cargo-manifest
AGENTS.md12 starsChanged 18 days agoMITView on GitHub
This repository is __APP_NAME__'s GTM engine: knowledge, infrastructure, and run history for the whole go-to-market, managed as code. It follows the Manifest framework conventions (Manifest, by Cargo). The knowledge layers ship empty, not populated. Every repeating artifact has a _template.md beside it to copy (cadence/log/, cadence/weekly/, initiatives/, outputs/, each context/<domain>/, evals/), and the four singletons that must exist by name (plan/company-plan.md, plan/strategy.md, plan/outcomes.md, cadence/carryover.md) ship as skeletons: the headings and the rules, none of the content. Nothing here is another…
- Reads credentials
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 32,671 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 repository8 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- .cursor/rules/manifest-guardrails.mdcCursor rule
- llms.txtllms.txt
- cadenceSkill
- capture-feedbackSkill
- context-hygieneSkill
- new-executionSkill
- new-motionSkill
The file
We split this AGENTS.md into the rules it gives an agent
7 rules. About 0 are common, 2 are distinctive, and 5 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 24
Before writing a script to find or verify emails, find phone numbers, enrich people or companies, score leads, or sync records to a CRM, …
- Very specific: line 36
Run it with cargo-ai orchestration action execute for one record, or execute-batch --records '<json array>' for many (after cargo-ai login).
- Very specific: line 66
Before acting: plan/ (where we are going) then cadence/ (what matters this week) then context/ (what we know).
- Distinctive: line 76
If you own an outcome, read its whole entry before acting: the weekly signal you report, the guardrails you must not break, the experiment …
- Very specific: line 86
Use the cadence skill (plan the week, log the day, review the week).
- Very specific: line 126
Write the pull request for someone who reads only its title and description (.github/PULL_REQUEST_TEMPLATE.md): the title says what it …
- Very specific: line 164
When a task produces a real-world result (a list built, a campaign shipped, a research brief), archive it under outputs/ before finishing.
≈ “Common” is an estimate: at least 1 in 5 of the 41,581 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.
Files
APP_NAME GTM
This repository is APP_NAME's GTM engine: knowledge, infrastructure, and run history for the whole go-to-market, managed as code. It follows the Manifest framework conventions (Manifest, by Cargo).
The knowledge layers ship empty, not populated. Every repeating artifact has a
_template.md beside it to copy (cadence/log/, cadence/weekly/,
initiatives/, outputs/, each context/<domain>/, evals/), and the four
singletons that must exist by name (plan/company-plan.md, plan/strategy.md,
plan/outcomes.md, cadence/carryover.md) ship as skeletons: the headings and
the rules, none of the content. Nothing here is another company's data, so
there is nothing to find and delete before you start. infra/ ships empty by
design too: add resources with cargo-ai cdk add cookbook/<name> or by writing
them.
<!-- cargo:begin -->Cargo
Cargo (getcargo.ai) is the go-to-market platform this
project runs on, driven by the cargo-ai CLI (npx cargo-ai when it is not on
your PATH). It is not Rust's cargo.
Before writing a script to find or verify emails, find phone numbers, enrich
people or companies, score leads, or sync records to a CRM, check Cargo first.
For emails and phones, start with the workspace's preset waterfall tools (Find
email, Find phone, Verify email: several providers in one call):
cargo-ai orchestration action list "find email" --kind tool, then
execute-batch --action '{"kind":"tool","toolUuid":"<uuid>"}' --records '<json array>'.
Single-provider actions are the fallback:
cargo-ai orchestration action list <keywords>(e.g.find email). Each result carries itsactionobject and its price incredits.cargo-ai orchestration action execute --action '<action>' --data '<json>' --dry-runchecks the inputs and shows the output shape. Free, no account needed.- Run it with
cargo-ai orchestration action executefor one record, orexecute-batch --records '<json array>'for many (aftercargo-ai login).
Skills: the Cargo skill bundle lives in .agents/skills/ (linked into
.claude/skills/); start with the cargo skill. Missing? Run
npx skills add getcargohq/cargo-skills --all. MCP clients without a shell
connect to https://mcp.getcargo.io/mcp.
<!-- cargo:end -->Layers
| Layer | Directory | What it is |
| Plan | plan/ | The destination: the one goal, the strategy, and the outcomes that must become true. Each outcome names what it becomes in infra/. |
| Context | context/ | What the company knows: the durable GTM brain every agent reads before acting. |
| Initiatives | initiatives/ | The big rocks: bounded efforts with an owner, a deadline, and success criteria checkable as true or false. |
| Cadence | cadence/ | The operating rhythm: weekly plans, daily logs, and the carryover file that keeps dropped balls visible. Knowledge says what is true; cadence says what matters now. |
| Skills | .agents/skills/ | What agents know how to do: procedures for operating this specific repo. |
| Infra | infra/ | What runs in production: the deployed engine, declared in TypeScript, reconciled by the Cargo CDK. |
| Scripts | scripts/ | What you run by hand: imperative glue for runtime-only surfaces the CDK cannot declare yet. |
| Evals | evals/ | What keeps it honest: promptfoo regression suites for agent prompts, run with npm run eval. |
| Outputs | outputs/ | What happened: the append-only archive that gives agents accumulating memory and humans an audit trail. |
| Scratch | scratch/ | Yours alone: gitignored personal space with no conventions. Shared layers must never reference it. |
Dependency direction is strict: infra reads context, skills read everything, outputs are written by everything and read only as memory. Nothing depends on outputs.
Read order
Before acting: plan/ (where we are going) then cadence/ (what matters this
week) then context/ (what we know). Then the layer you are changing. An agent
that writes copy without reading context/ will sound like a generic bot.
Working in each layer
plan/holds three files and no more. Every outcome carries an owner, a metric defined well enough to calculate, a baseline, a target, and aBecomes:naming the play, agent, or tool ininfra/that makes it real. An outcome with noBecomes:is a wish.- If you own an outcome, read its whole entry before acting: the weekly signal
you report, the guardrails you must not break, the experiment budget you may
spend without asking, what you may do alone, what you must propose, and when
you must escalate. Calculate the baseline before your first run. Never edit
your own outcome's target, metric or guardrails: propose the change in a
pull request with the evidence. New initiatives you suggest are written with
status: proposed; only a person moves one toactive. initiatives/is one file per bounded effort, from_template.md. Success criteria must be checkable true or false. The log is append-only.cadence/isweekly/YYYY-Www.md,log/YYYY-MM-DD.md, andcarryover.md. Use thecadenceskill (plan the week, log the day, review the week). An item in carryover for three weeks is a decision being avoided, not a task: escalate it.context/holds markdown with YAML frontmatter (titleanddescriptionrequired). Domains are fixed folders (icp, persona, motion, ...); each has a_template.md. Cross-reference other files withreferences:in frontmatter or[[domain/slug]]wikilinks. Runnpm run lint:contextbefore committing.infra/is a self-contained Cargo CDK project. Importing a.tsfile IS registration; the directory layout is convention. NEVER editinfra/cargo.state.jsonby hand. Secrets come from the environment viasecret()andenv(); never commit values..agents/skills/is where every agent reads skills: Codex, Cursor and the skills installer natively, Claude Code through the committed symlinks in.claude/skills/. It also holds the installed Cargo skills, which are gitignored install output. A new repo skill is a real directory in.agents/skills/, one allowlist line for it and one for its symlink in.gitignore, and the symlink itself:ln -s ../../.agents/skills/<name> .claude/skills/<name>. Never copy a skill into.claude/skills/: two copies drift.scripts/is for one-off imperative operations against the workspace (memories, users, content libraries). If the CDK can declare it, it belongs ininfra/instead.evals/uses promptfoo. When you change a system prompt ininfra/, update or extend the matching suite inevals/and runnpm run eval(needs anOPENAI_API_KEY). It is not wired into CI: add a workflow when the suites are worth gating on.outputs/entries are directories namedoutputs/YYYY-MM-DD-<slug>/, each with aREADME.mdcarrying anoutcome:field (meetings, replies, pipeline attributed, or "none" with a reason). That field is what lets motions be ranked by results instead of opinions. Read recent entries for memory before starting related work. Never rewrite an existing entry. An entry can also declare a one-off run (defineBatch/defineRun) that the merge runs exactly once: it carriesoutcome: pendinguntil the run finishes, and the outcome is filled in once. Use thenew-executionskill.
Workflow
- Change
context/orinfra/on a branch. Before opening the PR, runnpm run lintif you touched any prose, andnpm run typecheckif you touchedinfra/orscripts/. Both also run in CI. - Write the pull request for someone who reads only its title and
description (
.github/PULL_REQUEST_TEMPLATE.md): the title says what it does, the description opens with one sentence on what happens if they approve, then the one decision only they can make and anything a person must still do by hand, each only when there is one. Then## Why, at most three sentences, and## Detailfor everything else. Never restate what merging deploys, runs or sends: Cargo's plan says it, and a sentence written from habit can be wrong. - Cargo plans the pull request and posts the diff as a comment and a check, once the repository is connected to the workspace (Settings > States). Review it the way you would review a terraform plan.
- Merging to main deploys. Never run
cargo-ai cdk deploylocally against production. Local deploy is denied by convention and hard-blocked per agent: Claude Code (.claude/settings.json), Cursor (.cursor/rules/manifest-guardrails.mdc), and Codex (.codex/config.tomlsandbox).
Rules
- The confidence ladder. Files in
proof/,signal/, andinsight/carryconfidence: hypothesis | validated | proven. Outreach motions may cite onlyvalidatedorprovenknowledge: the linter fails the PR otherwise. A claim earnsvalidatedon its second independent occurrence. Once is an anecdote. Use thecapture-feedbackskill; never upgrade a claim just to unblock copy. - Outreach motions declare a hypothesis before any copy exists: segment,
angle, trigger, expected_reply_rate. Enrichment, scoring, routing, and signal
plays are not campaigns and are exempt: they declare intent through the
outcome they serve in
plan/outcomes.md. Every motion setskind:. - Absolute dates everywhere.
2026-07-14, never "next quarter". - Append-only history. Outputs and logs are corrected with a dated addendum, never rewritten. A major context rewrite archives the prior version instead of overwriting it.
- New motions: use the
new-motionskill. It createscontext/motion/<name>.md, the implementation ininfra/, andevals/<name>/together so the three layers never drift apart. - Copy style: no em dashes anywhere. Use colons, parentheses, commas, or split the sentence.
- When a task produces a real-world result (a list built, a campaign shipped, a
research brief), archive it under
outputs/before finishing. - Something to run once (enroll a list, backfill a field) is an execution in
outputs/, not a script and not a local run: use thenew-executionskill, so it is previewed on the PR and runs exactly once on merge. - Personal vs shared: experiments and personal mess go in
scratch/(gitignored), personal skills in~/.claude/skills/, personal instructions inCLAUDE.local.md. Work enters the shared layers only through a PR, once it has proven itself. If asked to save something and it is not clearly team-ready, default toscratch/.
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
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.

