docs/plans/rbac-plan.md

This invokes subagent-driven-development. It dispatches a fresh sub-agent per task, reviews the output in two passes (did they build what was asked? is the code good?), loops fixes until both gates pass, then moves to the next task.

"I think the docs are out of date"

/ecogo the deployment runbook looks wrong

The plugin re-reads the cited sources, checks claims against your declared code directories (source-roots), and:

  • Grounded claims stay
  • Unsupported claims get removed (never replaced with invented text)
  • Miscited-but-accurate claims get their citation fixed
  • Contradicted claims get replaced with the source-supported version plus a visible "corrected" note

"Same bug keeps biting the team"

Just keep using the plugin. Every correct, every lint, every compile emits observations. After 5+ observations of the same issue class within 30 days, the learn subsystem drafts a preventive rule. You accept or reject it. Accepted rules apply to every future run. Over weeks, the plugin learns your team's conventions and stops making the same class of mistake.


Using it in an existing project

If you already have a repository of engineering docs, point the plugin at it as a "source-root":

  1. Open ~/ObsidianVault/03-Resources/<your-base>/CLAUDE.md
  2. Find the Source Roots section
  3. Add the path to your code repo, e.g.:
    source-roots:
      - /Users/you/work/auth-service
      - /Users/you/work/internal-docs

Now when correct runs, it greps those directories to validate claims. Accurate-but-miscited becomes a first-class outcome: a claim about an API field that isn't in your RFC but IS in your code gets its citation fixed, not deleted.

This is the single biggest quality-of-life feature for teams where docs and code drift.


Thirteen operations (for when you're ready to go deeper)

The router maps plain English onto these. You rarely need to call them directly, but they're here if you want to.

Operation Purpose
init Scaffold a new knowledge base
migrate Upgrade an older base to the current schema (idempotent)
ingest Save a source (URL or file) — does not process yet
compile Read uncompiled sources, synthesise cross-linked pages, run self-critique
query Answer a question with [[wikilink]] citations
correct Re-ground a page against its sources and your code repos
rationale Extract the why — tradeoffs, alternatives, historical decisions
eval Score the base against a defined question set; track regressions
gap Stage open knowledge gaps from reports and follow-up prose
fetch Pull external sources to close open backlog items
learn Roll up recurring observations into proposed rules
lint Audit for dead links, orphan pages, missing sections
run Composer that chains ops into profiled pipelines (for scheduled runs)
remove Delete a knowledge base cleanly

Plus bundled skills:

  • /ecogo-workflow:brainstorming — design-first workflow with a hard gate
  • /ecogo-workflow:subagent-driven-development — execute a plan via subagents with two-stage review

Running it on a schedule

Let the plugin maintain your knowledge base overnight. In Claude Code:

# Daily upkeep at 3am local time
/schedule create --cron "0 3 * * *" --prompt "/ecogo-workflow:ecogo run"

# Hourly lightweight check during work hours, weekdays
/schedule create --cron "0 9-17 * * 1-5" --prompt "/ecogo-workflow:ecogo run cheap"

# Weekly deep pass Sunday morning (eval, fetch, distill)
/schedule create --cron "0 4 * * 0" --prompt "/ecogo-workflow:ecogo run expensive"

Each trigger fires a fresh Claude Code session, does the work, commits, and exits. The base compounds while you sleep.


Troubleshooting

"The slash command isn't found"

Restart your Claude Code session. Plugins load at session start.

"The plugin says my base isn't at v2"

You created the base before v2 schema. Run:

/ecogo-workflow:ecogo migrate

It's idempotent and purely additive — nothing existing will be deleted.

"I got an error about qmd"

qmd is the search tool this plugin auto-installs. It installs on first session after the plugin is added. If it didn't, reinstall the plugin:

/plugin install ecogo-workflow@ecogo-workflow --force

Then restart your session.

"I'm on Windows and paths look weird"

The plugin uses tilde (~) expansion for home paths, which works on both Mac and Windows PowerShell. If you see a path error, use the full form:

  • Mac: /Users/<you>/ObsidianVault/03-Resources/<base>/
  • Windows: C:\Users\<you>\ObsidianVault\03-Resources\<base>\

"My vault isn't under git"

Some features commit to your vault for history. If ~/ObsidianVault isn't a git repo, those commit steps silently skip — the plugin still works, you just don't get the git diff trail. To add it:

cd ~/ObsidianVault
git init
git add .
git commit -m "initial commit"

"The plugin is doing too much / too little"

Look at your ~/ObsidianVault/03-Resources/<base>/outputs/learnings.md. You'll see proposed rules from the learn subsystem. Accept the ones that match your preferences, reject the ones that don't. The plugin progressively tunes to your style.


What's inside (architecture, for the curious)

  • 13 operations covering the knowledge-base lifecycle
  • Natural-language router that maps plain English onto ops and bundled skills
  • Self-critique loop inside compile — bounded 3-iteration, catches fabrication at creation time
  • Three-way grounding in correct (grounded / miscited / ungrounded / contradicted) with source-roots fallback
  • Learning subsystem — observations → proposed rules → user accept → runtime overlay → schema promotion
  • Rationale extraction — regex pre-pass + LLM judge + promote-to-entity clustering; tradeoff / alternative / historical / gotcha classification
  • Model routing policy — Haiku for mechanical work, Sonnet for standard, Opus for deep reasoning
  • Effort-mode awareness — degrades gracefully in Claude Code's fast mode; full discipline in extended-thinking mode

Full technical details live in the skill file at skills/ecogo/SKILL.md.


Status

  • v0.3.0 — current release: single-word /ecogo invocation, intelligent skill orchestration, model routing, self-improvement loop
  • Full version history in CHANGELOG.md
  • MIT-licensed
  • Bundles the brainstorming and subagent-driven-development skills from superpowers (MIT © Jesse Vincent)

Contributing

Issues and PRs welcome. Every real-world use case surfaces a new signal to tune.

License

MIT. See LICENSE.

Acknowledgments

See ACKNOWLEDGMENTS.md.

About

Self-correcting engineering knowledge base for software teams. ADRs, runbooks, onboarding docs that stay synced with code. Natural-language commands, code-grounded self-correction, rationale extraction, learning-from-mistakes.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages