Skip to content

Where this sits relative to what you already run

Most of these are good at what they do, and several are worth using alongside Codee3 rather than instead of it. What follows is what each genuinely solves, and the specific thing it leaves open.

12 approaches, by relationship

Complements9of 12
Keep running it. It does something Codee3 does not do, and Codee3 governs a layer it does not reach. The two are used together.
Replaces2of 12
The same need, met a different way. Running both leaves two accounts of the same thing, and one of them is ungoverned.
Unrelated1of 12
Not a substitution decision at all. It answers a different question, and choosing between the two is a category error.

The relationship is this page’s reading of the two sentences beside each row, not a claim any of these tools makes about itself. Where a row is judged a complement, the recommendation is to keep running it.

01 · Your layer

Pick what you already run

Grouped by what the alternative is actually for, because “Codee3 versus everything” is not a real comparison. Most of what follows is a complement, and the two substitutions are named rather than implied.

Relationship to Codee3

Mostly complements

3 complement · 1 replaced

Keep writing where you already write. What the governed corpus adds is a declared authority grade per document and a merge-time check that the structure still holds — not another place to put prose.

  • READMEs and /docs

    Complements

    Keep them. A README is still the fastest way to tell a person what a directory is for. What it cannot do is mark one statement as binding and another as superseded — and that grading is what the governed corpus adds around it, not instead of it.

    What it solves
    Low-friction documentation, right beside the code, with no tooling at all.
    What it leaves open
    No graded authority and no structural enforcement, so it decays by default. An agent cannot distinguish a binding decision from an outdated proposal.
  • Wikis, Notion, Confluence

    Complements

    Keep them for the organisational knowledge that is not about one repository; the editing ergonomics are genuinely better. What has to move into the repository is the part that needs to branch with a change, be reviewed in the same pull request, and be readable from the working tree.

    What it solves
    Shared, searchable, well-formatted knowledge with good editing ergonomics.
    What it leaves open
    Lives outside the repository, so it cannot branch with a change, be reviewed in the same pull request, or be read by a tool that only has the working tree.
  • ADRs (Architecture Decision Records)

    Complements

    The clearest complement on the page, and compatible with everything here. Codee3 does not replace a decision record. It grades one, and registers the references to it so that a rename does not quietly break every mention.

    What it solves
    Captures decisions with their rationale, in the repository, in a format teams already understand. Genuinely good, and compatible with everything here.
    What it leaves open
    One category of knowledge only. No authority hierarchy across the wider corpus, and no structural guard against reference drift when a decision is renamed or superseded.
  • AI documentation generators

    Replaces

    The same need, met the opposite way. Both exist to give an agent repository context; one projects it from the code, the other requires a declared source and an authority grade before it counts as context at all. That is a substitution, not a stack.

    What it solves
    Very low cost to produce a large volume of documentation quickly.
    What it leaves open
    Projection with no declared source and no authority grading. The available evidence on generated repository context is not encouraging — it can raise cost and reduce task success while looking entirely plausible.
The full comparison as one table
All 12 approaches, grouped by what each one is for. Every row states what the approach solves before what it leaves open; the last column is the relationship to Codee3 derived from those two.
ApproachWhat it solvesWhat it leaves openRelationship
Documentation Where knowledge is written down.
READMEs and /docsLow-friction documentation, right beside the code, with no tooling at all.No graded authority and no structural enforcement, so it decays by default. An agent cannot distinguish a binding decision from an outdated proposal.Complements
Wikis, Notion, ConfluenceShared, searchable, well-formatted knowledge with good editing ergonomics.Lives outside the repository, so it cannot branch with a change, be reviewed in the same pull request, or be read by a tool that only has the working tree.Complements
ADRs (Architecture Decision Records)Captures decisions with their rationale, in the repository, in a format teams already understand. Genuinely good, and compatible with everything here.One category of knowledge only. No authority hierarchy across the wider corpus, and no structural guard against reference drift when a decision is renamed or superseded.Complements
AI documentation generatorsVery low cost to produce a large volume of documentation quickly.Projection with no declared source and no authority grading. The available evidence on generated repository context is not encouraging — it can raise cost and reduce task success while looking entirely plausible.Replaces
Retrieval How knowledge is found.
RAG and vector searchFinding text that resembles a query, across a large corpus, quickly.Cannot tell you whether what it found is current, superseded, or ever was a decision. Retrieval answers where; governance answers whether you may rely on it.Complements
Repo maps, code graphs and agentic grepNavigating source code structure — often better and more cheaply than semantic indexing.Reads the implementation, not the organisational knowledge about it. No amount of code analysis recovers a decision that was never written down.Complements
Agent configuration What a specific tool is told.
Per-tool rules files (AGENTS.md, CLAUDE.md, .cursor/rules)Gives one tool grounding for one repository, with no dependencies and no cost.Each tool keeps its own copy and they drift apart. No declared authority within a file, and no structural enforcement across several.Replaces
Instruction-file lintersValidating a single instruction file — paths that do not resolve, commands that do not exist, missing sections, stale dependency references. Free, focused, and worth running.Scoped to one file rather than a governed corpus. No authority model across documents, no reference registry with rename history, and no routing discipline.Complements
Agent memory featuresPersisting facts across sessions inside one tool, with no setup.Per-tool and typically per-person, not version-controlled, not reviewable in a pull request, and not shared with the rest of the team or the other tools.Complements
Platform and enforcement What is measured, and what is gated.
Internal developer portals and scorecardsCataloguing services, ownership and dependencies at organisation scale, with dashboards leadership can actually use.Standards are typically evaluated in the vendor backend on a schedule and surfaced as a score. That is visibility, which is valuable and different from a gate on your commit.Complements
Code linters, static analysis and policy enginesGating code at merge time — style, defects, vulnerabilities, licence policy. Many of these genuinely fail builds, and should.They gate the code. None of them gate the structure of the repository’s own knowledge layer, because that is not what they are for.Complements
Agent frameworks and orchestrationRunning, chaining and supervising agents.Agents still need repository knowledge to reason well. A framework moves work around; it does not supply the knowledge the work depends on.Unrelated
02 · The distinction

The one distinction worth keeping

Plenty of tools fail builds — linters, static analysis, policy engines, security scanners. They gate the code. Portals and scorecards measure standards and show them on a dashboard, on their own schedule. What nothing on this page does is gate the structure of the repository’s own knowledge layer at merge time, which is the narrow thing Codee3 is for.

Gate the code
Linters, static analysis, policy engines, security scanners. Many of them genuinely fail builds, and should.
Measure the standard
Portals and scorecards, evaluated on a schedule and surfaced as a score. That is visibility, and it is valuable.
Gate the knowledge layer
Codee3, at merge time, on the structure of the governed .ai/ directory — and on nothing outside it.
03Honestly

When you should not buy this

If an instruction-file linter and a habit of writing ADRs covers what your team needs, that combination costs nothing and you should use it. Codee3 is worth considering when the corpus is large enough that authority across documents, reference integrity across renames, and one destination for several agent tools have all become real problems.

That is the whole test. If authority across documents, reference integrity across renames and one destination for several agent tools are not yet real problems for you, then neither of the two substitutions above is worth making, and this page has done its job.