agentleFS
kbuleonkbuleon/agent-memory-powers/AGENTS.md

agent-memory-powers

AGENTS.md1 starsChanged 33 days agoMITView on GitHub

Project: TO BE COMPLETED * README.md: project overview and repository layout. Maintained by the agent. * CODEMAP.md: map of the repository's code, built from the sources only. Maintained by the agent. * ARCHITECTURE.md: software architecture only — everything there is to know about the project's technical and software architecture. Features do not belong here, nor does CODEMAP.md. This file is self-contained, with no outgoing links. Maintained by the agent when an architectural boundary is crossed, not on every change. *…

  • Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.

    Of the AGENTS.md files we read, the 31,700 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.

  • In this repository3 other files it gives its agents.

The file

We split this AGENTS.md into the rules it gives an agent

3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    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
  1. Very specific:

    When development starts — right after classification for bounded work, once the user approves the 30-plans for architectural work — …

    line 36
  2. Distinctive:

    Strict review: Check each delivery yourself.

    line 40
  3. Very specific:

    Every 10-work-items entry carries a front matter:

    line 45

≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.md files we read have a rule of the same kind, not the same words. “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
AGENTS.md

Project: TO BE COMPLETED

Key files and directories

  • README.md: project overview and repository layout. Maintained by the agent.
  • CODEMAP.md: map of the repository's code, built from the sources only. Maintained by the agent.
  • ARCHITECTURE.md: software architecture only — everything there is to know about the project's technical and software architecture. Features do not belong here, nor does CODEMAP.md. This file is self-contained, with no outgoing links. Maintained by the agent when an architectural boundary is crossed, not on every change.
  • TODO.md: maintained manually by me. The agent does not touch it.
  • CHANGELOG.md: history of software changes. Maintained by the agent. Each entry must be compact and reference a work item if one exists.
  • docs/
    • 00-knowledge-base: general knowledge base about the project, including 010-glossary.md, a domain glossary.
    • 05-draft: drafts in progress. Written by me. The agent neither reads nor modifies them.
    • 07-studies: one-off studies about how the repository works or about general topics. Written by the agent at my request. A study reflects the state of the repository at the date it was written: it is not kept up to date.
    • 10-work-items: analyses of requests. Written by me. Named <prefix>-<nnnn>-<slug>.md. In a Superpowers brainstorming, the agent will be told explicitly which work item to read. The agent does not modify them, except for the front matter. Two families of work items coexist:
      • Local work items, created by the /create-work-item skill, whose prefix is read from docs/work-items-settings.json and whose number is allocated by the creation skill.
      • Work items imported from Jira, created by the /create-work-item-from-jira skill, which carry the Jira project prefix and the Jira issue number padded to four digits. In this case, the agent also writes the body, at creation and only at creation.
    • 20-specs: Superpowers writes its specifications here, named YYMMDD-<slug>.md. For a brainstorming based on a work item, Superpowers names the spec YYMMDD-<prefix>-<nnnn>-<slug>.md.
    • 30-plans: Superpowers writes its execution plans here, named YYMMDD-<slug>.md. If a plan is based on a spec carrying a work item number, Superpowers names the plan YYMMDD-<prefix>-<nnnn>-<slug>.md.
    • 40-adr: Architecture Decision Records. As work goes on, as soon as a decision is adopted, the agent must propose creating an ADR, named YYMMDD-<slug>.md. The user must confirm the creation of the ADR.
    • 50-technical-references: technical reference documents. Each document covers one specific topic. Important: they must be kept up to date by the agent, as soon as the change happens, in a concise, non-verbose way. The agent does not create new documents on its own initiative, but may propose creating one if the topic is not already covered by another document; once the user accepts, the agent creates it.

Who to trust?

  • The source code is the only reliable source of truth.
  • The 10-work-items are useful because they describe the raw request for a change or a fix. But the solution adopted may differ from the initial request. These documents are kept as a record.
  • The 50-technical-references aim to offer thematic technical documents to someone discovering the project. But although the agent must maintain them whenever a change is made, the source code remains the only real truth.

Workflow

  • The user writes their documents in 00-knowledge-base and 05-draft, and maintains their TODO.md.
  • The user creates a new work item with the /create-work-item <title> skill, and the agent updates docs/10-work-items/README.md.
  • The user imports a Jira issue with the /create-work-item-from-jira <key> skill: the agent creates the work item, updates docs/10-work-items/README.md, and adds the label set by jiraLabel in docs/work-items-settings.json to the original Jira issue.
  • The user submits a 10-work-items entry to Superpowers brainstorming.
  • Superpowers iterates on the thinking with the user and classifies the work item in its superpowers-class field. The agent then sets its status to in-progress.
  • What follows depends on the class:
    • spike: the answer is given to the user, and the agent sets the work item to done.
    • bounded: Superpowers moves straight on to development.
    • architectural: Superpowers produces a specification in 20-specs, and the agent updates docs/20-specs/README.md. When the user approves the 20-specs, Superpowers produces an execution plan in 30-plans, and the agent updates docs/30-plans/README.md.
  • When development starts — right after classification for bounded work, once the user approves the 30-plans for architectural work — Superpowers follows these instructions:
    • Sequential only: One task at a time. No parallel runs.
    • Clean context: Give sub-agent only needed context with specs.
    • Adaptive effort: Match model power and reasoning depth to task difficulty.
    • Strict review: Check each delivery yourself. Reject if specs or edge cases fail. Fix before next step.
    • Compact memory: Summarize validated output into main context. Drop intermediate logs.
  • As decisions are made along the way, the agent creates an entry in 40-adr after the user confirms, and updates docs/40-adr/README.md.
  • As development progresses, the agent updates the relevant topic document in 50-technical-references; if it proposed creating a new one and the proposal was accepted, the agent updates docs/50-technical-references/README.md.

Work item front matter

Every 10-work-items entry carries a front matter:

---
id: <prefix>-<nnnn>
title: "The work item title"
status: new
superpowers-class: 
tags: []
---

The id prefix is configured in docs/work-items-settings.json. The prefix may also be free: typed in manually, or imported from Jira. An issue imported from Jira carries two extra marks: tags contains Jira, and a jira field holds the exact issue key — PROJ-558 for a work item PROJ-0558, the local number being padded to four digits while the Jira key is not. Its title is the Jira issue summary as is, followed by the summary of its parent in brackets when the issue has a parent. The front matter of an imported issue reads:

---
id: PROJ-0558
title: "The Jira issue summary [The parent summary]"
status: new
superpowers-class: 
tags: [Jira]
jira: PROJ-558
---

The possible status values are:

  • draft: set manually by the user.
  • new: status after the work item is created by the /create-work-item or /create-work-item-from-jira skill.
  • deferred: set manually by the user.
  • in-progress: set by the agent once Superpowers has classified the work item.
  • review: set by the agent when Superpowers reaches the code review step.
  • done: set by the agent at the final commit of the development — or, for a spike, once the answer has been given.
  • canceled: set manually by the user. The superpowers-class must be assigned by Superpowers during its classification, and may take the values:
  • spike: a question whose output is an answer, not code.
  • bounded: a well-scoped change in code that already exists in the repository.
  • architectural: a new project, a new subsystem, or an architectural change. The tags are updated manually by the user, but the agent may suggest some. The agent modifies only the front matter of a work item, never its body — with one exception: the import skill writes the body of a Jira issue at its creation, and only at its creation. After that, an imported issue belongs to its author like any other.

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.