customize.toml
# DO NOT EDIT -- overwritten on every update.
#
# Workflow customization surface for bmad-project-context.
# Team overrides: {project-root}/_bmad/custom/bmad-project-context.toml
# Personal overrides: {project-root}/_bmad/custom/bmad-project-context.user.toml
#
# Merge rules: scalars override (last layer wins); arrays append.
[workflow]
# --- Universal defaults ---
activation_steps_prepend = []
activation_steps_append = []
# Deliberately empty: this skill's own output (AGENTS.md) is loaded by the
# harness, not through this array. Users append their own standing facts.
persistent_facts = []
on_complete = ""
# Standing outside-the-repo sources offered at every setup/refresh run
# (untrusted until verified against the repo or user-confirmed).
# Append-only. Entries: "file:{project-root}/..." or "file:/abs/path" for
# docs, "skill:name" to consult a skill, plain text for a standing fact,
# "tool:name" for an MCP knowledgebase.
external_sources = []
module-manifest.toml
module = "method"
version = "6.13.0-next"
update_source = "github:bmad-code-org/BMAD-METHOD/skills"
knowledge = "`references/help.md` in the `bmad` skill"
references/best-practices.md
# What belongs in a repo's agent instructions
Rules for deciding what goes in the block, for judging what a repo already has, and for explaining both to the user.
## The test
Not *could an agent derive this* but *what does it cost when it doesn't*: how much exploration finding it takes, how likely the agent is to search the right place in time rather than guess, whether it is available at the point of use or only after the mistake, what a retrieval failure costs — a wasted search, or corrupt data — and whether it is a rule that must hold or a detail the code already shows.
A line that stops the same rediscovery every session earns its place, derivable or not. A stored copy of what the agent reads more accurately first-hand does not — it rots, and it is charged every session.
## Admit
- **Policy the code cannot express** — branch rules, frozen and protected paths, generated files, secrets, security and compliance. Stated by a human or read off an enforcing config, never inferred.
- **What a config file cannot say about running the project** — the root test script does nothing in this workspace, integration tests need a service up first, the suite takes eleven minutes so iterate on single files, the `Makefile` is the real entry point and `package.json` is vestigial, CI runs a typecheck the test script does not. An invocation the obvious guess gets right is already stated in `package.json`, `Makefile`, `pyproject.toml`, or CI config and does not earn a line — the correction, the caveat, and the right command to use do.
- **Conventions that differ from ecosystem defaults.** An agent follows the norm unless told otherwise, so only the divergences earn a line. Command invocations count: when the obvious command is wrong here — a bare-repo prefix, a required wrapper — the exact working invocation earns a line, and no observed mistake is needed to admit it.
- **Pitfalls with observed evidence** — a recorded lesson, the maintainer's recollection, the same mistake fixed repeatedly in history, or one this session made and caught. A repo yields hundreds of trap-looking facts and none of them predict real mistakes; only observed behavior does. A surprising scan finding is a question to ask, not a line to write.
- **Runtime behavior invisible from the repo** — replaying webhooks, lying health endpoints, environment quirks — once a human confirms it.
- **Cross-component rules**, admitted when getting one wrong in one file breaks something elsewhere — what must stay true across parts of the system the agent cannot see from the file it is editing: who owns what, how data must flow, what order a pipeline runs in. "Writes go through the dispatcher; direct store mutation skips the transaction." "The importer is two passes — validate every row, then commit; never write inside the parse loop." A six-line map of who owns what. Never an inventory written for completeness; the exclusions below still bind.
- **Required tool and runtime versions**, read from the project files that declare them, never from this session's environment — which answers faster, and wrongly, so the mistake arrives before the search.
- **Entry points and pointers** to where work lands.
Prefer prohibitions to advice, and name the permitted alternative in the same line.
## Exclude
| | Why |
|---|---|
| Repo overviews, directory trees, stack lists | Derived fresh, more accurately; stored copies rot |
| Anything included for being interesting | Interest is not need |
| Style rules an agent self-enforces | Belongs in a formatter, linter, hook, or CI check — propose the check instead |
| Platitudes | Already the default |
| Transcribed command lists whose obvious invocation is already right | Read from `package.json`, a `Makefile`, or CI config; a copy drifts the moment a script is renamed. The right command to use, and any command the obvious guess gets wrong, are admitted above |
| Pasted code, changelog content, fast-changing facts | Stale immediately |
| Aspirational state | Describe what is; intent belongs in specs |
| History and edit narration | Git holds it; state present truth |
## Retire
A policy or pitfall goes only when the thing it guards is gone, or the user retires it. Nothing failing lately is not evidence — a working rule erases its own evidence. Any other existing instruction goes only on one of the four grounds under "Judging an existing file".
Every line faces one question at each write: would removing it change agent behavior? If no, cut it — but for a line a human wrote, that answer only opens a candidate; a ground still has to carry it.
## Size
Every line is paid in every session, and instruction-following degrades as the loaded set grows. Count what other always-loaded files add. Over budget means cut the weakest lines or move them behind a trigger — never raise the budget. Ten lines of evidence means ten lines.
An adopted file must fit the budget too, but shrinking it works differently. Move the weakest instructions out first — into a child file, a linked doc, or a hook or check that enforces them. Deleting still needs one of the four grounds. If the file is still too big and no ground justifies another deletion, show the user and let them decide — an over-budget file they chose beats a gutted one they didn't. "Keep it small" disciplines what this skill writes, never what the maintainer already wrote.
## Retrieval
An index the agent must choose to fetch gets skipped; one already in context does not. Keep everything load-bearing in the block. A pointer out of it names a trigger the agent can observe — a path, a file type, a named task — never one it must judge ("when the task is complex") or track about itself ("before your first edit").
Rules bounded to a directory can go in a nested `AGENTS.md` there, attached by location rather than by pointer — but only when they are subtree-exclusive and substantial, the split materially reduces the root block, the user approves it, and **loading is verified for every harness in use**. Even with verified loading, keep a rule at the root when it must apply before a session enters that directory or when breaking it can affect work outside the child. Check, never assume: several harnesses build the instruction chain once at session start, root down to the working directory, so a nested file is invisible to the session that later edits into that subtree. Unverified means path-qualified lines at root instead — "in `src/importer/`: ..." — cheaper than a file nobody loads.
Use a linked file only when the trigger is not a path.
## Maintain
- Re-check that caveats still hold — a slow suite that got fast, a workaround for a bug that was fixed.
- Diff deletions and renames since the verified SHA against every line.
- Record provenance in the block so the next run knows what it is diffing from.
- Capture mistakes when they happen, not at review time. One occurrence is a note; recurrence earns a line.
- Route anything mechanically preventable to a hook, lint rule, or CI check. A check that lands deletes its line.
## Repo or home directory
This block belongs committed: shared by the team, consistent across machines, versioned with the code it constrains.
Two things belong in the user's global agent config instead — rules repeating across all their projects, and personal preferences that are theirs rather than the team's.
## Judging an existing file
Every instruction a human wrote is presumed intentional: someone paid for it, usually by watching an agent fail. The file is the baseline being improved, not raw material. Keep its phrasing where it works, and carry each instruction through a ledger entry — `retain | rewrite | relocate | automate | delete`, opened at retain or rewrite — so the user sees where all of it went.
**Deletion needs one of four grounds:**
1. **Stale or incorrect** — the referent is gone, or the instruction was never true; the evidence is named.
2. **Mechanically enforced** — a hook, linter, formatter, or CI check already fails the violation named by the instruction. A tool that only covers the same files or topic does not enforce the instruction.
3. **Harmful or contradictory** — it points agents at the wrong thing, or it contradicts another live instruction and loses the reconciliation.
4. **The user approved this deletion** — asked as a line item, never implied by approving a replacement block.
Grounds 1–3 are evidence the run carries itself, and ride the block approval; ground 4 is the ask-first path everything else takes. Nothing else deletes. Brevity is not grounds, nothing failing lately is not grounds, "the agent could derive it" is not grounds, and **"it is discoverable somewhere in the repository" is never, alone, grounds** — that is the reasoning that empties good files. Content the exclusions table rejects — a directory tree, a stack list, pasted code — has no ground of its own: propose the deletion and let it land under ground 4, asked rather than assumed.
Report, in this order: what is unverifiable or stale, what is missing against the sections above, what is already good, and the ledger, every relocation, automation, and deletion itemized. Recorded lessons are maintainer testimony — kept by default, challenged only with evidence that the thing they name is gone or wrong.
references/template.md
# Block shape
Sections in this order. Omit any section with nothing that passes its rule — never write an empty one. Admission rules: `best-practices.md`.
1. **Orientation** — three or four sentences: what this is, the stack, where planning and deeper docs live.
2. **Policy** — what the org requires.
3. **Where things are** — entry points, and pointers to children and linked files.
4. **Running and verifying** — the right commands to run and the required tool versions, plus what `package.json`, `pyproject.toml`, a `Makefile`, or CI config does not already say.
5. **Conventions that differ from defaults**
6. **Known pitfalls**
Terse imperative lines under plain headings. No prose beyond Orientation, no introduction, no summary. A bare fact appears only as the justification clause of an instruction — "Exclude `vendor/` from searches, it is 60% of tracked files", never "`vendor/` is 60% of tracked files". A prohibition names the alternative. At most two emphasis markers in the whole block.
## Worked example
````markdown
<!-- bmad:context -->
<!-- Verified 2026-08-08 against a1b2c3d. Managed by bmad-project-context; edits inside this block are replaced on refresh. Keep anything you want preserved outside the markers. -->
## acme-billing
Payment processing for Acme storefronts. TypeScript/Node, pnpm, Postgres. Planning lives in `docs/planning/`, tickets in Linear (ACME board).
## Policy
- Never push to main; PRs only, one approval.
- Never modify `legacy/` — frozen, being replaced. New work goes in `src/`.
- Never hand-edit `src/generated/` — run `pnpm codegen`.
## Where things are
- Webhook handling: `src/routes/webhooks.ts`; conventions in `docs/webhooks.md`
- Writing a migration? Read `docs/db-rules.md` first — ordering, transaction boundaries, pool limits.
- Billing service has its own guide: `services/billing/AGENTS.md`
## Running and verifying
- Run single test files while iterating; the full suite takes ~11 minutes.
- Integration tests need `docker compose up -d` first, and fail confusingly without it.
- CI also runs `pnpm typecheck`, which `pnpm test` does not cover.
## Conventions that differ from defaults
- Money is integer cents (`amountCents`), never floats — `src/lib/money.ts`
- All DB access goes through repositories in `src/repos/`; never call the client directly.
## Known pitfalls
- Stripe webhooks replay in staging every 6h — handlers must be idempotent.
- Use vitest matchers, not jest — agents repeatedly add jest syntax here.
<!-- /bmad:context -->
````
Fill the provenance line with the real date and the commit SHA verified against. Refresh diffs from that SHA.
SKILL.md
---
name: bmad-project-context
description: 'Set up, adopt, refresh, or audit a repository''s agent instructions (the AGENTS.md block) so AI agents work well in that repo. Also records observed agent mistakes as pitfalls. Use when invoked by name'
---
# Overview
A conversation that produces a repository's agent instructions: a small verified block inside `AGENTS.md`. The user brings rules they want followed — governance, security, standards — and the repository supplies the rest, verified.
Conversational always; the user approves every write.
**Args:** intent (`setup` | `adopt` | `refresh` | `record` | `audit`); a target repo or path; extra source paths or URLs.
## Resolution rules
- Bare paths and `{skill-root}` (e.g. `references/best-practices.md`) resolve from this skill's installed directory.
- `{project-root}` → the project working directory.
- **Target** → the repository being described, defaulting to `{project-root}`. If it resolves to more than one working tree, or to one the user cannot commit in, ask before writing.
## On Activation
1. Resolve customization: `uv run {project-root}/_bmad/scripts/resolve_customization.py --skill {skill-root} --project-root {project-root} --key workflow`. On failure, read `{skill-root}/customize.toml` directly and use defaults. Execute `{workflow.activation_steps_prepend}`; treat `{workflow.persistent_facts}` entries as standing context (`file:` = paths/globs to load, others verbatim).
2. Config: if `{project-root}/_bmad` exists, `uv run {project-root}/_bmad/scripts/resolve_config.py --project-root {project-root}` and read `{output_folder}`. Standalone: skip.
3. **Load `references/best-practices.md` and `references/template.md` before anything else.** Every decision below is made against them.
4. Detect intent and greet the user: **setup** (no instruction file in the target carries meaningful content — scaffolding alone, empty headings, a comment, a lone import line, is not meaningful; when unsure, adopt, since adopting a near-empty file costs one small ledger while setting up a meaningful one loses instructions), **adopt** (an instruction file has content but no managed block, whatever its state and whoever wrote it — the migration form of refresh; that file is the baseline and every instruction in it enters the ledger of step 1), **refresh** (a managed block exists), **record** (the user reports a mistake agents made), **audit** (re-verify and prune). A supplied intent that contradicts what detection finds — e.g. `setup` against a file with content — is surfaced and confirmed, never silently obeyed. Fold `{workflow.external_sources}` into the source list. Execute `{workflow.activation_steps_append}`.
## Setup, Adoption, and Refresh Steps
No writes until step 5!
### 1. Assess and report
Read `AGENTS.md`, harness or agent specific rule files, docs folders, and any notes carrying lessons. Report what exists and how it measures up, per `best-practices.md`.
Existing instructions are the baseline being improved, never raw material to discard. Open a **ledger**: one entry per existing section and per independently meaningful instruction, opened at `retain` or `rewrite`, carrying what an agent would get wrong without it. Entries settle as evidence arrives in steps 2–4 — `retain | rewrite | relocate | automate | delete`, each with its reason, its evidence, the risk if it goes, a destination for a relocation, and an approval flag. Deletion needs one of the four grounds in `best-practices.md`, and a relocation destination must itself be loaded or sit behind an observable trigger — a move into a file nothing reads is a deletion and needs its ground. Setup has nothing to map and opens no ledger; refresh opens entries for the lines it proposes to change or remove, the block's own included. A lesson found outside the instruction files — a warning in a README, a notes file — is an ordinary candidate, not a ledger entry.
If the target contains separable units — a workspace manifest listing members, or directories carrying their own build manifest — name them and ask whether this run covers the root only, all of them, or which. Absent that evidence, do not ask. Sibling repositories are not children; each is its own target, offered in turn.
### 2. Ask what they bring
Rules to follow regardless of what the repo does: governance, security and compliance, coding standards, style guides, frozen areas. Ask for outside documents too — handbooks, wikis, architecture docs, MCP knowledgebases. Note the paths; do not read them yet.
Greenfield: this is the whole content. Brownfield: it is the half no scan reaches.
### 3. Discover and verify
Fan out with parallel subagents against what the sections need — executable config and CI for policy and for what they already state, tracked source for conventions and boundaries, targeted history for constraints whose reason must still hold.
`package.json`, a `Makefile`, `pyproject.toml`, contribution guides, pull request templates, and CI config are read to know what the block must not repeat. Their caveats come from the human in step 4. Path-check every claim naming a file. For every claim the block will make about what a command does, read the target or script that runs it and verify the claim.
Each child agreed in step 1 is scanned as its own scope, against its own manifests.
### 4. Interview the gaps
Only what no scan reaches: what agents keep getting wrong here, what is off limits, what a domain term means, why a constraint exists.
- Never ask what a scan could answer. Asking the user to confirm a path-checked claim, or one a config file already states, is a defect.
- Ask recall questions, not review lists. Never hand the user a selection problem a scan created.
- A mistake this session made and caught is observed evidence — offer it.
- A repeatable command spotted in anything read this session — a log, a doc, its own runs — whose correct form is not the obvious guess is a candidate line: offer it. E.g. `uv run pytest` where plain `pytest` looks right but runs outside the project environment.
- Batches of at most eight; fewer is better. A batch yielding nothing new means write.
- When the repo contradicts the user, show the evidence and ask. Never write the claim as given, never drop it silently.
### 5. Show the block, then write it
Compose against `template.md`. For each candidate, ask first whether a hook, lint rule, or CI check enforces it better than prose; if so propose the check, and the line becomes the fallback if they decline. A ledger entry marked `automate` keeps its instruction until its check is in place (a later run deletes the line under ground 2 once the check is live).
**Show the complete block before writing it**, and every child block alongside it — one approval covers the set. **Present the settled ledger with it**: replacement text alone is an incomplete proposal, because it shows what the user gains and hides what they lose. Every existing instruction appears with its decision and reason. Retains and rewrites that keep the full rule may be grouped. If a rewrite weakens, narrows, or drops part of a rule, treat the lost part as a deletion and list it separately. Keep the rule itself; examples may explain it but cannot replace it. Every relocation, automation, and deletion is itemized. A deletion resting on none of the first three grounds is held for line-item approval — approving the block never approves it — and a declined deletion, relocation, or automation reverts to retain. On approval, splice between the markers — the splice itself touches nothing outside them. Text outside the markers changes only through a settled ledger entry or a proposed fix the user has seen, never as a side effect of the splice. Fill each provenance line with today's date and the verified SHA.
Where an instruction elsewhere contradicts the block in a way that changes behavior — a stale `CLAUDE.md` line, a retired command — propose the fix to that file. Two live contradictory instructions is a defect.
Never commit.
### 6. Close
- What went in, what was left out and why, and — after adoption or refresh — where each existing instruction landed.
- Why, in the user's terms, from `best-practices.md` — why it is small, why what the repo already states stays out, why a pitfall stays until its cause is gone.
- How it loads, and that other harness files can point at it.
- Any branch, ticket, commit, or pull request rules that apply when the user submits these instruction changes.
- Maintenance: re-run after significant change, `record` the moment an agent gets something wrong, prefer a check over a new line.
- Rules repeating across their projects, or personal rather than the team's, belong in their global agent config.
### Refresh
Same steps, step 1 as a diff. Read the provenance line, re-verify every path and every caveat, and run `git log --diff-filter=DR --name-only` since the recorded SHA against every line — update or remove lines whose evidence is gone. Every proposed removal is a ledger entry shown in step 5, never a silent edit, and handwritten instructions outside the block are treated as in adoption — any proposal touching them enters the ledger. Never re-ask what a prior run settled; the interview shrinks to what changed about how the team works. The block grows only on new evidence.
### Adoption
Refresh against instructions this skill has never touched. Nothing was settled by a prior run, so the full interview applies — and the file itself is maintainer testimony, so the ledger is the run's main output: the user should be able to read it and see where each of their instructions went.
The proposal states what remains of every file instructions were moved out of — commonly a `CLAUDE.md` reduced to `@AGENTS.md`, once that import is verified for every harness in use, like any loading mechanism. No instruction lives in two loaded files, where it is paid for twice; a duplicate, verbatim or reworded, is kept once — the block keeps the survivor — and that settles both entries.
### Greenfield
Seeded from a spec or planning document, or interview alone. Commands that do not exist yet are written as explicit TODOs naming the decided stack, never a guessed invocation stated as fact, and verified on the first refresh after code exists. A genuinely contested design decision — real tradeoffs, multiple viable shapes — goes to `bmad-architecture`.
### Migration
If the target has a `project-context.md` from the retired skills, commonly under `{output_folder}`, read it in step 1 and offer to absorb its content. Do not delete it without agreement, and do not silently orphan it.
## Record
Capture one observed agent mistake as it happens — the only admissible source for a pitfall.
Take the task, the mistake, the correction, and its evidence. Check the block for a line already covering it. One occurrence is noted; a recurring or costly mistake earns a line now — an exact invocation under **Running and verifying** when it is a command error, otherwise a pitfall. Write it and show the diff. If it is mechanically preventable, propose the hook, lint rule, or CI check instead.
## Audit
Re-check every caveat, path-check every file, follow every pointer, and ask of every line whether removing it would change agent behavior. Verify each command claim against the target or script that runs it. Check for contradictions with other instruction files.
Failing lines get fixed, move behind an observable trigger, or become ledger entries: a removal needs one of the four grounds in `best-practices.md`, presented and settled as in step 5 before anything is removed. **A policy or pitfall goes only when the thing it guards is gone or the user retires it; nothing failing lately is not grounds.** Audit ends smaller or equal.
## Children
A component, nested repository, or extracted rules file gets its own file under the same shape when work keeps landing there and every condition holds: its rules are subtree-exclusive, they are substantial (a handful of rules is not a file), the split materially reduces the parent block, the loading mechanism is verified for every harness in use — checked, never assumed — and the user approves the split. Even with verified loading, keep a rule at the root when it must apply before a session enters that directory or when breaking it can affect work outside the child. Otherwise the rules stay in the parent block as path-qualified lines ("in `src/importer/`: ..."), which cost less than a file nobody loads. Why the loading check: `best-practices.md`.
Use a linked file only when the trigger is not a path.
A chosen child that ends with nothing its parent does not already say gets no file. Say so and move on.
List every child in the parent's **Where things are** with one line and its path. Discovery never depends on the harness finding it.