Imported from av/facts (
skills/facts/SKILL.md). Install upstream withnpx skills add av/facts --skill facts. Copyright stays with the author.
facts
A CLI for fact-driven development. You use it to specify what must be true about a project, then validate that reality matches.
Project: https://github.com/av/facts
Installing
If facts is not installed, install it with one of:
curl -fsSL https://raw.githubusercontent.com/av/facts/main/install.sh | shnpm install -g @avcodes/factspipx install facts-cliVerify with facts --version. The command is always facts — never npx facts.
Core idea
A .facts file is a flat list of atomic truth statements about a project. Each fact can optionally have a shell command that verifies it. The fact sheet serves as both specification (what should be true) and documentation (what is true) — the difference is just which direction you're working from.
- the API returns JSON
- label: project builds
command: cargo build
- label: tests pass
command: cargo test
tags: [ci, core]
That's the entire format. Plain strings for simple facts, mappings when you need a command, tags, or explicit ID. Allowed mapping keys: id, label, command, tags — nothing else.
Essential commands
Recommended short aliases (all extra arguments are passed through):
ll=list --light— the most common "skim" view (markdown headings + dim IDs)ls=listrm=removeat <id> <tag>=edit <id> --add-tag <tag>(supports multiple IDs and--label/--new-idetc. after the tag)rt <id> <tag>=edit <id> --remove-tag <tag>
These are the highest-ROI shortcuts for daily and agent use.
See everything:
facts ll
facts ll --tags "not implemented"
facts ll --has-command
facts list --light # or the alias: facts ll
Validate:
facts check # default: command-facts only; manual stay ?
facts check --tags "ci"
facts check --id abc # check a single fact by ID
facts check --section cli/check # check one section
facts check --agent # opt in: .config agent or PATH preset
facts check --agent claude # opt in: specific agent
facts check --no-agent # explicit skip (redundant with default)
facts check --has-command # check only facts with commands
facts check --manual --agent claude # check only manual facts with agent
facts check --strict # fail on unverified manual facts (CI)
facts check --strict --agent claude -q # CI: verify everything, fail on gaps
facts check -q # quiet: exit code only (for CI)
facts check -v # verbose: show agent prompts and output
check is your primary feedback loop. It lints the files first (aborting on structural errors), then runs every command-fact and reports pass/fail/manual. Run it often. Exit 0 means all command-facts pass; manual facts don't affect the exit code unless an agent verifies them and finds failures, or --strict is used (which treats any unverified manual fact as a failure).
Manual facts (? in output) are your responsibility. They have no command — you verify them by reading the relevant code. For each ? fact: read what it claims, check the code, then report PASS or FAIL with a one-line reason. Reporting "N manual" without checking each one is not acceptable — those facts exist because they describe behavior that matters.
Configure the project for agent verification:
facts config set agent claude # persist agent choice in .facts
facts config set batch-size 5 # how many facts per agent call
facts config show # see current config
facts config rm agent # remove a config key
Add facts:
facts add "users can sign up" --section features/auth
facts add "signup returns 201" --command "curl -s -o /dev/null -w '%{http_code}' localhost:3000/signup | grep 201" --section features/auth
Edit / tag facts (lifecycle transitions):
facts at <id> "implemented" # most common: quick add tag
facts rt <id> "spec" # quick remove tag
facts at <id> "spec" --new-id xyz # extra flags still work
facts edit <id> --label "corrected statement"
facts edit <id> --command "new check command"
facts edit <id> --remove-command # convert back to manual fact
facts edit <id> --remove-command --add-tag "draft" # demote to manual + draft
Prefer at/rt (or the long --add-tag / --remove-tag forms) over the full --tags replacement. The latter replaces all tags silently. Use --remove-command to strip a validation command and convert a fact back to a manual fact (useful when demoting across the lifecycle).
Remove facts:
facts rm <id> # or the long form: facts remove <id>
Scaffold a new project:
facts init
Run facts <command> --help for the full flag reference.
How facts work
Files: .facts is the default. Additional sheets use semantic names (cli.facts, api.facts). All *.facts files in the project root are discovered automatically.
Sections: Markdown headings (#, ##, etc.) create hierarchical sections addressable by path (e.g. cli/subcommands). Sections are created when you add to them and removed when their last fact is deleted.
Tags: @word tokens for filtering. Inline for plain strings (- some fact @mvp), tags: key for mappings. Stripped from the label before display and ID hashing. Filter with boolean expressions: --tags "mvp and not blocked".
Lifecycle tags: Three well-known tags drive the agent workflow:
@draft— rough idea, needs refinement and atomization@spec— precise and actionable, ready to implement@implemented— true and backed by code- Untagged facts are ground truth — verified against the codebase
The lifecycle flows @draft → @spec → @implemented. Each companion skill owns one transition.
IDs: Every fact gets a short ID (3+ chars) derived from its label hash. Stable as long as the label doesn't change. Use --id or --new-id to override.
Validation: Commands run via $SHELL (fallback sh) in the project root. Exit 0 = fact holds. Write commands that are fast and idempotent — they run on every check.
Configuration: The .config fact (explicit id: .config) stores project-level settings in the main .facts file. Known keys: agent (preset name or shell template), batch-size (facts per agent call). Manage via facts config set/get/show/rm.
Agent verification
facts check can delegate manual facts (those without a command) to an AI agent for verification. The agent reads the codebase and reports whether each fact holds.
Opt-in only: agent verification runs only when you pass --agent. Without it, manual facts stay ?.
Resolution chain when --agent is passed (first match wins):
--agent <value>— preset name or shell template with%p- Bare
--agent—.config agentkey (set viafacts config set agent claude) - Bare
--agentwith no config — scans PATH for known agent binaries in preset order
Presets: claude, codex, pi, droid, opencode, mi. Each maps to a command template (e.g. claude becomes claude -p %p). Custom templates use %p as the prompt placeholder: --agent "my-agent run %p".
How it works:
- Command-facts run first (sequential, immediate output)
- Remaining manual facts are batched (default 10 per batch, configurable via
--batch-sizeor.config batch-size) - Each batch is sent to the agent as a single prompt with fact IDs and labels
- The agent's output is parsed for verdict XML tags
Verdict format — when you ARE the agent being called, respond with one XML tag per fact:
<fact-verdict id="abc">PASS</fact-verdict>
<fact-verdict id="def">FAIL the feature is missing from auth.rs</fact-verdict>Rules:
- Use
PASSfor facts that hold,FAIL reasonfor facts that don't - Include a one-line reason after FAIL
- Emit one verdict per fact — do not skip any
- The
idattribute must match the fact ID from the prompt - Markdown fences and ANSI codes are stripped before parsing
- If an ID appears multiple times, the last verdict wins
- Facts with no matching verdict are reported as "unresolved" (treated as failures)
Flags:
--no-agent— skip agent verification (default behavior without--agent)--id <ID>— check only specific facts (repeatable, composes with all filters)--section "path"— filter by section path (case-insensitive sub-path match, includes children)--tags "expr"— boolean tag filter (e.g."ci and not draft")--search "expr"— full-text search against section + label + tags--file "name"— filter to a specific.factsfile--has-command— check only facts with a validation command (skip manual facts)--manual— check only manual facts (no validation command). Useful with--agent--depth N— limit section nesting depth--timeout <secs>— per-command and per-agent-batch timeout--strict— treat unverified manual facts as failures (non-zero exit). Essential for CI pipelines that require every fact to be verified by command or agent--quiet/-q— no output, exit code only (0 = all pass)--verbose/-v— show agent prompts and raw output for debugging
Writing good facts
Facts should describe what the project does — its behaviors, features, and contracts — not what files it has or what libraries it uses. The test: if an agent reimplemented this project using only the fact sheet, would the result behave correctly?
- Behavioral — describe what happens from the user's perspective ("expired tokens are rejected with 401") not what exists in the code ("there is an auth module")
- Atomic — one truth per fact, independently verifiable
- Declarative — state what is true, not what to do ("uses PostgreSQL" not "set up PostgreSQL")
- Stable — shouldn't change with every commit ("tests pass" not "there are 47 tests")
- Falsifiable — you could imagine a broken implementation where this fact would not hold
Structural facts (dependencies, file layout) are supporting detail, not the main content. A fact sheet full of "uses X library" and "has Y directory" tells an agent nothing about how the project actually works.
Good validation commands are fast, idempotent, and test one thing. Prefer test -f, grep -q, and short script checks over running full builds.
What makes a good fact
Each quality below is the inverse of a failure mode observed in real agent-maintained projects.
- Spec-first — the fact was written before the code, not after. A fact sheet is a specification, not a changelog. Record completion by tagging
@implemented, not by adding new facts about work you already did. - Behavioral label — the label describes what a user or caller observes. Source file names, line numbers, branch names, and coverage percentages never belong in a fact label — those belong in tests.
- Black-box check — the command tests from the outside: runs the tool, calls the API, checks output. If your command parses source files for variable names, function signatures, or specific string positions, rewrite it to test behavior.
- Independent value — removing the fact would lose something. If an existing test already validates the same claim, the fact is a pass-through wrapper — state a different claim or skip it.
- Actively curated — before adding a fact, check whether existing facts already cover the behavior. Merge related facts instead of letting the sheet grow without bound. A single file past ~80 facts needs splitting.
- Contract-level — one fact per feature contract, not one per code path. If three facts describe the same endpoint's success, error, and edge-case paths, consolidate into one fact about the endpoint's contract.
Domain vocabulary
The ## domain section in the main .facts file defines the project's key entities and their relationships. These are facts too — atomic truth statements — but they describe the conceptual model rather than specific behaviors.
Entity facts name and define a concept — the common pattern is a <Name> is <definition>.
Relation facts state how entities connect. Use the defined entity names in natural declarative statements — there is no rigid grammar, but the connection should be specific.
## domain
- a FactSheet is a parsed *.facts file containing sections and facts
- a Section is a heading-delimited group of facts, nestable via heading depth
- a Fact is an atomic truth statement: a label, optional command, optional tags
- a FactSheet contains a preamble and nested Sections
- check validates command-bearing Facts by running them in the project root
- Tags filter Facts via boolean expressions
Use these names consistently across the fact sheet. The domain section lives in the main .facts file. It is built by facts-discover, refined by facts-refine, and extended by facts-implement when new concepts emerge during implementation.
Agent workflows
Start of work — always do this first:
facts ll # read the full spec (skim view)
facts check # see what holds and what doesn't
Use the fact sheet to orient before writing code. It is the source of truth for what the project should look like and what is already validated.
Define the spec (most common user intent): When the user says "work on facts", "add facts", or "define the spec", they want to collaboratively define what should be true — not audit what already is.
facts add "users can sign up" --section features/auth --tags "draft"
facts add "signup returns 201" --command "curl -s ..." --section features/auth --tags "spec"
- Discuss with the user what the project should look like
- Add rough ideas as
@draft— they'll be refined into precise specs later - Add precise, actionable facts as
@spec— they're ready to implement - Do NOT remove
@draftor@specfacts — they represent intended future work - Do NOT run
facts-discoverunless the user explicitly asks to sync with reality
Track lifecycle progress:
facts ll --tags "draft" # rough ideas to refine (ll = list --light)
facts ll --tags "spec" # ready to implement
facts ll --tags "implemented" # done
facts ll --tags "not implemented" # all remaining work
facts check # verify
Maintain accuracy during coding work: When you add a feature, add corresponding facts. When you fix a bug, verify related facts still hold. When you remove code, remove obsolete facts. Before adding a fact, check if the existing sheet already covers the behavior — consolidate near-duplicates rather than accumulating them.
facts check # find failing facts
facts at <id> "implemented" # tag done (alias for edit --add-tag)
facts rm <id> # remove obsolete facts
facts add "new truth" --section foo # add discovered truths
Companion skills
Each skill owns one lifecycle transition:
- facts-discover — scan the codebase and classify every fact by lifecycle stage (
@draft,@spec,@implemented). Use to triage or bootstrap. - facts-refine — pick up
@draftfacts and refine them into precise@specfacts with the user. Use when drafts need sharpening. - facts-implement — pick up
@specfacts and implement them in code, then tag@implemented. Use when specs are ready to build.