intent-replay
Skill29 starsChanged 7 months agoMITView on GitHub
Tracks intent, decisions, and handoffs in .project when this repository's code or other source material changes. Use before that work. First check whether .project exists. If it does not, bootstrap it before continuing. If it does, update intent when the goal changes and record the decisions later work must follow. A commit or push is not a reason to write these files. After a commit or push, leave the tree as it is.
In this repository13 other files it gives its agents.
- agent-creatorSkill
- agent-skill-creatorSkill
- art-engineSkill
- capability-architectureSkill
- copy-sanitizerSkill
- md-compressorSkill
- md-design-systemSkill
- opportunity-gap-finderSkill
- project-graphSkill
- skill-scanSkill
- story-architectSkill
- token-estimatorSkill
- vector-drawingSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
14 rules. About 0 are common, 8 are distinctive, and 6 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
6
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Distinctive: line 22
Validate continuity: After bootstrap and before the final handoff, run the continuity check in references/lifecycle.md.
- Distinctive: line 27
Respect permissions and conflicts: Do not commit, push, merge, deploy, install integrations, or run destructive commands without …
- Very specific: line 35
If .project/ or the Intent Replay block is missing, follow references/bootstrap.md.
- Very specific: line 36
If .project/ and the Intent Replay block already exist, do not bootstrap again.
- Very specific: line 42
Read .project/NOW.md first, then .project/INTENT.md and the referenced active task.
- Distinctive: line 50
Keep acceptance criteria, progress, links to changed artifacts, and validation evidence on that task.
- Very specific: line 51
Add a consequential decision to DECISIONS.md with reasoning and scope.
- Distinctive: line 52
If an agent lacks write access, provide a proposed change or patch.
- Distinctive: line 60
Run the continuity check, resolve any consistency errors, then tell the user what changed, what was validated, where the work lives, and …
- Distinctive: line 61
Do this recording before a requested commit, never after one.
- Distinctive: line 86
Before a change to this repository's code or other source material.
- Very specific: line 87
When .project/ is missing, so it is bootstrapped before that work continues.
- Very specific: line 88
When .project/ already exists, so intent and decisions stay current.
- Distinctive: line 89
When resuming in another agent, completing work, or handing off.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill 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
Intent Replay
Purpose
Maintain one durable, agent-independent source of project intent, decisions, active work, and verified handoffs. The repository owns project context; no agent's conversation is authoritative. Intent Replay coordinates context, not model selection, commits, or deployments.
Non-Negotiable Rules
- Start with Intent Replay. Before other work in the repository, look for
.project/and the Intent Replay block inAGENTS.md. If either is missing, bootstrap. If both exist, read them and keep the tracking files current as the work changes. Do not skip this step, and do not start a second tracking system. TheAGENTS.mdblock is the standing order that tells the next agent to do the same. A commit or push does not start this update. - Bootstrap safely. At the intended project root, create missing Intent Replay files and the
AGENTS.mdintegration. Never reset, truncate, replace, or delete existing project content. If initialization cannot be performed, explain the blocker and provide the exact proposed changes. - Read before acting. Inspect
AGENTS.md,.project/NOW.md,.project/INTENT.md, the relevant active task, and only the decisions needed for the task. Reconcile the snapshot with the repository and accessible Git state before trusting it. - Validate continuity. After bootstrap and before the final handoff, run the continuity check in references/lifecycle.md. Fix structural problems. A clean check is not proof the implementation works.
- Never fabricate progress. Label facts as observed, reported, planned, or unverified. Never represent an unrun test, remote-only view, or prior agent claim as verified.
- Update at meaningful boundaries, not per message. Update
INTENT.mdwhen the goal, audience, scope, or a constraint actually changes. Record a decision only when a later agent would choose wrong without it. Update tasks when scope, implementation, blockers, or validation changes. RefreshNOW.mdso it matches the current intent and the decisions still in force. Keep both files short. A commit, push, or sync to GitHub is not one of these updates. - Stop after a commit. When the user asks to commit or push, or when a commit or push has just happened, do not edit
.project/or the Intent Replay block afterward. Do not record that the commit happened in intent, decisions, orNOW.md. Leave any uncommitted Intent Replay edits out of that follow-up. The tree stays as the commit left it. - Archive tasks; compact old decisions. Move accepted and verified tasks from
tasks/active/totasks/done/. Keep task IDs and evidence. WhenDECISIONS.mdgrows past about 120 lines, compact superseded decisions, and accepted decisions that no longer affect the work, down to one line each. Keep every decision ID. Do not silently complete work or remove task files. - Respect permissions and conflicts. Do not commit, push, merge, deploy, install integrations, or run destructive commands without authorization. Detect concurrent edits and discrepancies; preserve both sides and request a resolution rather than overwriting.
- Do not store secrets or personal data.
.project/records where the project is and what it is trying to do. Never write secret keys, credentials, tokens, raw environment values, private chat transcripts, or personal data into these files.
Workflow
1. Bootstrap, or continue
- Identify the actual workspace/repository root. In a monorepo use the intended repository or explicitly scoped subproject, not an arbitrary parent directory.
- If
.project/or the Intent Replay block is missing, follow references/bootstrap.md. Copy missing files from this skill'sassets/templates. Create only missing paths. - If
.project/and the Intent Replay block already exist, do not bootstrap again. Continue with the resume, record, and handoff steps below. - Inspect existing
AGENTS.mdinstructions for conflicts. A generated Intent Replay block supplements, never supersedes, existing rules. - Report what was added and what already existed. For a new project, ask for the intent when unknown; placeholders are explicitly unconfirmed, not project facts.
2. Resume reliably
- Read
.project/NOW.mdfirst, then.project/INTENT.mdand the referenced active task. Read matching entries in.project/DECISIONS.mdonly as needed. - Inspect actual code, files, and available Git state (
status, branch, recent changes) within available permissions. Distinguish local uncommitted changes from remote history and record the visibility boundary. - Identify stale claims, conflicting plans, missing tasks, or unverified results. Reconcile only what can be established; otherwise flag the uncertainty and stop any action that could overwrite work.
- State the current objective and next safe action briefly before executing.
3. Work and record
- For substantial work, create or reuse an active task using references/state-contract.md. Do not invent tasks for trivial edits.
- Keep acceptance criteria, progress, links to changed artifacts, and validation evidence on that task. Preserve the original user intent and decisions.
- Add a consequential decision to
DECISIONS.mdwith reasoning and scope. Supersede an older decision by a new ID. Compact the log when it grows past the limit in references/state-contract.md. - If an agent lacks write access, provide a proposed change or patch; never claim the shared project state was updated.
4. Complete, clean, and hand off
- Apply references/lifecycle.md: verify acceptance criteria, record actual check results, and confirm approvals if required.
- Move a task to
done/only when genuinely complete; otherwise leave it inactive/with the precise blocker or remaining work. Do not delete completed tasks automatically. - Rewrite
NOW.mdas a compact, accurate current snapshot: goal, active task pointers, verified state, blockers, next action, handoff provenance, and visibility limits. Remove stale completed work from the active snapshot, not from history. - Compare files against their pre-write state to avoid lost updates. If concurrent work changed the same file, reconcile or stop instead of clobbering it.
- Run the continuity check, resolve any consistency errors, then tell the user what changed, what was validated, where the work lives, and what remains. Never imply a commit or synchronization occurred unless confirmed.
- Do this recording before a requested commit, never after one. If the user asked only to commit or push, skip the recording and leave the files untouched.
Working set
AGENTS.md
.project/
INTENT.md
NOW.md
DECISIONS.md
tasks/
active/
done/Reference Index
Load a reference only when that operation is needed.
- Initialization and existing-file protection: references/bootstrap.md
- File contracts, task IDs, and decision records: references/state-contract.md
- Completion, cleanup, and cross-agent handoffs: references/lifecycle.md
When To Use This Skill
- Before a change to this repository's code or other source material.
- When
.project/is missing, so it is bootstrapped before that work continues. - When
.project/already exists, so intent and decisions stay current. - When resuming in another agent, completing work, or handing off.
- A request to commit or push is not a trigger to update these files.
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.

