agentleFS
anivaranivar/agentic-sdlc/skills/evidence/SKILL.md

evidence

Skill0 starsChanged 37 days agoMITView on GitHub

Calibrate every claim before it leaves the session: tag each factual clause Verified, Measured, Inferred or Planned, verify behaviour at the production sha, treat agent and bot output as a lead rather than evidence, read sources in full, compute a headline number twice, sweep every surface after one defect, and record what a source taught in a dated document. Triggers on: "verify", "is this true", "check the source", "what is the evidence", "at which sha", "production data", "sweep", a number about to be reported, a finding about to be filed, a summary or digest offered in place of a source, "confirmed", "resolved", "live", "100%", "on track".

8 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

4 rules. 0 are common, 4 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 8 edited versions we read keep these.

  • 4

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Never upgrade a level. Merged is not deployed.

    line 53
  2. Distinctive:

    Before answering on a topic, open what is already established (the topic's memory or dated documents), lead with that, then add what is new.

    line 80
  3. Distinctive:

    Before showing a deliverable, run the self-review in review-change (REVIEW-7).

    line 84
  4. Distinctive:

    After a research task, write the dated document (EVID-10) before moving on.

    line 85

“Common” is measured: kept by at least 6 in 10 of the 8 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Evidence

A claim is worth exactly what was done to check it. This skill fixes the vocabulary for how much was done, and the handful of checks that turn a plausible sentence into a verified one. The other skills cite these rules by id where they make a claim; the ids are stable and live in rules/registry.json. The decider named across this set is the role that approves and gives go-aheads: a role, not a person.

When to use

SituationRule
A sentence states a fact about code, data, a system or a person's workEVID-1
The fact is about running behaviourEVID-2
The fact came from an agent, a retrieval system or a review botEVID-3
A summary, digest or excerpt is on offer instead of the sourceEVID-4
A number is about to lead a message or a documentEVID-5
One defect was just foundEVID-6
A finding is about to be filed against someone else's codeEVID-7
A file is about to be asserted or editedEVID-8
Custom code is about to be writtenEVID-9
A research task read a source that taught somethingEVID-10

The tags

TagMeaningWhat earns it
Verifiedthe source was opened in this sessionthe file, log, record or page, read at the cited location
Measureda number with the query that produced itthe command or query, repeatable
Inferredreasoning from evidence, not the evidence itselfthe premises named
Plannedintended, not donea step in a plan

Never upgrade a level. Merged is not deployed. Deployed is not verified in data. "Live", "100%", "PASS", "confirmed", "resolved" and "on track" need matching evidence; "green" is never a status. Give the sha, the date and the lane, or the gate.

The rules

IdRuleWhat it means in practice
EVID-1Tag every claimEach factual clause carries a tag; a reader treats an untagged clause as Inferred, which is rarely what was meant. Calibration is labelled honestly: "unverified", never blank.
EVID-2Verify at the production shafile:line at the sha found from the live artifact or the deployment record, never a branch head or the latest tag. Behaviour comes from production data; dates from the deployment record. Code alone is a hypothesis; an absence claim needs data.
EVID-3Agent output is a leadA sub-agent's quote, a retrieval hit or a bot's finding points at a line; the author opens that line before it is used. A quote is verbatim only after the author compared it.
EVID-4Read sources in fullA summary is a pointer to the source. Slice a large source rather than skip it. Say which sources were read in full and which were not.
EVID-5Compute a headline number twiceSecond method, different path. Never divide counts from two systems with different identity models without the join. A number that moves between identical runs is a harness artefact (TEST-6).
EVID-6Sweep after one defectName the pattern, list every surface, report each as confirmed, not affected, or not checked. "Not checked" is a valid, honest row.
EVID-7Refute before filingThe second pass argues the other side. If the finding survives, file it with both passes shown.
EVID-8Re-read before assertingEspecially after checkout, stash or rebase, and before updating a change request whose state may have moved.
EVID-9Survey before buildingTool, data source, maintenance, fit: record it, name the pick. Custom code covers only the gap. A vendor's or partner's definition of a mechanism wins over a new mode.
EVID-10Record what a source taughtA dated document in the topic folder: question, sources (full or partial), method, findings with evidence, open questions, exact commands. Never overwrite the earlier one.

Weighting evidence

Weight evidence by what it cost to produce: shipped code, then a written design, then a documented observation, then an opinion. A claim resting on the cheapest kind is Inferred until the costlier kind is opened.

How to apply in a session

  1. Before answering on a topic, open what is already established (the topic's memory or dated documents), lead with that, then add what is new.
  2. While working, tag as you write. An untagged clause reads as Inferred, which is rarely what was meant.
  3. Before showing a deliverable, run the self-review in review-change (REVIEW-7).
  4. After a research task, write the dated document (EVID-10) before moving on.

What a tagged paragraph looks like

The retry loop re-sends non-idempotent requests (Verified: src/client/send.py:142 at the production sha 3f9c2ab, from the deployment record of 2026-10-08). It produced 17 duplicate orders last week (Measured: SELECT count(*) … WHERE created_at > '2026-10-01' on the production replica, re-computed from the payment provider's export: 17). The cause is the timeout after the server accepted the write (Inferred from the request logs; no direct evidence of the accept). A fix with an idempotency key is in review (Planned).

Four clauses, four tags, one number computed twice (EVID-5), the inference marked as an inference rather than upgraded to a finding (EVID-1).

Companion skills

NeedSkill
The adversarial read before showing workreview-change
The words that need evidence, and the form of a verdict-first messageplain-voice
Where decision records and glossaries livedecision-records
Plans with assumptions tied to testsplan-project

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.